# Decosa

> Applied AI on open models: tools that each do one real job (check a brief, run the visit, fill a form), and the open
> models under them, by API in the OpenAI format. Every hosted model call returns a receipt signed with Decosa's key.
> 86 tools. Use them hosted, or, on request, self-host them on your own hardware.

Site: https://decosa.ai (canonical; cite these URLs). This file is the index; [/llms-full.txt](https://decosa.ai/llms-full.txt) adds every
tool with its details and the API routes.

## Start here

- Base URL: `https://api.decosa.ai` (the OpenAI chat format under `/v1`, so any OpenAI SDK works with `base_url` set to `https://api.decosa.ai/v1`).
- Auth: `Authorization: Bearer dk_…` on every call.
- Key: create one at [API keys](https://decosa.ai/account/keys). It is shown once; store it as `DECOSA_API_KEY`. A new key has a free budget of $0.02 of hosted usage a day at list price, reset every UTC day. Past it, calls get HTTP 402 until the next UTC day, or until the key has credits. A key with credits has no daily cap.
- First call:

```bash
curl -s -i https://api.decosa.ai/v1/chat/completions \
  -H "Authorization: Bearer $DECOSA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.8-27b", "messages": [{"role": "user", "content": "Hello"}]}'
```

- Receipt: the response header `x-decosa-receipt` holds the receipt id; `GET https://api.decosa.ai/receipts/<id>` returns the signed record. [What a receipt proves](https://decosa.ai/docs/receipts).
- Credits: One credit is $1 of usage at list price: hosted API calls, the coding agent, Studio and tool runs all draw on the same balance. Bought with US dollars, by card. Not transferable, and no cash value. They last 12 months from purchase. Usage is metered from each call's signed receipt; your ledger shows every debit and the receipt behind it. Requests stop when the balance reaches zero, and we email you before that. At zero credits the API answers HTTP 402; over a rate limit, 429 with `Retry-After`.
- [Quickstart](https://decosa.ai/docs) ([markdown](https://decosa.ai/docs.md)): key, call, receipt, verify, with no browser.

## Machine-readable

- [/api/catalog.json](https://decosa.ai/api/catalog.json): every tool (86), with its home, status, verification, data handling and links (page, JSON, metrics, prompts). Schema 1.
- [/api/contract.json](https://decosa.ai/api/contract.json) (route index) and [/api/contract.md](https://decosa.ai/api/contract.md) (the full API contract; the source of truth).
- [/api/metrics.json](https://decosa.ai/api/metrics.json): per tool, the eval summary (split, held out or not, caveats, date), cost per run and models.
- `/use-cases/<id>.json`: one tool in full (stack, tiers, samples, data handling, prompt URLs), e.g. [/use-cases/clinical.json](https://decosa.ai/use-cases/clinical.json).
- `/prompts/<id>-hosted.md`, `/prompts/<id>-selfhost.md`, `/prompts/<id>-assemble.md`: copy-paste prompts for a coding agent. Each tool's are in catalog.json `links.prompts` (null where a tool has none).
- [/api/hardware.json](https://decosa.ai/api/hardware.json): device specs and each tool's memory needs, for self-hosting.
- [/llms-full.txt](https://decosa.ai/llms-full.txt): this index plus every tool, its details and the API routes.
- [For AI agents](https://decosa.ai/agents): stable selectors, deep links, rate limits and what agents may drive.

## Homes

- [Legal](https://decosa.ai/legal): Briefs, depositions, discovery, records, client calls.
- [Healthcare](https://decosa.ai/clinics): The visit, then notes, audits and denials.
- [Finance & insurance](https://decosa.ai/tools/finance): Alerts, SARs, disputes, claims, recorded calls.
- [HR](https://decosa.ai/hr): Investigations, leave and I-9 audits.
- [Compliance & trust](https://decosa.ai/industries/compliance-trust): Security questionnaires, control evidence, AI audits.
- [Operations](https://decosa.ai/tools/operations): Forms, shop pages, imports, listings: any industry.
- [Life sciences](https://decosa.ai/tools/life-sciences): Trial numbers, safety cases, labels, papers.
- [Create](https://decosa.ai/studio): Stories, songs and dubs, made with consent and labelled as AI.
- [Developers](https://decosa.ai/developers): Add Decosa to the coding agent you already use, move prompts to open models with evidence, and prove what your AI did.
- More industries: [Public sector](https://decosa.ai/industries/public-sector), [Sales and marketing](https://decosa.ai/industries/sales-marketing), [Science and research](https://decosa.ai/industries/science-research), [Music](https://decosa.ai/industries/music), [Film, TV and games](https://decosa.ai/industries/entertainment).
- [All tools](https://decosa.ai/tools): filters in the query string, e.g. `/tools?industry=legal&deploy=selfhost`.

## Docs

Each docs page is also plain markdown at the same path plus `.md`.

- [Quickstart](https://decosa.ai/docs) ([markdown](https://decosa.ai/docs.md))
- [API reference](https://decosa.ai/docs/api) ([markdown](https://decosa.ai/docs/api.md))
- [Receipts and verification](https://decosa.ai/docs/receipts) ([markdown](https://decosa.ai/docs/receipts.md))
- [Sealed tier (paused at launch)](https://decosa.ai/docs/sealed-tier) ([markdown](https://decosa.ai/docs/sealed-tier.md))
- [Self-hosting](https://decosa.ai/docs/self-host) ([markdown](https://decosa.ai/docs/self-host.md))

## Trust

- [Where your data goes](https://decosa.ai/data): per tool, what stays on your machine when self-hosted, what the hosted service processes, retention, and every outside call.
- Hosted calls travel encrypted to Decosa's API server; the open models are run by NEAR AI through OpenRouter, with Reka AI as the only fallback, under Decosa's account and with zero data retention. Nothing is used for training. [Hosted API](https://decosa.ai/developers#serving): who serves each model.
- Hosted video input, hosted renders other than UGC ads, live speech recognition, speaker separation, document reading, retrieval and the sealed tier are off at launch; their tool pages show recorded runs. About [the sealed tier](https://decosa.ai/docs/sealed-tier).
- [Verify a run](https://decosa.ai/docs/receipts): what a signed receipt proves, and what it doesn't.
- [Accuracy](https://decosa.ai/metrics): eval results and cost per run for every tool. [Status](https://decosa.ai/status): every hosted model and service. [Security](https://decosa.ai/security).
- [Terms](https://decosa.ai/terms) · [Privacy](https://decosa.ai/privacy) · [Acceptable use](https://decosa.ai/acceptable-use) · [AI notice](https://decosa.ai/ai-notice).

## Self-host

- [Self-host](https://decosa.ai/self-host): run a tool on your own GPU with Docker Compose, or [on a Mac](https://decosa.ai/self-host/mac). The source code and container images are not public yet: self-host is on request ([ask here](https://decosa.ai/contact?topic=self-host)).
- [Will it run on my hardware?](https://decosa.ai/self-host/hardware): pick a GPU, Mac or cloud box and see which tools fit.
- [Self-hosting docs](https://decosa.ai/docs/self-host) ([markdown](https://decosa.ai/docs/self-host.md)).

## Apps by Decosa

- [Everdust](https://decosa.ai/family#everdust): Calm 3D worlds that hear music. Apple Vision Pro, iPhone, iPad, Mac and Apple TV; coming soon. Site: everdust.ai.
- [SottoMD](https://decosa.ai/family#sottomd): A visit copilot for independent clinicians: a live transcript, the questions still worth asking, and a note checked against what was said. Web; coming soon.
- [Cue](https://decosa.ai/family#cue): Finds where each song starts in a long DJ mix, so you can skip song by song, from the lock screen or the car. iPhone and Mac; in testing.
- [Snug](https://decosa.ai/family#snug): Scan a room with a LiDAR iPhone, then walk a measured, furnished 3D model and change it by asking. iPhone and iPad (LiDAR); in testing. Site: snug3d.com.
- [AutoTalent](https://decosa.ai/family#autotalent): Give it a role: it searches widely, ranks the people it finds with the evidence for each, and writes the outreach. Web. Site: autotalent.ai.
- [ElmoSEO](https://decosa.ai/family#elmoseo): Shows how ChatGPT, Google and Perplexity see a business, and a short list of what to fix first, in plain English. Web. Site: elmoseo.com.
- [Dr. Grey](https://decosa.ai/family#drgrey): Pick a goal and see the supplements with the strongest research, what the studies found, and a grade for each. Web. Site: drgrey.ai.

## Driving the site with a browser harness

- Landmarks: one <header>, <nav aria-label="Main">, <main id="main">, <footer>. One h1 per page.
- Stable hooks (data-testid):
  catalog: use-case-card / use-case-row (data-use-case-id, data-status), catalog-search, catalog-filter (data-facet, data-value);
  search: search-open, search-input, search-result (a link; data-id, data-href), search-ask-guide;
  site guide (Ask Decosa): guide-open, guide-open-menu, guide-panel, guide-opener, guide-profession (data-id), guide-job (data-id = tool id), guide-nav, guide-undo, guide-input, guide-send, guide-starter, guide-answer (aria-busy while streaming), guide-source, guide-request, guide-request-send;
  use-case page: tab-stack, tab-live, tab-watch, tab-build, tab-self-host; copy-prompt-assemble (Stack tab),
  copy-prompt-hosted, copy-prompt-selfhost, copy-snippet, and copy-text for the text beside each copy button;
  keys: get-api-key (compact form only), key-name, key-create, key-value;
  console: console (root, data-use-case-id), console-picker, console-picker-option, console-sample, console-run,
  console-record (live audio), console-stop, console-output (aria-busy while running; on consoles that stream a result);
  Watch: watch-sample, watch-play, watch-seek, watch-speed, watch-output;
  data: data-badge (data-self-host), data-badge-hosted (data-hosted), and on /data data-level, data-flow-row (data-use-case-id);
  hardware: hardware-check / hardware-check-row (data-device, data-verdict), hardware-fit (data-use-case-id, data-verdict),
  hardware-device, hardware-count, hardware-custom-arch, hardware-custom-vram, hardware-custom-count, hardware-custom-ram,
  hardware-recommendation (data-tier), hardware-tier-row (data-tier, data-fit), hardware-substitution, hardware-speed,
  copy-prompt-hardware, and on /self-host/hardware hardware-use-case, hardware-catalog, hardware-device-row.
- Hidden tab panels stay in the DOM, so scope locators to the visible panel (#panel-<tab>).
- Every copy button also shows its text in a <pre> next to it, and the same text is at /prompts/<id>-<kind>.md,
  so you never need clipboard access.
- ⌘K / Ctrl+K opens search (a combobox with a listbox of links). The Search button in the header does the same.
- Tabs on /apps/<id> are addressable by hash: #stack, #live, #watch, #build, #self-host.
- Sample deep links: <tool page>?sample=<n>&autorun=0 (n is 1-based, or a sample id). autorun=1 starts the run (slow tools play their recorded real run first; add &mode=live or &mode=replay to choose). The tool page carries the console hooks (console, console-sample, console-run, console-stop, console-output).
- ?reduce-motion=1 turns animations off for the rest of the tab session.

## Facts

Platform facts, checked 2026-09-29. These are the facts the site's guide and the Decosa agent use.

- [ways-to-run] Decosa tools can run up to three ways; each tool page's 'Where it runs' says which it offers. Hosted: through Decosa's hosted API with a dk_ key, with a signed receipt for every model call; the hosted site and API are for made-up or public material (a tool's page may allow more, such as your own non-confidential drafts) and never for real patient data or privileged client files, and a few tools also call outside services (each page's 'What leaves the box' names them). Self-hosted: on your own hardware from pinned containers and a copy-paste setup prompt for a coding agent; some tools are self-host only. Confidential access: a sealed, hardware-attested enclave set up per customer on request; it is not self-serve, not instant, and not yet running for customers.
- [own-material] On the hosted site, "Use your own" is for made-up or public material only. Every tool can also be run on your own hardware (its page explains how).
- [phi] Real patient data (PHI): hosted Decosa Clinics tools take synthetic (made-up) patients only, and confidential access can't take patient data yet either, because no business associate agreement (BAA) is in place. Today the way to use real patient data is to run the tool on your own hardware. A copy of a chart is still patient data. You can still ask about it on [Talk to us](/contact?topic=clinic).
- [baa] Decosa has no business associate agreement (BAA) in place and doesn't offer one today. Nothing Decosa ships is HIPAA-certified or 'HIPAA-ready'; running a tool on your own hardware keeps patient data under your own controls.
- [privileged] Privileged legal material or confidential business data: run it yourself, or request confidential access (a sealed enclave), which is set up on request [Request confidential access](/contact?topic=confidential). Confidential access is not available instantly.
- [confidential-tier] Confidential access runs the model inside a hardware-attested enclave (Intel TDX or AMD SEV-SNP with an NVIDIA GPU in confidential mode), so Decosa as operator can't read prompts and the customer can check the attestation offline. It was proven in a pilot on rented hardware; it is set up per customer on request and is not yet running for customers. Legal teams are the first market; patient data waits for a BAA.
- [sealed-tier] The sealed tier (end-to-end encrypted chat to a model server Decosa operates) is paused at launch, because outside providers run the hosted models. It comes back as a confidential tier inside an attested enclave; the confidential pilot is by request.
- [receipts] Every hosted model call gets a signed receipt: the model and its pinned weights, a hash of the request and of the answer, token counts and cost, signed by Decosa's gateway key. Receipts hold hashes, not text. Anyone can fetch one at https://api.decosa.ai/receipts/<id> and check it offline against the pinned key, for example with `decosa verify <id>`. When a coding agent uses tool calling, the receipt also covers the tool definitions and the exact tool calls.
- [first-party] Hosted model calls (every tool's calls and the OpenAI-compatible chat route, including the Decosa agent) run on Decosa's hosted service: on Decosa's own servers or with inference providers under Decosa's account; /data says which, per model. Each receipt names the serving machine. A receipt proves what Decosa's gateway signed (model, hashes, tokens, cost, serving machine); where the call ran rests on Decosa's signed word. Hardware-attested proof of where a call ran is what confidential access adds.
- [records] Most tools also return a signed, hash-chained record of the run (what went in, what came out, which model calls, with their receipts). Check one at https://api.decosa.ai/record/verify or with `decosa verify <record.json>`.
- [prices] Plans and prices aren't published yet. Each tool page shows the measured cost of a run at the gateway's list price (what its model calls, and any partner render, cost): that is a cost, not a price.
- [list-price] Measured run costs on tool pages use the gateway list price for the hosted model: $0.30 per million input tokens and $1.50 per million output tokens.
- [keys-limits] Hosted API keys (dk_...) are created at /account/keys. A key allows 200,000 generated model tokens a day and 60 requests a minute; a demo session allows 20,000 generated tokens. Use a key as `Authorization: Bearer <key>` against https://api.decosa.ai.
- [openai-compatible] The hosted chat model is OpenAI-compatible: point any OpenAI SDK or agent at https://api.decosa.ai/v1 with model qwen3.8-27b and your dk_ key as the bearer token. It streams, supports tool calling (tools and tool_choice), and every response carries its receipt id in the x-decosa-receipt header. Plain chats take up to 48,000 characters and 2,048 output tokens; requests with tools take up to 400,000 characters and 4,096 output tokens.
- [hosted-images] The hosted chat route also takes images (PNG, JPEG or WebP, sent as data in the request, up to 4 per request) in user messages. They go to the model on Decosa's hosted service, the receipt covers each image's hash, and they aren't kept after the request. Decosa's browser extension sends screenshots only if you choose its screenshot mode; its text-only and hash-only modes are the defaults.
- [accounts] "My runs" are saved in the visitor's own browser only; there are no accounts yet.
- [retention] Each tool page states its own data retention and what leaves the machine. The hosted chat model route stores no prompts or answers, only the receipt (hashes, token counts, cost). Tools that return files keep them only for the stated time: for example, the form filler keeps a hosted run's files for 24 hours, readable only with your key or demo token, and 'Delete this run now' deletes them at once. Self-hosted, your data stays on your machine except for the public lookups a tool names on its page (for example the visit copilot's code lookups send search terms from the visit to NLM Clinical Tables, RxNav and openFDA; if you block them, those lanes report no match and the note is still written).
- [hosted-model] The hosted language model is Qwen3.8-27B (Apache-2.0). At launch Decosa's hosted API doesn't run it on Decosa's own GPUs: inference providers run it under Decosa's account with zero data retention, and Where your data goes (/data) names them. It reads text and images; hosted video input is off at launch. Speech, translation and document-reading tools show recorded runs until their hosted services are switched on. Each tool page lists its models and their licences. The hosted API serves Qwen3.8-27B only: models such as GLM-5.3-Flash and DeepSeek-V4.1-Flash appear on some tool pages as "Wanted" tiers, not served yet.
- [studio-fal] Studio renders run on Decosa's hosted service, except two older video tools that render on fal, a partner GPU provider, under Decosa's account: the Studio video queue (the text prompt only) and the disclosed UGC ad (shot prompts, frames of its invented presenter, and a stock house-voice voice-over). No photo, voice or consent recording from a Studio project goes to fal. fal keeps request data for 30 days by default.
- [hardware-sizing] Self-host sizing for the hosted model (Qwen3.8-27B in NVFP4): Decosa measured it on one RTX PRO 6000 Blackwell (96 GB) at the full 262k context. A 32 GB RTX 5090 (Blackwell) is the Lite tier: same weights at a shorter context (16k–32k depending on the tool), not yet benchmarked by Decosa. On Ada, Hopper or Ampere cards it runs in FP8 and needs 40 GB or more, so a 24 GB RTX 4090 or 3090 runs only the smaller models. Check a machine at /self-host/hardware, or with `decosa selfhost <tool> --gpu <GB>`.
- [licences] Licences: the tools' self-host code (decosa-api) is AGPL-3.0-or-later since 29 Sep 2026: anyone can read, run, self-host and change it, including commercially, and only someone offering a modified version as a service must publish their changes. The client libraries and verifiers (including the Decosa agent and the receipt verifiers) and the gateway are Apache-2.0. The models are open-weight and each tool page lists every model's licence (for example ACE-Step 1.5 is MIT; MiniMax-Music3 has its own commercial terms); non-commercial models are never the default. Models Decosa trains are released under Apache-2.0 after review. The site itself is not open source, and "Decosa" is a trademark.
- [workflows] Today each tool runs on its own, and nothing hands off to the next tool automatically. Decosa is designing chains of tools as workflows; the leading ones are prior-auth pre-check → packet → denial appeal on the same policy and chart, and a personal-injury matter from intake through records and the medical chronology to deposition prep. None is sold as one package yet.
- [computer-use] Tools that work with forms and websites stop before anything is sent: they fill or prepare, a person checks and submits. The form filler never signs, never ticks a certification or consent, never fills passwords, card, bank, Social Security or other ID numbers, and never fills a live website, hosted or self-hosted: for a web form it gives a copy list in the site's order to paste from. Every answer shows the line of your papers it came from. Browser test runs (`test-runs`) run only on their test app, a public test site and sites whose owner verified them.
- [not-compliant] Decosa tools check, draft and prepare work for a qualified person to decide. Decosa never says an output is compliant, certified, approved, cleared or court-ready; each tool page says what it checks and what it doesn't.
- [decosa-agent] The Decosa agent (`decosa agent`) is a coding agent on Qwen3.8-27B with the Decosa tools built in. Hosted, it uses Decosa's hosted chat route: calls run on Decosa's hosted service, prompts and code aren't stored, and every call has a receipt; the hosted route follows the code tool's rule, so don't send proprietary code to it. For proprietary or regulated code, run it on your own GPU with `--local`, where nothing leaves your machine. The same install adds the Decosa tools to Claude Code, Cursor, Codex or any MCP client (`decosa-mcp`). It is in private beta; see /developers/agent.
- [tool-map] Some jobs have purpose-built Decosa tools that are easy to miss: the Endpoint auditor ("Audit a model endpoint", `auditor`) checks whether an OpenAI-compatible endpoint really serves the model and precision it claims and signs the result; "Check an AI note" (the Clinical AI assurance monitor, `clinical-ai-monitor`) checks an AI scribe's or clinical AI's notes against the visit; "Check answers against sources" (`grounding`) checks any draft's sentences against sources; Verified end-to-end test runs ("Run an end-to-end browser test", `test-runs`) run plain-language browser tests and sign a certificate; "Fill a form from your own papers" (`fill-and-stop`) fills a form without submitting it. `decosa list <words>` searches all the tools.
- [tariff-coverage] The tariff classification tool ("Draft a tariff classification memo", `tariff-classification`) it covers CBP New York rulings in 35 headings of 11 chapters. Chapters: 39, 42, 61, 62, 63, 64, 73, 84, 85, 94, 95. Headings: 3923, 3924, 3926, 4202, 6104, 6109, 6110, 6114, 6204, 6211, 6307, 6402, 6403, 6404, 7323, 7326, 8414, 8419, 8471, 8479, 8481, 8504, 8516, 8517, 8518, 8528, 8543, 8544, 9401, 9403, 9405, 9503, 9504, 9505, 9506. A product outside them comes back as "out of coverage", naming the heading it would need; a licensed customs broker can classify it. Inside them, a description that lacks a deciding fact comes back as "needs a broker".
- [advice] Decosa gives no legal, medical, tax or financial advice. Tools prepare drafts and checks for a qualified professional.
- [family] The Decosa family (decosa.ai/family): separate apps from the same team. "Built on Decosa" means an app runs Decosa's engine code as a real part of what it does (not necessarily Decosa's hosting); "Built by Decosa" means Decosa, Inc. makes the app itself; "Joining Decosa" means its engine is coming over. Everdust (everdust.ai; Apple Vision Pro, iPhone, iPad, Mac and Apple TV; coming soon): Calm 3D worlds that hear music. Built by Decosa: Decosa, Inc. makes Everdust. Its scene Creator, in beta, builds scenes with open models that Decosa can serve. SottoMD (sottomd.com; Web; coming soon): A visit copilot for independent clinicians: a live transcript, the questions still worth asking, and a note checked against what was said. Built on Decosa: runs Decosa's clinical engine. Cue (github.com/sipratt-p/cue-engine; iPhone and Mac; in testing): Finds where each song starts in a long DJ mix, so you can skip song by song, from the lock screen or the car. Joining Decosa: its mix engine coming over. Snug (snug3d.com; iPhone and iPad (LiDAR); in testing): Scan a room with a LiDAR iPhone, then walk a measured, furnished 3D model and change it by asking. Joining Decosa: its room engine coming over. AutoTalent (autotalent.ai; Web): Give it a role: it searches widely, ranks the people it finds with the evidence for each, and writes the outreach. Joining Decosa: its signed decision record coming over. ElmoSEO (elmoseo.com; Web): Shows how ChatGPT, Google and Perplexity see a business, and a short list of what to fix first, in plain English. Joining Decosa: its site-check engine coming over. Dr. Grey (drgrey.ai; Web): Pick a goal and see the supplements with the strongest research, what the studies found, and a grade for each. Joining Decosa: its evidence engine coming over.

## Tools

- [Run the visit with a copilot](https://decosa.ai/clinics/visit-copilot) (`clinical`, live; home: Healthcare; runs: selfhost, crp-phi-research): The visit copilot: a live transcript with speaker labels, guidance on what is still worth asking (each item with its source), a note checked sentence by sentence against the visit, codes, prior-auth prep, and the work note, FMLA sections and patient instructions drafted, never signed or sent. JSON: https://decosa.ai/use-cases/clinical.json
- [Run an open model behind the OpenAI API](https://decosa.ai/tools/developer/code) (`code`, live; home: Software and AI ops; runs: hosted, selfhost): A coding agent on an open model that never sends your source to a third party. JSON: https://decosa.ai/use-cases/code.json
- [Make a story, a song or a music video](https://decosa.ai/studio) (`studio`, live; home: Creative and media; runs: hosted, selfhost): Make things with your people in them: a picture book read in your own voice, a song, a music video, an animatic or a dub. Add your cast once and every pack uses it. JSON: https://decosa.ai/use-cases/studio.json
- [Write the inspection report](https://decosa.ai/tools/operations/field) (`field`, live; home: Operations; runs: hosted, selfhost): Talk through an inspection or site walk. Get the checklist, issues by severity, measurements and a finished report. JSON: https://decosa.ai/use-cases/field.json
- [Translate a talk live](https://decosa.ai/tools/operations/translate) (`translate`, live; home: Operations; runs: hosted, selfhost): Speak in one language and read it in another, sentence by sentence, with a running glossary and a summary at the end. JSON: https://decosa.ai/use-cases/translate.json
- [Make a tamper-evident meeting record](https://decosa.ai/tools/developer/record) (`record`, live; home: Software and AI ops; runs: hosted, selfhost): Live captions now. Afterwards, a transcript with speakers and a cited summary, sealed in a signed record that anyone can re-check. JSON: https://decosa.ai/use-cases/record.json
- [Check your provider serves the model you pay for](https://decosa.ai/tools/developer/auditor) (`auditor`, live; home: Software and AI ops; runs: hosted, selfhost): Checks whether an OpenAI-compatible endpoint really serves the model it claims, at the quality it claims, and signs the result. JSON: https://decosa.ai/use-cases/auditor.json
- [Make a disclosed UGC ad](https://decosa.ai/studio/brand) (`ugc`, preview; home: Creative and media; runs: hosted, selfhost): Short UGC-style product videos with an original AI presenter. Labelled as AI ads on screen and in a content credential. Fake reviews and claims without evidence are refused. JSON: https://decosa.ai/use-cases/ugc.json
- [Catch AI answers your sources don't back](https://decosa.ai/tools/developer/grounding) (`grounding`, live; home: Software and AI ops; runs: hosted, selfhost, crp-deal-research): Checks every sentence of a draft against the sources you give it: supported, partial, unsupported or contradicted, with the source text it rests on. A claims gate can stop an email or page before it ships. JSON: https://decosa.ai/use-cases/grounding.json
- [Digest a deposition](https://decosa.ai/legal/deposition) (`deposition`, live; home: Legal; runs: selfhost, crp-legal-matter): A page:line digest of deposition or hearing transcripts. Every sentence cites its lines and is checked against them; with two or more transcripts it finds testimony that conflicts, with both cites. JSON: https://decosa.ai/use-cases/deposition.json
- [Check a brief before filing](https://decosa.ai/legal/filing-preflight) (`filing-preflight`, live; home: Legal; runs: hosted, selfhost, crp-legal-matter): One check before a brief is filed: every cited case, statute and rule exists, the cited opinion says what the brief claims, record cites resolve to the excerpts, quotations are verbatim, and no full SSN, birth date, account number or minor's name is left in. Ends in a signed record a partner can review. JSON: https://decosa.ai/use-cases/filing-preflight.json
- [Pre-check promo claims for MLR](https://decosa.ai/tools/life-sciences/promo-claims-check) (`promo-claims-check`, live; home: Life sciences; runs: hosted, selfhost): Checks every claim in a drug, device or supplement promotional piece against its label and references before MLR review: unsupported or off-label claims, missing fair balance, comparisons without head-to-head data, and supplement disease claims. Returns a signed pre-review packet. JSON: https://decosa.ai/use-cases/promo-claims-check.json
- [Check if an open model can take over your prompt](https://decosa.ai/tools/developer/migration-check) (`migration-check`, live; home: Software and AI ops; runs: hosted, selfhost): Paste a production prompt and the outputs your closed model already gave. The same inputs run on an open model; you get agreement per example with an interval, the failure clusters, cost per 1,000 requests, and a go or no-go in a signed record. JSON: https://decosa.ai/use-cases/migration-check.json
- [Classify with a confidence you can act on](https://decosa.ai/tools/developer/typed-judgment) (`typed-judgment`, live; home: Software and AI ops; runs: hosted, selfhost): Yes/no, one-of-N, score and label questions that come back as typed answers with a calibrated probability, a short reason and a receipt. It runs on pinned open weights, so a request can be re-run next quarter against the same model, and confident verdicts reproduce. An eval mode grades an output against a rubric or compares two. JSON: https://decosa.ai/use-cases/typed-judgment.json
- [Answer a security questionnaire](https://decosa.ai/tools/finance/security-questionnaire) (`security-questionnaire`, live; home: Finance and insurance; runs: hosted, selfhost): Fills a buyer's security questionnaire from answers you have already approved. The model chooses and may trim, but anything it changes is checked against the source, stale answers are caught against your current policies, and questions with nothing approved behind them are left for a person. JSON: https://decosa.ai/use-cases/security-questionnaire.json
- [Prove what your agent did](https://decosa.ai/tools/developer/flight-recorder) (`flight-recorder`, live; home: Software and AI ops; runs: hosted, selfhost): A signed, step-by-step record of what a browser or computer-use agent saw, decided and did. Replay it step by step; change one screenshot or one decision and verification names the step. JSON: https://decosa.ai/use-cases/flight-recorder.json
- [Run an end-to-end browser test](https://decosa.ai/tools/developer/test-runs) (`test-runs`, live; home: Software and AI ops; runs: hosted, selfhost): Write a test as plain-language steps with assertions. A browser agent carries out the steps; pass or fail comes only from the assertions, checked in code. Each run ends in a signed certificate you can keep as release evidence, and a failed assertion fails the build. JSON: https://decosa.ai/use-cases/test-runs.json
- [Build a model-risk evidence pack](https://decosa.ai/tools/finance/model-risk-pack) (`model-risk-pack`, live; home: Finance and insurance; runs: hosted, selfhost): Signed evidence that an LLM a bank, insurer or lender relies on still behaves as validated: identity probes, graded test cases, stability against the validated run, counterfactual fairness pairs and a model card, with a re-check before any alert. Validators can recompute every number from the record. JSON: https://decosa.ai/use-cases/model-risk-pack.json
- [Check an AI note](https://decosa.ai/clinics/clinical-ai-monitor) (`clinical-ai-monitor`, live; home: Healthcare; runs: selfhost, crp-phi-research): Checks any commercial AI scribe's note against the visit transcript: every sentence typed as not in the visit, a changed detail, the wrong speaker or an invented exam finding, with the transcript line, plus the medicines, allergies and plan items the note left out. Tightens the note without adding a word. Buyer mode scores vendors into a signed scorecard. JSON: https://decosa.ai/use-cases/clinical-ai-monitor.json
- [Build a privilege log](https://decosa.ai/legal/privilege-log) (`privilege-log`, live; home: Legal; runs: selfhost, crp-legal-matter): Typed privilege calls for each email or memo in a production set: attorney-client, work product, not privileged, or needs attorney review, with the lawyer involved, the reason checked against the document and a calibrated probability. It drafts log entries that describe without disclosing, catches the ones that leak, gives duplicates the same call, flags outsiders who could waive, and seals it all in a signed record. JSON: https://decosa.ai/use-cases/privilege-log.json
- [Log human edits and editor sign-off](https://decosa.ai/tools/media/editorial-ledger) (`editorial-ledger`, live; home: Media and marketing checks; runs: hosted, selfhost, crp-source-protection): Keeps the AI first draft, every human edit as a diff with edit metrics, and a named editor's sign-off in one signed record, with a public page per published piece. It is the evidence behind the EU AI Act's exception for AI-assisted text that had human review and editorial responsibility (Art. 50(4)). JSON: https://decosa.ai/use-cases/editorial-ledger.json
- [Check a police report against bodycam](https://decosa.ai/legal/report-integrity) (`report-integrity`, live; home: Legal; runs: selfhost): Checks an incident report (police, security, EMS, workplace) against the transcript of its recording, sentence by sentence, and lists key events the report leaves out, with times you can cue up. Or drafts the report from the recording, keeps a signed first draft, and records which sentences a person changed before signing. JSON: https://decosa.ai/use-cases/report-integrity.json
- [Redline from firm precedents](https://decosa.ai/legal/legal-drafting-editor) (`legal-drafting-editor`, live; home: Legal; runs: selfhost, crp-legal-matter): Applies a firm's playbook to a contract and returns a Word file with real tracked changes and a margin comment on each, naming the rule, the position and the precedent. New wording comes only from the firm's own precedent clauses: code traces every inserted word back to one, a grounding check and a second read of the rule confirm each edit, and reject-all gives back the original word for word. JSON: https://decosa.ai/use-cases/legal-drafting-editor.json
- [Redact a public-records release](https://decosa.ai/legal/foia-desk) (`foia-desk`, live; home: Legal; runs: selfhost): For a records officer answering a FOIA or state public-records request: turns the request into a search scope, finds the responsive records, and proposes each redaction with its exemption and a reason that names the harm without repeating the withheld words. The release PDF is built without the withheld text and read back to prove nothing under a box can be copied out. It drafts the index and the response letter, and seals every call and every officer decision in a signed record. JSON: https://decosa.ai/use-cases/foia-desk.json
- [Score an oral exam against a rubric](https://decosa.ai/tools/operations/oral-assessment) (`oral-assessment`, live; home: Operations; runs: hosted, selfhost): Vivas and structured interviews scored against a rubric. While it runs, the examiner sees which criteria still lack evidence and neutral follow-ups. Afterwards each criterion gets a draft score with the candidate's words cited and checked, the examiner confirms or overrides every score with a reason, and both steps are sealed in signed records that survive an appeal. Examiner side only. JSON: https://decosa.ai/use-cases/oral-assessment.json
- [Chart claim support in a patent spec](https://decosa.ai/legal/patent-claim-support) (`patent-claim-support`, live; home: Legal; runs: selfhost): Maps every element of every claim to the specification paragraphs that describe it, with a support strength and the model's reason, one receipted call per element. Code checks antecedent basis, claim references and claim terms the specification never uses. The output is a claim chart, an issue list and a signed record, for a registered practitioner to review before filing or in a response. JSON: https://decosa.ai/use-cases/patent-claim-support.json
- [Make a cleared song](https://decosa.ai/apps/music-gen-cleared) (`music-gen-cleared`, live; home: Creative and media; runs: hosted, selfhost): Music for ads, games and creators with a paper trail. A prompt guard refuses or strips named artists, songs, labels and voice imitation (code rules, then one receipted typed judgment), the track renders on MiniMax-Music3 or ACE-Step, and a CPU check scores it against a catalogue of openly licensed recordings. Each track ships with a C2PA credential and a signed certificate that states the model, seed, checks, similarity score and the exact licence. JSON: https://decosa.ai/use-cases/music-gen-cleared.json
- [Pre-check samples and lyrics](https://decosa.ai/tools/media/sample-clearance) (`sample-clearance`, live; home: Media and marketing checks; runs: hosted, selfhost): Checks a track before release against your own catalog: samples (even pitched, stretched, filtered or looped), re-played melodies and lifted lyric lines, each with its time in the track and in the reference. It compares what it finds with what the artist declared and marks each item clear, replace or remove, with a signed record. Triage for clearance staff, not legal advice. JSON: https://decosa.ai/use-cases/sample-clearance.json
- [Check split sheets and metadata](https://decosa.ai/tools/media/split-sheet-check) (`split-sheet-check`, live; home: Media and marketing checks; runs: hosted, selfhost): Reconciles split sheets, release metadata (DDEX ERN or a distributor CSV), contract excerpts and PRO or MLC registration exports. Code checks every ISRC, ISWC, IPI and UPC check digit and adds up every split; an open model reads the messy documents, and every value it reads must be found in the document. You get each discrepancy with the lines it came from, a proposed credit sheet to confirm, and a signed record. JSON: https://decosa.ai/use-cases/split-sheet-check.json
- [Check a paper's citations](https://decosa.ai/tools/life-sciences/paper-claim-check) (`paper-claim-check`, live; home: Life sciences; runs: hosted, selfhost): Checks every citing sentence of a manuscript against the cited paper's open full text and quotes the passage: supported, overstated, not supported or contradicted, one receipted call per citation. Each reference is looked up in Crossref and OpenAlex for retractions, corrections, wrong DOIs and years, and self-citation stacks are flagged. The output is an issue list and a signed report for an editor, reviewer or integrity office to act on. JSON: https://decosa.ai/use-cases/paper-claim-check.json
- [Keep a signed lab notebook](https://decosa.ai/tools/life-sciences/signed-lab-notebook) (`signed-lab-notebook`, live; home: Life sciences; runs: hosted, selfhost): A tamper-evident lab notebook with an AI-analysis log. Entries hold text and instrument files by SHA-256, amendments point at the original and give a reason, and witnesses and approvers sign with their own keys. When a model analyses data, the prompt, the exact data files, the model and its signed receipt go on the record, so a reviewer sees which conclusions were AI-assisted. RFC 3161 timestamps from an independent authority, an audit-trail export and a verify page anyone can run. JSON: https://decosa.ai/use-cases/signed-lab-notebook.json
- [Check green claims in copy](https://decosa.ai/tools/finance/green-claims-check) (`green-claims-check`, live; home: Finance and insurance; runs: hosted, selfhost): Checks every environmental claim in pack copy, a product page or an ad script against the EU bans that apply from 27 September 2026 (Directive (EU) 2024/825) or the UK CMA Green Claims Code: generic claims, offset-based climate neutral, own-brand labels, whole-product claims, legal requirements sold as features and future targets without a plan. Each claim gets a verdict with the rule cited, the evidence-file line that backs it, and a rewrite; you get a CSV claim table and a signed report. A triage, not a compliance sign-off. JSON: https://decosa.ai/use-cases/green-claims-check.json
- [Record F&I add-on disclosures](https://decosa.ai/tools/finance/fi-disclosure-record) (`fi-disclosure-record`, live; home: Finance and insurance; runs: selfhost): Checks a recorded F&I-office conversation against its deal jacket under California's CARS Act (SB 766, in force 1 Oct 2026). For each add-on it asks whether the add-on was said to be optional and whether it was presented as required, with a quote and a time. It flags add-ons that were declined or never discussed but are on the contract, and prices or payments that differ. It reads the written disclosures, the 3-day cancel right and banned add-ons from the jacket, and seals it all in a signed record to keep for two years. JSON: https://decosa.ai/use-cases/fi-disclosure-record.json
- [Dub a video in your own voice](https://decosa.ai/apps/consented-dubbing) (`consented-dubbing`, live; home: Creative and media; runs: hosted, selfhost): A Spanish track for a creator's video in the creator's own consented voice, or in a consented dubber's. The consent ledger is checked when the job starts, before the voice is rendered and again at approval. The track is cut to the video's exact length, the original track is never replaced, and subtitles get a QA pass. Nothing is published until a person approves. Approval seals a signed record and a C2PA label: "AI-dubbed, voice consented by" the named person. JSON: https://decosa.ai/use-cases/consented-dubbing.json
- [Pre-flight a synthetic-performer ad](https://decosa.ai/tools/media/disclosure-preflight) (`disclosure-preflight`, live; home: Media and marketing checks; runs: hosted, selfhost): Checks an ad or short before it ships. It finds synthetic performers from the file's own provenance (its C2PA credential, Decosa's render receipt or watermark) and what you declare, and reads the frames to see whether an AI label is on screen. It burns in the label and signs a C2PA marking when the New York or EU rules need them, checks each real performer's consent by ledger id, and flags real people, other brands, profanity, rating triggers and music cues in the script. A named reviewer signs off, and the sign-off is sealed into an AI-use attestation. Advisory, not a legal clearance. JSON: https://decosa.ai/use-cases/disclosure-preflight.json
- [Turn a script into an animatic](https://decosa.ai/apps/animatic-studio) (`animatic-studio`, live; home: Creative and media; runs: hosted, selfhost): Paste a short script or ad brief and get a timed animatic. Qwen3.8-27B breaks it into shots that each cite their script lines; code checks that every line is covered and inserts the dialogue word for word, and the grounding check flags shots that add events the script does not have. Frames come from Wan2.2-VACE-Fun-A14B, optional motion from the same model, temp dialogue from stock voices the consent ledger allows. Edit any shot and re-render just that one. The MP4 carries shot numbers, timecode, a C2PA credential and a signed record of which steps were AI and which were your edits. Previsualisation that saves a storyboard artist's first pass, not a replacement for one. JSON: https://decosa.ai/use-cases/animatic-studio.json
- [Produce an audio drama](https://decosa.ai/apps/audio-drama-studio) (`audio-drama-studio`, live; home: Creative and media; runs: hosted, selfhost): Paste a radio script or a prose chapter and review the parse: every speaker, line, direction and sound cue, before anything is voiced. Cast each role from consented house voices; every line passes the consent ledger right before it is spoken. Music comes from cleared cues with licence certificates, and sound from CC0 or public-domain recordings or code. The mix is mastered to a podcast or ACX-style audiobook spec, with chapters, captions, show notes with an AI disclosure, a C2PA credential, a signed record and sides for human actors. JSON: https://decosa.ai/use-cases/audio-drama-studio.json
- [Make a music video for your track](https://decosa.ai/apps/music-video-studio) (`music-video-studio`, live; home: Creative and media; runs: hosted, selfhost): A first music video for an independent artist's own track. It finds the beats, bars and sections, times the artist's lyrics to the audio, runs the sample-clearance pre-check, and has an open model write scenes that each cite the lines they illustrate; code and a receipted yes/no check test every citation. Clips render with an open video model, cut on the bar lines, with the lyrics as karaoke captions, in 16:9 and 9:16. Each file carries a C2PA credential and a burned-in AI label, and a signed record holds the rights statement, the checks and the measured cut timing. The visuals are 480p upscaled: a stylised first video, not a studio shoot. JSON: https://decosa.ai/use-cases/music-video-studio.json
- [Write a SAR narrative](https://decosa.ai/tools/finance/sar-narrative-desk) (`sar-narrative-desk`, live; home: Finance and insurance; runs: selfhost): Drafts a suspicious-activity narrative from an alert's case file (a transaction ledger, the customer profile and the investigator's notes) in the structure FinCEN asks for: who, what, when, where, why and how. Every sentence cites the rows, fields or notes it rests on. Every amount, date, count and total is recomputed in code from the cited rows, and a wrong one is held back with the right figure beside it. It also writes a signed no-file rationale for alerts closed without a SAR, or checks a draft you wrote. A drafting aid: an investigator decides and files. JSON: https://decosa.ai/use-cases/sar-narrative-desk.json
- [Review a claim file for conduct](https://decosa.ai/tools/finance/claims-conduct-pack) (`claims-conduct-pack`, live; home: Finance and insurance; runs: selfhost): Reads a claim file and its policy the way a market-conduct examiner would, under the NAIC model, California (10 CCR 2695) or Texas (Ins. Code 542). It computes the acknowledgement, decision, status-letter and payment deadlines from the file's own dates, checks each denial reason for a cited provision and grounds it in the policy text, and looks for the Department of Insurance review notice. It gives typed judgments with quotes on misrepresentation, low offers and missing investigation, records any AI use, and seals it all in a signed file review a reviewer can sign off. JSON: https://decosa.ai/use-cases/claims-conduct-pack.json
- [Review a collections call](https://decosa.ai/tools/finance/collections-call-qa) (`collections-call-qa`, live; home: Finance and insurance; runs: selfhost): QA for recorded debt-collection and mortgage-servicing calls under Regulation F and Regulation X. On the call it answers yes, no or unclear, with a quote and a time: debt-collector disclosure, identity before the debt, third-party disclosure, threats, a stop request or dispute brushed aside, and loss-mitigation options. From the call log, in code: the 7-in-7 call count, calling hours in every time zone on file, calls after a stop request, and live contact by day 36. Signed record kept with the recording for three years. JSON: https://decosa.ai/use-cases/collections-call-qa.json
- [Draft breach notices and deadlines](https://decosa.ai/tools/finance/incident-notification-pack) (`incident-notification-pack`, live; home: Finance and insurance; runs: selfhost): For the CISO and general counsel during a cyber incident. From the team's incident log it builds a signed, hash-chained timeline and a deadline clock per report (SEC 8-K Item 1.05, NIS2, DORA, the Cyber Resilience Act, California), computed in code from the entries you tag. It drafts each notice or checks yours sentence by sentence against the log entries it cites: times and counts in code, other facts by a grounding judge. It lists the required elements each notice lacks and the facts two notices state differently. Counsel decides and files. JSON: https://decosa.ai/use-cases/incident-notification-pack.json
- [Tie out MD&A figures](https://decosa.ai/tools/finance/filing-tieout) (`filing-tieout`, live; home: Finance and insurance; runs: selfhost): Ties every figure in a 10-K or 10-Q's MD&A to the filing's tagged statements and notes, in code: the cell it equals, or the change, percentage change or margin it is computed from. A figure that does not tie is flagged with the cell it was compared with: a different value, last year's figure, million for billion, the other direction, a note reference to the wrong note, or a note table that disagrees with a statement. Figures the tables do not show are left for a person. You get a workpaper and a signed record. A preparer's and reviewer's aid, not an audit. JSON: https://decosa.ai/use-cases/filing-tieout.json
- [Disposition a sanctions alert](https://decosa.ai/tools/finance/sanctions-disposition-record) (`sanctions-disposition-record`, live; home: Finance and insurance; runs: selfhost): Clears one sanctions screening alert at a time, for the analyst. Code compares the customer with the OFAC, EU or UK list entry field by field: name across transliterations and name order, date of birth, nationality, ID numbers, IMO number, address and the kind of party. It proposes true match, false positive or needs more information by a named rule. An open model gives an independent reading and writes a rationale, and every sentence of it is checked against the fields it cites. The analyst decides, and the decision is sealed in a signed, hash-chained record that pins the list version, for a 10-year log you keep. JSON: https://decosa.ai/use-cases/sanctions-disposition-record.json
- [Review a Medicare sales call](https://decosa.ai/clinics/medicare-call-record) (`medicare-call-record`, live; home: Healthcare; runs: selfhost): QA and a signed record for recorded Medicare Advantage and Part D sales calls. On the call it answers yes, no or unclear, with a quote and a time: the TPMO disclaimer and whether it came before benefits were discussed, the Scope of Appointment, "free" for a $0 premium, claiming to be Medicare, gifts, pressure, and the steps of an enrollment. Every benefit the agent states is checked against the plan's Summary of Benefits. Signed record with the retention dates of the 1 Jun 2026 rule. JSON: https://decosa.ai/use-cases/medicare-call-record.json
- [Appeal a denial](https://decosa.ai/clinics/denial-appeal-packet) (`denial-appeal-packet`, live; home: Healthcare; runs: selfhost): For billing, revenue-cycle and denials staff, and patient advocates. Give it the denial (EOB, remittance or letter with its CARC/RARC codes), the chart excerpts and the payer's coverage criteria. It computes the next appeal level's deadline with the federal rule and the arithmetic (ERISA and ACA, Medicare, Medicare Advantage, Medicaid managed care) and marks each criterion met, not met or undocumented, with a quote from the chart. It drafts a letter only when the chart supports every criterion, and checks every sentence against the chart and the policy. When the chart doesn't support it, it says don't appeal, or what to get first. The packet is signed. JSON: https://decosa.ai/use-cases/denial-appeal-packet.json
- [Check if a video is made for kids](https://decosa.ai/tools/media/kids-content-preflight) (`kids-content-preflight`, live; home: Media and marketing checks; runs: hosted, selfhost): Before a video, short or app listing goes out, checks whether it is directed to children. Each of the FTC's COPPA factors (subject matter, characters, music, the age of the people on screen, language, ads) gets its own typed answer and probability, with the lines that show it. The call is compared with your made-for-kids setting: kids-directed but set not-for-kids by a channel default is the pattern in the FTC's $10M Disney order. When it is kids-directed, it lists what has to change: comments, personalised ads, links, requests for children's details, undisclosed paid promotion, AI labels, and for apps purchases, third-party ads and analytics. A named reviewer makes the designation, sealed in a signed review record. A back-catalogue audit checks many uploads at once. Not a COPPA compliance determination. JSON: https://decosa.ai/use-cases/kids-content-preflight.json
- [Check HCC codes for MEAT evidence](https://decosa.ai/clinics/hcc-evidence-file) (`hcc-evidence-file`, live; home: Healthcare; runs: selfhost): For Medicare Advantage risk-adjustment and compliance teams. Send a member's submitted diagnosis codes and the notes for the year. For each code that maps to a CMS-HCC V28 payment HCC it says supported, not supported or insufficient, quotes the sentences that show the condition was monitored, evaluated, assessed or treated (MEAT) or says none do, and checks each record the way a RADV reviewer would: face to face, signed with an acceptable credential, in the year, no audio-only calls, radiology reports or late addenda. Deletions come first. It never suggests codes to add. Out comes the net effect, an evidence file and a signed record. A review aid: certified coders decide. JSON: https://decosa.ai/use-cases/hcc-evidence-file.json
- [Draft VEX for scanner findings](https://decosa.ai/tools/developer/vex-triage) (`vex-triage`, live; home: Software and AI ops; runs: hosted, selfhost): For product security teams and software vendors. Give it a Grype or Trivy report for a container image, its SBOM, and an evidence bundle from a small collector that runs next to the image. For each CVE, code checks the package, the version against the fix and the advisory ranges, the distro changelog, what actually loads the library, the settings the advisory names, and the platform. An open model then picks a VEX status with quoted evidence. Not affected needs positive evidence; everything else stays affected, fixed or under investigation. You get signed OpenVEX and CycloneDX VEX drafts and a signed evidence record for a security engineer to sign off. JSON: https://decosa.ai/use-cases/vex-triage.json
- [Triage a device complaint for MDR](https://decosa.ai/tools/life-sciences/device-mdr-triage) (`device-mdr-triage`, live; home: Life sciences; runs: selfhost): For medical device regulatory and quality teams. Give it a complaint and the product's labeling. It answers the 21 CFR 803.50(a) questions one by one, each with the words from the complaint that decide it: death or serious injury, could the device have caused or contributed, a malfunction, likely to cause serious harm if it recurred. It suggests reportable, not reportable or needs investigation, and starts the 30-day or 5-work-day clock from the day any employee heard, which the complaint itself may name. It checks the complaint record against 820.35(a), drafts the 3500A event description with every sentence checked, and signs a triage record and the named person's decision. A trend view groups complaints by failure mode. A triage aid: a person decides. JSON: https://decosa.ai/use-cases/device-mdr-triage.json
- [Map evidence to NIST 800-171](https://decosa.ai/tools/finance/cmmc-evidence-map) (`cmmc-evidence-map`, live; home: Finance and insurance; runs: selfhost): For small defence contractors and the consultants who prepare them for CMMC Level 2. Send the SSP's statements and the evidence behind them: policies, config exports, logs, records, scans, screenshots and verified test runs. For each NIST SP 800-171 Rev 2 requirement and each of its NIST SP 800-171A objectives it says evidence present, partial or missing, quotes the lines, and flags artefacts that cannot count: drafts, stale exports, systems outside the scope. Gaps come first. Out come a gap map, a draft POA&M marking what may never go on one, and a signed evidence index an assessor can check. An evidence-gap map, never a certification or a score. JSON: https://decosa.ai/use-cases/cmmc-evidence-map.json
- [Check a lay summary's numbers](https://decosa.ai/tools/life-sciences/eu-trial-lay-summary) (`eu-trial-lay-summary`, live; home: Life sciences; runs: hosted, selfhost): For sponsors' clinical disclosure teams, medical writers and CROs. Give it a trial with results on ClinicalTrials.gov and it drafts the lay summary the EU Clinical Trials Regulation requires (Article 37(4), content in Annex V), or checks the one you wrote. Every number is traced in code to a results cell, or to arithmetic on the cells the sentence cites, and a wrong one is flagged with the right figure. Comparisons are checked against the table, a result that could be chance has to say so, serious side effects and deaths must be given for every group, and promotional or softening words are flagged. Each claim goes through a grounding judge; Annex V's ten elements and the reading grade are reported. Out come a review file and a signed record of every check. It never says a summary is compliant: the sponsor's medical writer approves. JSON: https://decosa.ai/use-cases/eu-trial-lay-summary.json
- [Investigate a Reg E dispute](https://decosa.ai/tools/finance/reg-e-dispute-file) (`reg-e-dispute-file`, live; home: Finance and insurance; runs: selfhost): For bank, credit union and fintech disputes teams, and the compliance officers who answer for their files. Give it the consumer's notice of an electronic fund transfer error (a chat, call transcript, form or letter), the account's transactions as CSV, and the investigation notes, events and results letter. It computes every Regulation E clock in code with the rule cited: notice within 60 days of the statement, 10 or 20 business days, provisional credit and its notice, 45 or 90 days, the report within 3 business days and the correction within 1. It quotes the notice's required elements, lists the file's gaps, and checks that a no-error letter explains the findings for this case and notes the right to request the documents relied on. It never decides a claim: a named investigator does, and the signed record keeps their decision as entered. JSON: https://decosa.ai/use-cases/reg-e-dispute-file.json
- [Stage a listing photo with disclosure](https://decosa.ai/tools/operations/virtual-staging) (`virtual-staging`, live; home: Operations; runs: hosted, selfhost): Stages an empty-room listing photo with open image models, adding movable furniture only, then checks that the walls, windows, doors, views, floors and visible damage were left alone, pixel by pixel and with a side-by-side model comparison. Requests to remove power lines, hide cracks or make a room look bigger are refused. It also checks images staged anywhere else. A photo that passes gets a burned-in 'Virtually staged' label linking to the unaltered original on a public page, a C2PA credential naming the original as its parent, and a signed pair record: the disclosure California's AB 723 has asked for since 1 Jan 2026. Not legal advice. JSON: https://decosa.ai/use-cases/virtual-staging.json
- [Write a due-diligence red-flag memo](https://decosa.ai/legal/ma-dd-redflags) (`ma-dd-redflags`, live; home: Legal; runs: selfhost): For deal counsel and corporate development teams. Give it a data room: contracts, board minutes, the disclosure schedule, revenue by customer, financial summaries. For ten red-flag categories (change of control, anti-assignment, exclusivity and non-compete, MFN pricing, uncapped liability, debt covenants, IP assignment gaps, litigation missing from the disclosure schedule, customer concentration, key person), a reranked search pulls the passages that could hold one and an open model says which do, quoting the exact words. Code checks every quote word for word against the passage and byte span it cites, and a grounding check reads each finding against that passage. You get a red-flag memo with what to ask the seller, and a signed record. A triage aid for counsel, not legal advice. JSON: https://decosa.ai/use-cases/ma-dd-redflags.json
- [Draft a tariff classification memo](https://decosa.ai/tools/finance/tariff-classification) (`tariff-classification`, live; home: Finance and insurance; runs: hosted, selfhost): For customs brokers and importers' trade-compliance teams. Describe a product. It searches public CBP rulings on similar products and the HTS text of the candidate headings, and an open model proposes a heading, subheading and statistical number with the General Rules of Interpretation it applied. Code checks that every code is printed in the current HTS release and that every quote from a ruling is found word for word in the text it retrieved. When the description leaves out a fact the code turns on, or the evidence is thin, it says so and sends it to a broker. You get a memo and a signed record naming the ruling set and the index it searched. A research aid: a licensed broker decides. JSON: https://decosa.ai/use-cases/tariff-classification.json
- [Check CSR numbers against tables](https://decosa.ai/tools/life-sciences/csr-number-verifier) (`csr-number-verifier`, live; home: Life sciences; runs: selfhost): For medical writing and regulatory operations teams at sponsors and CROs. Give it a clinical study report: the narrative of sections 10 to 12 with its in-text tables and TLFs, as text and rows or as a PDF (born-digital or scanned, read by the document reader). Every number in the narrative is traced to the table cell it reports: an open model points each number at the cell for the right arm, row and timepoint, and code compares the value at the printed decimals and recomputes percentages, differences, totals and relative reductions. It flags numbers that don't match their cell, numbers with no source, and derived numbers that don't compute, with the likely slip (the other arm's value, a stale value from an earlier data cut, transposed digits, rounding, a wrong denominator). In-text tables are checked against their source TLF. You get a QC report and a signed record. A medical-writing QC aid: the medical writer and the QC reviewer decide. JSON: https://decosa.ai/use-cases/csr-number-verifier.json
- [Build an EU GPSR listing](https://decosa.ai/tools/finance/gpsr-listing-pack) (`gpsr-listing-pack`, live; home: Finance and insurance; runs: hosted, selfhost): Turns a supplier's product sheet and the product label (text or a photo) into the listing information Article 19 of the EU General Product Safety Regulation (EU) 2023/988 asks every online offer to show: manufacturer, EU responsible person when the maker is outside the Union, product identification, and the warnings in the buyer's language. Every value is looked up in the documents, each Article 19 element is marked found or missing, a label that disagrees with the sheet is flagged, and the warnings come back in up to five languages with every number, unit, negation and name checked. A preparation aid with a signed record, never a statement that an offer meets the Regulation. JSON: https://decosa.ai/use-cases/gpsr-listing-pack.json
- [Build a medical chronology](https://decosa.ai/legal/medical-chronology) (`medical-chronology`, live; home: Legal; runs: selfhost): For personal-injury and med-mal attorneys, life and disability insurers and APS reviewers. Give it a packet of medical records: typed notes, scanned and faxed forms, handwriting, lab tables. The document reader reads every page with boxes, the pages go into a retrieval index as chunks that keep their page and box, and an open model lists the encounters, diagnoses, procedures, imaging, labs and medications on each page. Code keeps only events whose words are on the page and whose date is written there. Copies from different providers are merged, conflicting dates for the same procedure are flagged, and so are gaps in treatment and conditions documented before the injury. Every line cites the file, page and box. A paralegal and reviewer aid, not medical or legal advice. JSON: https://decosa.ai/use-cases/medical-chronology.json
- [Screen a claim for Federal IDR](https://decosa.ai/clinics/nsa-idr-packet) (`nsa-idr-packet`, live; home: Healthcare; runs: selfhost): For out-of-network emergency, anesthesia, radiology and pathology groups, their billing companies, and the health plans on the other side. Give it the EOB or remittance, the open negotiation notice and the supporting documents. It reads the claim facts with quotes, then screens the claim for Federal IDR: coverage, qualified service, state law, notice and consent, the 30-business-day open negotiation start, the 4-business-day initiation window, cooling-off and batching, each check with its dated rule and the business-day arithmetic. When nothing fails, it states the offer as a dollar amount and a percentage of the QPA and drafts a brief from facts it can quote; billed charges, usual and customary charges and Medicare rates are left out. A screen-only mode makes one model call. The packet is signed. A preparation aid: a person files, and a certified IDR entity decides. JSON: https://decosa.ai/use-cases/nsa-idr-packet.json
- [Quote a job from a walkthrough video](https://decosa.ai/tools/operations/walkthrough-to-quote) (`walkthrough-to-quote`, live; home: Operations; runs: hosted, selfhost): For painters, flooring, remodel, restoration, landscaping and moving crews. The customer or estimator walks the job with a phone video and talks. It drafts a line-item scope: rooms and areas, quantities with how each was estimated, materials mentioned, and damage, hazards and access problems seen. Every line cites a time range and a keyframe and is marked seen, said or assumed. Quantities from the video are ranges; it never invents a dimension and says what to measure first. Prices come only from your own price sheet, the sums are done in code and signed. A draft for the estimator: you set the final price. JSON: https://decosa.ai/use-cases/walkthrough-to-quote.json
- [Turn an expert video into an SOP](https://decosa.ai/tools/operations/expert-to-sop) (`expert-to-sop`, live; home: Operations; runs: hosted, selfhost): Turns a screen or bench recording of an expert doing a procedure into a draft SOP. Every step cites a time range and a keyframe from the recording and is checked against what the expert said. Steps that were said but never shown are marked needs confirmation, steps shown but never said are marked, and a named reviewer signs each revision into a record anyone can verify. A draft for a person to approve, not a certified procedure. JSON: https://decosa.ai/use-cases/expert-to-sop.json
- [Check generated product images](https://decosa.ai/tools/media/honest-product-imagery) (`honest-product-imagery`, preview; home: Media and marketing checks; runs: hosted, selfhost): For e-commerce sellers and brands that make product and lifestyle images with AI. Send photos of the real product and the AI image: it checks the net quantity, colour, pack text and claims, logo, warnings, and features or extra units the product does not have, with deterministic measurements (colour difference in Lab, the document reader for the pack text, logo and unit matching) plus a side-by-side open-model comparison. Any person in the image must be a consent-ledger identity whose consent covers this advertising use. An image that passes gets the AI metadata and label each marketplace or law you pick asks for, a C2PA credential and a signed record. It checks images made anywhere; it does not generate them. Not legal advice. JSON: https://decosa.ai/use-cases/honest-product-imagery.json
- [Draft an adverse-event case](https://decosa.ai/tools/life-sciences/pv-intake) (`pv-intake`, live; home: Life sciences; runs: selfhost): For drug safety teams at marketing authorisation holders and the CROs that run case intake for them. Send an adverse event report as it arrived: an email, a web form, a call to the medical information line, or a scanned MedWatch or CIOMS form, in any of nine languages. It drafts an E2B(R3)-shaped case with every field quoted to its source line (and the reporter's own words for a translated report), checks the four minimum criteria, marks the seriousness criteria per event, checks expectedness against the reference label you send, and works out day 0, which the report itself may move earlier, and the US 15-day and EU 15- and 90-day dates. Coding suggestions are free text, not MedDRA. Anything missing comes back as follow-up questions in the reporter's language. The safety physician decides and signs. JSON: https://decosa.ai/use-cases/pv-intake.json
- [Compare safety sections across labels](https://decosa.ai/tools/life-sciences/label-consistency-check) (`label-consistency-check`, live; home: Life sciences; runs: hosted, selfhost): For regulatory affairs and labelling teams. Give it the label documents for one medicine: the company core data sheet (CCDS), the US Prescribing Information, the EU SmPC and package leaflet in several languages, and the carton text or a scan of the carton. It aligns the sections (indications, dosing, contraindications, warnings, adverse reactions, storage, strengths) and flags every difference with a quote from each document: a dose, interval or strength that disagrees, a warning or contraindication that is missing, a different storage temperature, a translation that dropped a negation or changed a number, an SmPC heading that is not the QRD template's. Each flag is triaged as a likely error, likely deliberate (a regional convention such as US Fahrenheit storage wording, or a line in your deviation log) or unclear. A labelling QC aid with a signed record; regulatory affairs decides. JSON: https://decosa.ai/use-cases/label-consistency-check.json
- [Find accessibility fixes for a shop](https://decosa.ai/tools/operations/storefront-accessibility-pass) (`storefront-accessibility-pass`, live; home: Operations; runs: hosted, selfhost): For online shops selling into the EU or the US, and the agencies that build them. A private headless browser opens the shop's pages, runs axe-core's automated WCAG 2.2 A/AA rules, tries add-to-cart and the checkout with the keyboard alone (reachable, answers Enter, focus visible, no trap), and asks an open model the questions a rule engine cannot answer: is this alt text right for this photo, is "More" a clear link name here, does "Invalid input" say how to fix the field, is the offer only inside a banner image. You get a fix list, purchase blockers first, each with its WCAG success criterion and a code fix, and a signed record of what was checked. An audit aid, not a certification; not legal advice. JSON: https://decosa.ai/use-cases/storefront-accessibility-pass.json
- [Fill a form from your own papers](https://decosa.ai/tools/operations/fill-and-stop) (`fill-and-stop`, preview; home: Operations; runs: hosted, selfhost): For anyone re-typing the same facts into a long form: a claim, a benefits renewal, a W-9, a school form. Give it the form and your papers. A PDF form (fillable, flat or scanned) comes back filled, with a source sheet showing the line each answer came from and why each blank is blank. For a form on a website it never touches the site (none of the insurer, benefits, school or payer portals we checked allows automated tools): you get a copy list in the site's own order, each answer with its source and a copy button. It never signs, dates your signature, ticks a statement for you or fills ID, bank or card numbers, and you can review in your language. JSON: https://decosa.ai/use-cases/fill-and-stop.json
- [Turn a client call into notes](https://decosa.ai/legal/privileged-call-notes) (`privileged-call-notes`, live; home: Legal; runs: hosted, selfhost): For solo and small-firm lawyers who take client calls all day. After the call you get an intake memo where every line cites the transcript and is checked against it, a conflicts-check list of every name heard, a time entry in 0.1-hour units with a billing narrative, follow-up tasks and an email draft in the client's language. No notetaker bot joins the call: open models run on the server or your own box, the audio is deleted once the transcript is made, and the session ends in a signed record. JSON: https://decosa.ai/use-cases/privileged-call-notes.json
- [Check the other side's brief](https://decosa.ai/legal/check-their-brief) (`check-their-brief`, live; home: Legal; runs: hosted, selfhost): For the litigator answering an opposition: upload the other side's filing and get findings for your reply. Every cited case, statute and rule is looked up in public sources (a case none of them has is "not found in the sources searched", with the search trail, never "fake"), quotations are matched word for word against the opinion, holdings are checked against the opinion's text, and the file is scanned for text the judge cannot see (white, tiny or off-page text, invisible characters, metadata, comments) and for instructions aimed at AI tools. Each finding cites the page of their filing and the source, in a Word memo and a signed record. JSON: https://decosa.ai/use-cases/check-their-brief.json
- [Check their discovery responses](https://decosa.ai/legal/discovery-deficiency) (`discovery-deficiency`, live; home: Legal; runs: hosted, selfhost): For litigation associates and paralegals on either side of written discovery. Paste or upload the other side's interrogatory answers or responses to requests for production or admission. It flags objections with no specifics, incorporated general objections, responses that don't say whether anything is being withheld, promises to produce with no date, evasive or incomplete answers, admission responses that neither admit nor deny, privilege claims with no log, a missing verification or signature, and late service, with the deadline worked out. Each flag quotes the response and cites the rule text (federal courts, California, Texas). Then it turns the flags into a first-draft deficiency letter with the rule for each point: a starting point to rewrite, not a letter ready to send. Run it on your own draft responses before you serve them: it suggests fixes and never removes an objection. JSON: https://decosa.ai/use-cases/discovery-deficiency.json
- [Answer a payer audit](https://decosa.ai/clinics/payer-audit) (`payer-audit`, live; home: Healthcare; runs: selfhost): For the owner or manager of a small practice, or its billing service, when a payer or its review contractor asks for records or audits paid claims. Give it the auditor's letter and the charts. It works out the respond-by date from the letter in code, reads the claim list, matches each claim to its note, and checks every documentation requirement of the payer's published policy: found, with the chart, page and line, or missing. Weak claims are listed first, as plainly as the supported ones. It drafts a cover letter whose sentences are each checked against the sources, builds an indexed, bookmarked response packet, and signs a record of what was checked. It never suggests changing a record, and it never touches a payer portal. The same check runs as a self-audit before a payer asks. JSON: https://decosa.ai/use-cases/payer-audit.json
- [Check a prior auth before you send it](https://decosa.ai/clinics/prior-auth-check) (`prior-auth-check`, preview; home: Healthcare; runs: selfhost): For prior-authorization coordinators and medical assistants, before a request goes out. Give it the payer's own published medical policy (pasted, or the PDF) and the chart. It finds the criteria section for the service, then marks each criterion met with a chart quote, missing with the record to get, or not met with the line that contradicts it, and says plainly: ready to send, get these records first, or not supported on this chart. It answers the payer's form questions with chart quotes for you to copy into the portal yourself, writes a one-page peer-to-peer brief, and drafts a letter of medical necessity only when every criterion is met, each sentence checked against the chart and the policy. Licensed criteria (InterQual, MCG) are never reproduced: a policy that points to them gets a plain "can't check". It never suggests codes or changes to a note, and never touches a payer portal. JSON: https://decosa.ai/use-cases/prior-auth-check.json
- [Write a note from my jottings](https://decosa.ai/clinics/jottings-note) (`jottings-note`, live; home: Healthcare; runs: selfhost): For therapists in private practice with a notes backlog. After the session, give it what you already jot: typed bullets, a photo of your handwritten notes, or a dictation of up to a minute. Out comes a DAP, SOAP or BIRP progress note tied to your treatment-plan goals, where every sentence links to the jotting it came from. Nothing is added: numbers, diagnoses, moods, risk and recommendations must be in your jottings or the sentence is removed. A completeness check lists what payers audit (start and stop times, modality, interventions, goals, response, plan, your credential, and risk when you mentioned it) and marks what is missing instead of guessing. Catch-up mode drafts a backlog in one go. There is no recording, ever. The note is a draft you review, edit and sign. JSON: https://decosa.ai/use-cases/jottings-note.json
- [Capture audit evidence from your admin screens](https://decosa.ai/tools/finance/evidence-runner) (`evidence-runner`, preview; home: Finance and insurance; runs: selfhost): For the compliance lead who screenshots 30 to 150 admin settings screens every quarter for SOC 2 or CMMC. Name the screens once (2-step verification, password policy, audit logging, your own product's admin pages). The agent finds each one, read-only, and code reads the settings you named; you approve those values as the baseline. Every quarter after that it runs with no model: it opens each screen from its saved route, re-reads every setting, captures it on a fresh load stamped with the URL, the time and who is signed in, and lists what changed since the baseline, all in a signed certificate. The browser session blocks every write, so it cannot change a setting. It runs on your side, in your own signed-in browser or inside your network. JSON: https://decosa.ai/use-cases/evidence-runner.json
- [Put the visit into your EHR as drafts](https://decosa.ai/clinics/ehr-drafts) (`ehr-drafts`, preview; home: Healthcare; runs: selfhost): For clinicians in small practices who finish every visit by re-typing it into the EHR. Take what the visit produced (diagnosis codes, lab and imaging orders, medicine changes, the note) and an agent in your own browser enters it into the chart you have open as drafts: problems marked unconfirmed, prescriptions and orders saved but not sent, the note unsigned. It checks the chart is the visit's patient (name and date of birth read off the screen) before anything and again before every save, reads each medicine against the transcript and holds any that differ, never types a controlled substance, and stops before sign. Signing, transmitting, sending and e-prescribing are held in the browser's network layer whatever the button says. You get a checklist of what went where with the transcript line behind each value, and a copy list for what it could not enter. Where an EHR's terms forbid tools in your session, you get the copy list only. JSON: https://decosa.ai/use-cases/ehr-drafts.json
- [Read a demand before the deadline](https://decosa.ai/tools/insurance/demand-reader) (`demand-reader`, preview; home: Insurance; runs: selfhost): For bodily-injury adjusters at carriers, TPAs and self-insureds. Give it an injury demand package: the lawyer's letter, the medical bills and the records. It shows whether the letter is a time-limited demand and the respond-by date the letter sets, computed in code from its own words and the day you received it, next to the statute's minimum where the state is encoded (California, Georgia, Missouri, Montana, Utah; Florida's safe harbour). Every condition of acceptance is quoted with its page, the state's required elements are checked, each bill line is tied to its page and summed, and the total is reconciled with what the letter claims. It lists providers billed with no records, records with no bill, gaps in treatment, earlier care, and notes on the letter's citations (a statute from another state, a case that isn't at its cite). It never values the claim or recommends paying. JSON: https://decosa.ai/use-cases/demand-reader.json
- [Check a certificate request](https://decosa.ai/tools/insurance/certificate-check) (`certificate-check`, preview; home: Insurance; runs: selfhost): For commercial-lines account managers at independent agencies. When a general contractor, landlord or lender asks for a certificate of insurance and attaches a long contract, it pulls every insurance requirement out of the contract and the request email and checks each against the insured's policy papers: met by the policy (with the wording quoted), needs an endorsement or a policy change from the carrier, a certificate can't do this, or the papers don't say. It drafts the request to the carrier, a note to the holder, and plain text to paste into your own licensed certificate system. A certificate can't add coverage, so it never suggests putting anything on the certificate that the policy doesn't have. No ACORD form is drawn or reproduced. JSON: https://decosa.ai/use-cases/certificate-check.json
- [Check a bank-detail change before you pay](https://decosa.ai/tools/finance/bank-change-check) (`bank-change-check`, preview; home: Finance and insurance; runs: selfhost): For AP clerks, controllers and the bookkeepers who pay vendors for clients. Forward the email that asks to change a vendor's bank details, with your vendor file. It lists the warning signs with their evidence: a look-alike domain, replies routed elsewhere, failed sender checks, a sender who isn't on file, a new phone number, a bank in another country, an account holder who isn't the vendor, pressure, secrecy, "don't call". It always says to hold the change and call the number already on file, never the one in the email, because a real vendor's hacked mailbox passes every check the email allows. You record the call-back and a second person approves, in a signed record. It never says an email is safe and never releases or blocks a payment. JSON: https://decosa.ai/use-cases/bank-change-check.json
- [Turn a family interview into a film](https://decosa.ai/apps/family-film) (`family-film`, live; home: Creative and media; runs: hosted, selfhost): Record Grandma telling her stories, in her language, and add her old photos. Studio hears who is speaking, finds the chapters of her life and her best lines (each traced to the second she said it), subtitles her in English under her own words, reads the backs of her photos, and cuts a trailer and a film: her real voice over her real photos, with maps, dates and a quiet score. A book view puts a QR code next to every quote that plays her saying it. Her consent is recorded in her own words and her own link takes it back. No model ever draws, animates or changes a photo, and there is no voice clone. JSON: https://decosa.ai/use-cases/family-film.json
- [Make a music video starring you](https://decosa.ai/apps/music-video-starring-you) (`music-video-starring-you`, preview; home: Creative and media; runs: hosted, selfhost): Record a 15-second consent clip (a selfie reading a sentence that ends in three fresh words), add your track and a line of direction, and get 8 to 12 shots of you cut on the beat: the beat grid and sections of the song, a shot list per section you can edit, a re-roll on any shot, a 16:9 and a 9:16 export and a Spotify Canvas loop. Your face comes only from your consent clip: there is no photo upload, and adults only. Every frame says AI video and the end card credits the music to you. Moving shots run on MiniMax H3 when a render GPU is attached; otherwise the shots are a storyboard of drawn stills with camera moves. JSON: https://decosa.ai/use-cases/music-video-starring-you.json
- [Make a film of your story together](https://decosa.ai/apps/our-story-film) (`our-story-film`, preview; home: Creative and media; runs: hosted, selfhost): A 60-second film of the two of you, cut to your song. Each partner records their own consent clip, you both write three memories in your own words, and Studio makes an opening shot, the three memories in order with a line from your own words on screen, and a closing shot of the two of you now, cut on the beat. Both faces come only from your consent clips; adults only; nobody else in focus. Private by default: share it with the people you name, or download it. JSON: https://decosa.ai/use-cases/our-story-film.json
- [Make a settlement video](https://decosa.ai/legal/settlement-video) (`settlement-video`, live; home: Legal; runs: selfhost): For plaintiff's personal-injury lawyers taking a case to mediation. Start from the matter's medical chronology (every entry already cited to a page and box), the itemized bills and the client's signed statement and photos. An open model drafts a 7-scene script and names the entries each line rests on; the cites are attached in code. Every line is checked against what it cites, every date, count and amount is checked against the record, and the bills are added up in code, with a statement's wrong total, duplicate charges, charges before the injury and charges with no matching record flagged. The lawyer edits, reorders and approves each scene and picks the narration: their own recording, their voice with their recorded consent, or a stock house voice. Frames are drawn from the exhibits (no video-generation model), and the MP4 ships with content credentials and a cite sheet. A drafting aid; the lawyer decides. JSON: https://decosa.ai/use-cases/settlement-video.json
- [Find the themes in your interviews](https://decosa.ai/tools/research/interview-themes) (`interview-themes`, preview; home: Science and research; runs: hosted, selfhost): For PhD students, academic researchers, UX researchers and program evaluators with 10 to 60 recorded interviews. Recordings come in as transcripts with every speaker kept apart; a voice check moves segments credited to the wrong person and marks them, and interviewer turns are never coded or quoted. An open model proposes a codebook of 8 to 12 codes that you rename, merge, split and approve; then every passage of participant talk is coded against it. Themes come with participant counts worked out in code and three quotes each, every quote checked word for word against the transcript and the person it is credited to, and playable from its timestamp. Export a REFI-QDA project, CSVs, a methods paragraph and a signed record of every model call; after you review a few interviews, a small model of your own codes the rest your way. JSON: https://decosa.ai/use-cases/interview-themes.json
- [Find the songs in your mix](https://decosa.ai/tools/media/mix-cue-sheet) (`mix-cue-sheet`, live; home: Media and marketing checks; runs: hosted, selfhost): Drop in a DJ mix or any long recording, with the tracklist if you have it and the track files you played if you want exact starts. Every song start appears on the timeline; click one and the player jumps there. You get a timestamped tracklist to paste under your YouTube, Mixcloud or SoundCloud upload, a chaptered .m4a your phone skips song by song, and a signed cue sheet that names the audio it describes. It runs on CPU with no model call, and it only names songs you give it a title or a file for. JSON: https://decosa.ai/use-cases/mix-cue-sheet.json
- [Reply to a review without breaking patient privacy](https://decosa.ai/tools/operations/review-reply) (`review-reply`, live; home: Operations; runs: hosted, selfhost): Paste a Google or Yelp review. The open model drafts a short, warm reply, and code checks it before you see it: no offers or refunds in public, no contact details or numbers you didn't give, no arguing, and for a dentist, clinic or therapist never a word that confirms the reviewer is a patient: no visit, treatment, wait, bill, family member or name. A second model check reads health and care replies on their own. A blocked draft is redrafted once, then a fixed safe reply. Nothing is posted and the review isn't stored. JSON: https://decosa.ai/use-cases/review-reply.json
- [See what studies found for a supplement](https://decosa.ai/tools/life-sciences/what-studies-found) (`what-studies-found`, preview; home: Life sciences; runs: hosted, selfhost): Type a supplement and an outcome, such as probiotics and antibiotic-associated diarrhea. It searches PubMed for trials, meta-analyses and systematic reviews and gives you a table of what they found: one row per study with its design, how many people, who they were, whether the result was better, no different or worse, the finding in numbers where the abstract gives them, and a short quote. Each row links to its PubMed record, each quote is found word for word in the abstract by code, and each finding is judged against its own abstract. A published rubric in code then grades the body of evidence, a separate reading of every abstract flags possible harm, and the studies it left out are listed with why. It reports what studies found: no doses, no advice on what to take. JSON: https://decosa.ai/use-cases/what-studies-found.json

## Building blocks

- Grounding (`grounding`): Checks each sentence against its sources: supported, partial, unsupported or contradicted, with the span it rests on. Used by: clinical, record, grounding, deposition, filing-preflight, promo-claims-check, security-questionnaire, clinical-ai-monitor, editorial-ledger, report-integrity, legal-drafting-editor, foia-desk, split-sheet-check, paper-claim-check, green-claims-check, animatic-studio, sar-narrative-desk, claims-conduct-pack, incident-notification-pack, filing-tieout, medicare-call-record, denial-appeal-packet, hcc-evidence-file, device-mdr-triage, cmmc-evidence-map, eu-trial-lay-summary, reg-e-dispute-file, ma-dd-redflags, nsa-idr-packet, expert-to-sop, privileged-call-notes, check-their-brief, payer-audit, prior-auth-check, demand-reader, certificate-check, settlement-video, what-studies-found.
- Typed judgment (`typed-judgment`): Yes/no, one-of-N, score and label questions answered as typed values with a calibrated probability and a receipt. Used by: hiring-screen, migration-check, typed-judgment, model-risk-pack, privilege-log, foia-desk, music-gen-cleared, sample-clearance, split-sheet-check, green-claims-check, fi-disclosure-record, disclosure-preflight, music-video-studio, claims-conduct-pack, collections-call-qa, sanctions-disposition-record, medicare-call-record, denial-appeal-packet, kids-content-preflight, hcc-evidence-file, vex-triage, device-mdr-triage, cmmc-evidence-map, reg-e-dispute-file, medical-chronology, nsa-idr-packet, prior-auth-check, geo-audit.
- Signed record (`signed-record`): A hash chain over every entry, signed with Ed25519, that anyone can re-verify in the browser. Used by: record, auditor, grounding, deposition, hiring-screen, filing-preflight, promo-claims-check, migration-check, typed-judgment, security-questionnaire, flight-recorder, test-runs, model-risk-pack, clinical-ai-monitor, privilege-log, editorial-ledger, report-integrity, legal-drafting-editor, foia-desk, oral-assessment, patent-claim-support, music-gen-cleared, sample-clearance, split-sheet-check, paper-claim-check, signed-lab-notebook, green-claims-check, fi-disclosure-record, consent-ledger, consented-dubbing, disclosure-preflight, animatic-studio, audio-drama-studio, music-video-studio, sar-narrative-desk, claims-conduct-pack, collections-call-qa, incident-notification-pack, filing-tieout, sanctions-disposition-record, medicare-call-record, denial-appeal-packet, kids-content-preflight, hcc-evidence-file, vex-triage, device-mdr-triage, cmmc-evidence-map, eu-trial-lay-summary, reg-e-dispute-file, virtual-staging, ma-dd-redflags, tariff-classification, csr-number-verifier, gpsr-listing-pack, medical-chronology, nsa-idr-packet, walkthrough-to-quote, expert-to-sop, honest-product-imagery, pv-intake, label-consistency-check, no-training-receipts, storefront-accessibility-pass, fill-and-stop, privileged-call-notes, discovery-deficiency, payer-audit, prior-auth-check, jottings-note, evidence-runner, ehr-drafts, demand-reader, certificate-check, bank-change-check, settlement-video, interview-themes, mix-cue-sheet, geo-audit, review-reply, what-studies-found, child-welfare-notes, crp-legal-matter.
- Live speech to text (`live-asr`): Streaming transcription from a microphone or a sample script, sentence by sentence. Used by: clinical, sales, field, translate, record, oral-assessment, consented-dubbing, pv-intake, privileged-call-notes, family-film, music-video-starring-you, child-welfare-notes.
- Speaker diarization (`diarize`): Who said what: transcripts with speakers from recorded audio. Used by: clinical, record, clinical-ai-monitor, report-integrity, oral-assessment, fi-disclosure-record, collections-call-qa, medicare-call-record, kids-content-preflight, walkthrough-to-quote, expert-to-sop, pv-intake, privileged-call-notes, interview-themes.
- Studio render (`studio-render`): Music, image and video generation on pinned checkpoints, queued, with a seeded render receipt. Used by: studio, provenance, ugc, characters, music-gen-cleared, animatic-studio, audio-drama-studio, music-video-studio, virtual-staging, family-film, music-video-starring-you, our-story-film.
- Content credentials (`content-credentials`): C2PA credentials and watermarks on rendered files, and a checker for any file. Used by: provenance, ugc, characters, editorial-ledger, music-gen-cleared, consent-ledger, consented-dubbing, disclosure-preflight, animatic-studio, audio-drama-studio, music-video-studio, virtual-staging, honest-product-imagery, family-film, music-video-starring-you, our-story-film, settlement-video.
- Agent flight recorder (`flight-recorder`): A signed, step-by-step record of what a browser or computer-use agent saw, decided and did. Used by: flight-recorder, test-runs, cmmc-evidence-map, fill-and-stop, evidence-runner, ehr-drafts.
- Endpoint audit (`endpoint-audit`): Probes an OpenAI-compatible endpoint for model identity, quality and speed, and signs the verdict. Used by: auditor, model-risk-pack.
- Numeric grounding (`numeric-grounding`): Every amount, date, count, span and percentage in a sentence recomputed in code from the table rows it cites; a mismatch comes back with the figure the rows do support. Used by: sar-narrative-desk, incident-notification-pack, filing-tieout, eu-trial-lay-summary, csr-number-verifier, nsa-idr-packet, walkthrough-to-quote, settlement-video.
- Document reader (`document-reader`): PDFs (born-digital or scanned), images, forms and handwriting read into elements with page and box: text in reading order, tables as cells with spans, form fields and checkboxes, and a signed receipt over the page pixels. Used by: hcc-evidence-file, csr-number-verifier, medical-chronology, honest-product-imagery, label-consistency-check, discovery-deficiency, settlement-video.
- Evidence retrieval (`evidence-retrieval`): Finds the passages that answer a question in your documents: an embedder, BM25 and a reranker, each hit a chunk with byte offsets and a Merkle proof into a signed index snapshot. Used by: security-questionnaire, ma-dd-redflags, tariff-classification, csr-number-verifier, medical-chronology, what-studies-found.
- Language pack (`language-pack`): Translation into all 24 EU official languages with a number lock, the model chosen per language from measured scores: every number, unit, date, code, negation and required term checked in each language's own format, drift flagged with its source span. Speech in consented house voices, speech recognition, and a signed receipt for every model call. Used by: eu-trial-lay-summary, gpsr-listing-pack, label-consistency-check.
- Video understanding (`video-understanding`): One recording per request to Qwen3.8-27B, sampled at 1 frame a second (up to 240 frames, 32,768 video tokens), answers that cite time ranges, and a receipt whose request hash covers the video's sha256 and the sampling settings. Used by: walkthrough-to-quote, expert-to-sop.
- Evidence tables from PubMed (`pubmed-evidence`): What the trials, meta-analyses and reviews in PubMed found for a supplement and an outcome: one row per study with its PMID, design, n, direction, finding and a quote checked word for word, each finding judged against its own abstract, graded by a published rubric in code, with a signed record. Used by: what-studies-found.
- Record AI decisions about people (`decision-record`): A signed, hash-chained record of every automated step your app takes about a person (score, screen, rank, filter, a message sent without approval) and the human decisions after it, holding hashes and your own ids, never personal data. Explain one person's steps in plain words and export a retention log. Open-source TypeScript; nothing is sent to Decosa. Used by: none yet.
- Song starts in long audio (`song-starts`): Where each song starts in a mix, a radio show or any long recording, from a tracklist (with or without times), your own track files and the audio itself, on CPU with no model call: a chaptered .m4a, a timestamped tracklist and a signed cue sheet that names the audio's SHA-256. Used by: mix-cue-sheet.
- Consent gate (`consent-gate`): A signed consent ledger for voices, faces and movement: every render is checked against a specific, revocable consent first, and the decision is signed either way. Used by: consent-ledger, consented-dubbing, animatic-studio, audio-drama-studio, honest-product-imagery, settlement-video.
- Form filling (`formfill`): Fills a PDF or web form from your own documents, shows the source line for every value, never types passwords or card numbers, and stops before Submit. Used by: clinical, denial-appeal-packet, fill-and-stop, prior-auth-check.
- Citation index (`citation-index`): Looks up case citations in a local index of public court opinions first (a remote lookup only on a miss), so invented citations are caught without sending the brief out. Used by: filing-preflight, check-their-brief.
- Small checker models (`checker-models`): Two small models of our own, Apache-2.0 on Hugging Face, that run on a CPU: citation support (does the cited passage back the sentence?) and note-detail checking (did a drug, dose or date change between transcript and note?). Used by: none yet.
- Site checks (`site-checks`): Deterministic checks for AI answers: robots.txt for AI crawlers by RFC 9309, llms.txt, JSON-LD property by property, and a review-reply guard that never confirms a patient. Open source, in TypeScript and Python. Used by: geo-audit, review-reply.

## Ways to run it

- [Decosa hosted](https://decosa.ai/developers): The API and console, on Decosa's hosted service.
- [Self-host kits](https://decosa.ai/self-host): The same apps and pinned models, on your GPU.
- [Confidential Research Platform](https://decosa.ai/products/crp): A private research workspace that runs on your computer.
- [Decosa Studio](https://decosa.ai/studio): Make things with your people in them: stories, songs, films and voices from one project.
  - Story (`studio-story`, https://decosa.ai/studio#studio-story): Print-ready picture book (coming)
  - Music (`studio-music`, https://decosa.ai/studio#studio-music): music-gen-cleared, music-video-studio, music-video-starring-you, mix-cue-sheet, Make it a song (from your Story) (coming), Lyric videos and visualisers (coming)
  - Film (`studio-film`, https://decosa.ai/studio#studio-film): animatic-studio, family-film, our-story-film, Real moving shots (not just camera moves) (coming)
  - Voice (`studio-voice`, https://decosa.ai/studio#studio-voice): consented-dubbing, audio-drama-studio, Audiobook chapters in your own voice (coming), Podcast episodes (coming)
  - Characters & Worlds (`studio-characters`, https://decosa.ai/studio#studio-characters): characters, Theme songs and expression sheets (coming)
  - Brand & Ads (`studio-brand`, https://decosa.ai/studio#studio-brand): ugc, honest-product-imagery, virtual-staging, Market variants and a mascot kit (coming)
  - Studio checks (`studio-checks`, https://decosa.ai/studio#studio-checks): provenance, sample-clearance, split-sheet-check, kids-content-preflight, AI music release check (coming)

## Tool details

### Run the visit with a copilot (`clinical`)

The visit copilot: a live transcript with speaker labels, guidance on what is still worth asking (each item with its source), a note checked sentence by sentence against the visit, codes, prior-auth prep, and the work note, FMLA sections and patient instructions drafted, never signed or sent.

- What it is: The visit copilot: transcribes with speaker labels, reminds you what is still worth asking (each item with its source), writes a note checked sentence by sentence before you see it, codes it, and drafts the paperwork, on your own GPU.
- Who it's for: Family physicians, NPs and PAs, and clinics that want the note, codes and forms from the visit on their own hardware.
- Page: https://decosa.ai/clinics/visit-copilot
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/clinical):
  - 8.4%: Medical-term miss rate, live ASR (Voxtral Mini 4B Realtime) (held out, n = 57)
  - 99.71%: Speaker labels during the visit, word-level role accuracy (rolling MOSS windows) (held out, n = 11)
  - 14/15 and 13/15: Considerations: warning-feature encounter recall (synthetic, held-out) (held out, n = 15)
  - 60.0 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: live. Models: Voxtral 4B · MOSS diarizer · Qwen3.8-27B · M17 detail checker. Where: Self-host (patient data stays on site).
- Verification: Note checked before it is final (receipt coverage: full). Hosted QA 2026-09-29: pass. Self-host QA 2026-09-29: pass.
- Hardware: 1× RTX PRO 6000 (96 GB), or 2× RTX 5090
- Output: The self-checked note, validated codes, the paperwork drafts, and the clinical considerations to accept or dismiss.
- Page: https://decosa.ai/clinics/visit-copilot · JSON: https://decosa.ai/use-cases/clinical.json
- Prompts: https://decosa.ai/prompts/clinical-hosted.md · https://decosa.ai/prompts/clinical-selfhost.md · https://decosa.ai/prompts/clinical-assemble.md · https://decosa.ai/prompts/clinical-mac.md
- Tiers:
  - Lite · one 48 GB card, live pass only (1x L40S or RTX 6000 Ada 48 GB (FP8 weights; starting-point settings, not measured)): Live transcript, lanes and a final note written from the streaming transcript. No speaker labels, no citations, no verifier; the pipeline the benchmark calls vanilla.
  - Standard · one 96 GB Blackwell card, two passes (1x RTX PRO 6000 Blackwell 96 GB): The hosted demo: live captions, speaker labels, guidance and lanes during the visit; then the cited note, its self-check (with M17 on CPU), assessment codes and paperwork, all on one Qwen3.8-27B.
  - Best · adds DeepSeek V4 Flash as the note writer on 2x 96 GB (2x RTX PRO 6000 96 GB for the writer (about 83 GB weights per GPU) plus the standard card for the live pass): Standard stack plus the loop's best writer: higher term precision and, with citations, the most grounded notes; costs two extra 96 GB cards and some plan recall.
- Buyer facts:
  - Measured latency: The note, its self-check, the assessment codes and the paperwork drafts arrive about a minute after the audio ends on the shared gateway (measured), sooner on the direct route. Speaker labels follow the captions after a short delay; guidance updates every few seconds of speech.
  - Typical run cost: Several cents per visit for the synthetic sample visit: model calls at the gateway list price plus a little speech-recognition GPU time (measured). The "in model calls" figure after a run is the first part only. Cost grows with visit length (the guidance and the live view reread the transcript every tick).
  - Data retention: Audio and transcript live in server memory for the session and are dropped when it ends; paperwork PDFs are rendered on request and not stored. The receipt store keeps hashes of each model call, not the text.
  - What leaves the box (self-host): Only the codes lane's lookups: a condition name to NLM Clinical Tables and a drug name to NLM RxNav and openFDA. Block egress for the api container and the lanes report no match; the note is still written.
  - Input: 16 kHz mono PCM16 over WebSocket (about 100 ms frames), or a canned sample through /demo/replay. US English visits.
  - Clinical considerations: After the visit: up to 6 items (warning features first, then differentials) from a 26-topic public-domain pack, each with the patient's words, the source excerpt and URL, and what the note does not record. Clinician only; nothing enters the note; each accept or dismiss is sealed in a signed record. About 10 model calls and USD 0.006 for the chest-pain sample transcript (text route, 2026-09-26). Not a diagnosis and not for urgent decisions.
  - Guidance during the visit: Warning-feature considerations, checklist items of a triggered topic not yet asked, and history elements not yet covered (CMS 1995 Documentation Guidelines). Every item shows the transcript words and a public source, in fixed wording, and turns 'covered' when asked. No sound, no alarm colours, no deadlines. A blind clinician judge found 66% of checklist items useful at that moment but only 11% of history items, so history folds away until you ask ('Only when I ask' mode).
  - Paperwork: Patient instructions (the clinician's own words), a work or school note (dates said in the visit are worked out in code: 'through Friday, October 2nd' becomes 10/02/2026, citing the turn), and the Department of Labor's WH-380-E provider sections. Signatures, signing dates and the employer's section are never filled; nothing is sent. WH-380-E drafts need every box checked: 91.4% of the boxes it filled were right on 8 held-out visits.
- Samples:
  - 1. Back pain visit (`clinical-back-pain`): https://decosa.ai/clinics/visit-copilot?sample=1&autorun=0
  - 2. Diabetes follow-up (`clinical-diabetes-followup`): https://decosa.ai/clinics/visit-copilot?sample=2&autorun=0
  - 3. Chest burning on the stairs (`clinical-chest-pain`): https://decosa.ai/clinics/visit-copilot?sample=3&autorun=0
- Regulatory note: HIPAA: run it on the clinic's own hardware so audio, transcripts and notes stay on-site; there is no hosted BAA yet, so the hosted demo takes synthetic visits only. Drafts for clinician review; not a diagnostic device. The guidance lane is designed to stay outside FDA's device definition under FD&C Act 520(o)(1)(E) (FDA guidance on software functions, January 2026): no image or signal analysis (the audio is only transcribed), clinician only, considerations rather than directives, the basis (transcript words and a public source) shown on every item, no alerting or time pressure, nothing written into the note. That is our reading, not a legal opinion: counsel review of the in-visit use is pending before the guidance lane is marketed. Paperwork drafts never fill signatures, signing dates or attestations and are never sent; the clinician reviews, signs and sends. California AB 3030 exempts only clinician-reviewed AI messages to patients. CPT codes are AMA-licensed and hidden in the demo.

### Run an open model behind the OpenAI API (`code`)

A coding agent on an open model that never sends your source to a third party.

- What it is: An OpenAI-compatible coding model on your own GPU, so your source code never leaves the machine.
- Who it's for: Teams in software and ai ops.
- Page: https://decosa.ai/tools/developer/code
- Last reviewed: 2026-09-30
- Key numbers (details and caveats: https://decosa.ai/metrics/code):
  - 478 / 500: Coding benchmark, 5 core tasks, direct API (Qwen3.8-27B, vLLM NVFP4 + MTP) (test split, n = 5)
  - vLLM FP8 eager 387; vLLM FP8 + MTP 193; llama.cpp CUDA Q8_0 63 (/500): Same weights on other engines (test split, n = 5)
  - 12/12: Code smoke on the exact pinned stack (test split, n = 12)
  - 4.3 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: live. Models: Qwen3.8-27B. Where: Hosted or self-host.
- Verification: Strong (receipt coverage: full). Hosted QA 2026-09-30: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB), or 2× RTX 5090
- Output: Streamed answers, each with a receipt.
- Page: https://decosa.ai/tools/developer/code · JSON: https://decosa.ai/use-cases/code.json
- Prompts: https://decosa.ai/prompts/code-hosted.md · https://decosa.ai/prompts/code-selfhost.md · https://decosa.ai/prompts/code-assemble.md · https://decosa.ai/prompts/code-mac.md
- Tiers:
  - Lite · runs on one 32 GB Blackwell card (1x RTX 5090 32 GB (Blackwell, NVFP4)): Same Qwen3.8-27B NVFP4 weights and engine as Standard. You give up context length and concurrent sessions, not model quality.
  - Standard · one 96 GB card (hosted demo) (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B NVFP4 + MTP3 on the pinned vLLM 0.29.0 with a full 262k context and room for about 32 concurrent streams. This is the hosted demo.
  - Best · two 96 GB cards (2x RTX PRO 6000 Blackwell 96 GB): DeepSeek-V4-Flash (284B MoE, MIT) across both GPUs for longer agentic runs. It takes the whole box. The 5-task suite is near its ceiling, so it doesn't clearly beat Standard one-shot; its edge shows with an agent harness.
  - Wanted · the two most-used open coding models (Network providers: 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash; an 8x H200-class node for DeepSeek-V4.1-Flash. Estimate.): GLM-5.3-Flash and DeepSeek-V4.1-Flash, served by network providers. Neither fits one workstation card: GLM needs two 96 GB cards (its sm_120 runtime hangs on our server today) and V4.1-Flash a datacentre node.
- Buyer facts:
  - Data retention: Hosted: prompts and answers are not stored; only the receipt (hashes, token counts, cost) is kept. Self-hosted: nothing leaves the machine unless you turn on the network profile.
  - What leaves the box (self-host): Nothing. vLLM is bound to 127.0.0.1 and the optional api service signs receipts with a local key.
  - Inputs: OpenAI-compatible chat messages, up to 48,000 characters per request on the hosted API; model qwen3.8-27b.
  - Typical hosted cost: A fraction of a cent per short answer at the gateway's list price (measured). Each run shows its own measured cost.
  - Hosted limits: Demo session: 20,000 generated tokens. API key: 200,000 generated tokens a day, 60 requests a minute.
- Regulatory note: Self-hosted, source code and prompts stay on your machine. The hosted demo sends prompts to a shared GPU and our gateway, so don't paste proprietary code there. Model licences: Apache-2.0 (Qwen3.8-27B), MIT (optional DeepSeek-V4-Flash tier).

### Make a story, a song or a music video (`studio`)

Make things with your people in them: a picture book read in your own voice, a song, a music video, an animatic or a dub. Add your cast once and every pack uses it.

- What it is: Local AI image generation, plus video and music, from open weights on your own GPU: a render queue for images, songs and short videos, with models loading one at a time, private checkpoints and LoRAs kept on your disk, and a C2PA credential and signed receipt on every render.
- Who it's for: Teams in creative and media and personal and family.
- Page: https://decosa.ai/studio
- Last reviewed: 2026-09-25
- Key result: There is no quality eval yet, on any tier. The one check so far is reproducibility: a same-seed re-render was bit-identical for MiniMax-Music3 and Qwen-Image-2512, and not for ACE-Step. That shows reproducibility, not quality.

- Status: live. Models: Qwen3.8-27B · FLUX.2 [klein] 4B · Chatterbox · ACE-Step · MiniMax-Music3 · MiniMax H3. Where: Hosted or self-host.
- Verification: Partial (seeded re-render) (receipt coverage: partial). Hosted QA 2026-09-25: partial. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB)
- Output: Rendered audio, images and video.
- Page: https://decosa.ai/studio · JSON: https://decosa.ai/use-cases/studio.json
- Prompts: https://decosa.ai/prompts/studio-hosted.md · https://decosa.ai/prompts/studio-selfhost.md · https://decosa.ai/prompts/studio-assemble.md · https://decosa.ai/prompts/studio-mac.md
- Tiers:
  - Lite · drafts on one 24–32 GB card (1× 24–32 GB card (not tested)): Fast music drafts, draft images and small draft video on one consumer card. Quality is below the standard tier, and Z-Image-Turbo and Wan 1.3B are not tested yet.
  - Standard · the hosted demo, about 50 GB of one 96 GB card (1× RTX PRO 6000 Blackwell 96 GB, about 50 GB free): The hosted gallery and queue: MiniMax-Music3 songs, Qwen-Image-2512 images and Wan2.1-T2V-14B video (silent), with ACE-Step clips as well. Every model's licence allows hosted use. Hosted video runs on MiniMax H3 Max through fal at list price with no markup (API key, 2 per key per day, daily cap); Wan2.1 stays the fully open option.
  - Best · self-host only, a whole 96 GB card (1× RTX PRO 6000 Blackwell 96 GB, 256 GB RAM for MiniMax-H3): Adds video with sound: MiniMax-H3 (licence pending; self-host available where its current licence covers you) or LTX-2.5 (below USD 10M revenue, not in a competing hosted service). Not in the hosted demo until the H3 licence lands.
  - Wanted · MiniMax H3 on two cards, no offload (2x RTX PRO 6000 96 GB, no CPU offload (about 133 GB of BF16 weights; estimate)): The studio's best video model, faster. Both halves of MiniMax-H3 in BF16 on two 96 GB cards, so nothing is offloaded to RAM: faster renders (estimate, not measured). Self-host now where H3's current licence covers you; once the pending licence lands we add this compute for the hosted tier ourselves. Not a community-provider model.
- Buyer facts:
  - Data retention: Rendered files stay under /studio/media on the server; job status is kept in a file without tokens, and without the prompt once a job has started. Logs carry prompt hashes, never prompts. On fal (hosted video only): fal keeps request data (prompts and file links) for 30 days by default and serves uploaded and generated files from public CDN links until they are deleted [V: fal docs, Data Retention & Storage, read 29 Sep 2026]; Decosa downloads each result at once, and fal's no-storage and short-expiry options are not switched on yet.
  - What leaves the box: Hosted: your prompt reaches Decosa's render service and the text model's prompt check through our gateway; hosted video prompts (text only: no photos, voices or consent recordings) go to fal, a partner GPU provider, under Decosa's account. Self-host: nothing.
  - Inputs: A prompt of up to 1,000 characters (the console takes 600); music takes lyrics, seconds and a seed; images take an aspect (square, landscape, portrait).
  - Outputs: MP3 songs (MiniMax-Music3 must be credited: 'Powered by MiniMax-Music3'), PNG images, MP4 video; each with a C2PA credential and a signed render receipt.
  - Typical time: Measured: a song in about a minute, an image in a couple of minutes, Wan2.1 video in over half an hour.
- Samples:
  - 1. Music3 harbor lights (`music3-harbor-lights`): https://decosa.ai/studio?sample=1&autorun=0
  - 2. Music3 neon drive (`music3-neon-drive`): https://decosa.ai/studio?sample=2&autorun=0
  - 3. Music3 paper garden (`music3-paper-garden`): https://decosa.ai/studio?sample=3&autorun=0
  - 4. Music lofi evening (`music-lofi-evening`): https://decosa.ai/studio?sample=4&autorun=0
  - 5. Music orchestral rise (`music-orchestral-rise`): https://decosa.ai/studio?sample=5&autorun=0
  - 6. Music indie morning (`music-indie-morning`): https://decosa.ai/studio?sample=6&autorun=0
  - 7. Music ambient tide (`music-ambient-tide`): https://decosa.ai/studio?sample=7&autorun=0
  - 8. Image lighthouse dawn (`image-lighthouse-dawn`): https://decosa.ai/studio?sample=8&autorun=0
  - 9. Image greenhouse cafe (`image-greenhouse-cafe`): https://decosa.ai/studio?sample=9&autorun=0
  - 10. Image fox watercolor (`image-fox-watercolor`): https://decosa.ai/studio?sample=10&autorun=0
  - 11. Image poster typography (`image-poster-typography`): https://decosa.ai/studio?sample=11&autorun=0
  - 12. Image isometric city (`image-isometric-city`): https://decosa.ai/studio?sample=12&autorun=0
  - 13. Image macro dew (`image-macro-dew`): https://decosa.ai/studio?sample=13&autorun=0
  - 14. Video wan forest sunrise (`video-wan-forest-sunrise`): https://decosa.ai/studio?sample=14&autorun=0
  - 15. Video wan rainy window (`video-wan-rainy-window`): https://decosa.ai/studio?sample=15&autorun=0
- Regulatory note: Hosted demo models: MiniMax-Music3 (show "MiniMax-Music3" prominently; abuse and IP safeguards; authorization above USD 20M revenue), ACE-Step (MIT), Qwen-Image-2512 (Apache-2.0) and Wan2.1-T2V-14B (Apache-2.0). LTX-2.5 and MiniMax-H3 are self-host only because of their licences. Voice cloning needs documented consent from the speaker; never clone a real person's voice or likeness. Excluded: Qwen-Image-2.1 (research-only licence) and OmniVoice (CC-BY-NC weights).

### Write the inspection report (`field`)

Talk through an inspection or site walk. Get the checklist, issues by severity, measurements and a finished report.

- What it is: Talk through an inspection, site visit or job; get the checklist, issues by severity, measurements and a finished report.
- Who it's for: Teams in field and trades.
- Page: https://decosa.ai/tools/operations/field
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/field):
  - 40/41 (98%); before 38/41 (93%): Expected issues in the report (grounded report + safety sweep + claim check) (synthetic, n = 41)
  - 40/40; all 6 critical-only hazards rated critical (before 35/38 and 6/6): Severity within the accepted range, of issues found (synthetic, n = 40)
  - 51/51; before 32/51 (63%): Stated facts kept (synthetic, n = 51)
  - 26.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Voxtral 4B · Qwen3.8-27B. Where: Hosted or self-host.
- Verification: Strong on text lanes (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB), or 2× RTX 5090
- Output: A structured report (JSON) plus markdown.
- Page: https://decosa.ai/tools/operations/field · JSON: https://decosa.ai/use-cases/field.json
- Prompts: https://decosa.ai/prompts/field-hosted.md · https://decosa.ai/prompts/field-selfhost.md · https://decosa.ai/prompts/field-assemble.md · https://decosa.ai/prompts/field-mac.md
- Tiers:
  - Lite · 4-bit on a 32 GB card, plus a small card for speech (1× 32 GB Blackwell card (NVFP4 LLM) + 1× ≥16 GB card (Voxtral)): Same models and weights as Standard, so same output quality; you give up context length (16k) and concurrency (1–2 walk-throughs at a time).
  - Standard · one 96 GB card (the hosted demo) (1× RTX PRO 6000 Blackwell 96 GB): Live Voxtral transcript plus Qwen3.8-27B NVFP4 with MTP; about 4 walk-throughs at once, 64k context.
  - Best · two 96 GB cards (2× RTX PRO 6000 Blackwell 96 GB (TP2)): DeepSeek-V4-Flash writes the lanes and report; a final MOSS-Transcribe-Diarize pass cleans the transcript. Needs the whole two-card box and a community vLLM build.
  - Wanted · the largest open flash models (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). Speech recognition stays local. Estimate.): GLM-5.3-Flash or DeepSeek-V4.1-Flash writes the report, served by network providers; speech recognition stays on your machine. Not served yet.
- Buyer facts:
  - Measured latency: Hosted p50 below is the time from the end of the audio to the final `done`, over 3 replays of the sample (2026-09-25, gateway route, one session at a time).
  - Typical run cost: A few cents for the roof sample (a few dozen model calls) at the gateway list price. Each run shows its own measured cost.
  - Data retention: Audio and transcript in memory for the session only; the report comes back in done and is not stored.
  - What leaves the box (self-host): Nothing, with DECOSA_LLM_ROUTE=direct.
  - Outputs: Structured JSON (issues with severity, quote, line and time; measurements; passed checks) and Markdown; PDF with pandoc.
- Samples:
  - 1. Roof inspection (`field-roof-inspection`): https://decosa.ai/tools/operations/field?sample=1&autorun=0
  - 2. Electrical panel check (`field-electrical-panel`): https://decosa.ai/tools/operations/field?sample=2&autorun=0
- Regulatory note: Drafting aid: the report is a draft for the inspector to review and sign; it does not replace a licensed inspection or code-compliance determination. Self-host keeps site audio, transcripts and reports on your hardware.

### Translate a talk live (`translate`)

Speak in one language and read it in another, sentence by sentence, with a running glossary and a summary at the end.

- What it is: Speech in one language, captions in another, one sentence at a time, with a running glossary and a summary at the end.
- Who it's for: Teams in any industry.
- Page: https://decosa.ai/tools/operations/translate
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/translate):
  - 55.3 / 29.9 / 87.2: FLORES-200 devtest en→es: chrF++ / BLEU / COMET-22 (held out, n = 1012)
  - 59.7 / 31.8 / 87.7: FLORES-200 devtest es→en: chrF++ / BLEU / COMET-22 (held out, n = 1012)
  - 68.8 / 49.1 / 88.9: FLORES-200 devtest en→fr: chrF++ / BLEU / COMET-22 (held out, n = 1012)
  - 2.3 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Voxtral 4B · Qwen3.8-27B. Where: Hosted or self-host.
- Verification: Strong on text lanes (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB), or 2× RTX 5090
- Output: The translation and a summary.
- Page: https://decosa.ai/tools/operations/translate · JSON: https://decosa.ai/use-cases/translate.json
- Prompts: https://decosa.ai/prompts/translate-hosted.md · https://decosa.ai/prompts/translate-selfhost.md · https://decosa.ai/prompts/translate-assemble.md · https://decosa.ai/prompts/translate-mac.md
- Tiers:
  - Lite · one 48 GB card (1× L40S / RTX 6000 Ada 48 GB (or any 48 GB+ card)): A smaller mixture-of-experts model on cheaper hardware. Translations are likely less polished and no quality has been measured; no signed receipts.
  - Standard · one 96 GB card (hosted demo) (1× RTX PRO 6000 Blackwell 96 GB): The stack the hosted demo runs: Qwen3.8-27B NVFP4 with MTP. Every lane call can carry a gateway-signed receipt.
  - Best · two 96 GB cards (2× RTX PRO 6000 Blackwell 96 GB (TP2)): DeepSeek-V4-Flash (284B MoE) on two cards: the strongest model measured on the owner's own benchmarks, at twice the hardware and without signed receipts.
  - Wanted · the largest open flash models (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). Speech recognition stays local. Estimate.): GLM-5.3-Flash and DeepSeek-V4.1-Flash translate, served by network providers; speech recognition stays on your machine. Not served yet.
- Alternates (smaller models for the same job):
  - Specialist translation models (1x 24 GB card (7B) or 1x 48 GB card (30B-A3B FP8)), runs today: Hy-MT2-7B runs today in the language pack (served on Decosa's hosted service; this tool's hosted demo doesn't use it yet); Hy-MT2-30B-A3B would be a one-card best tier (page 32 #8) and is not served. Re-run our FLORES COMET-22 harness before any quality claim.
- Buyer facts:
  - Measured latency: The hosted median below is the time from the end of the audio to the final `done`, over replays of the sample (gateway route, one session at a time). Each sentence's translation arrives about a second after the sentence ends.
  - Typical run cost: A fraction of a cent for the Spanish repair call (under two dozen model calls) at the gateway list price.
  - Data retention: Nothing stored: transcript and translations live in memory for the session.
  - What leaves the box (self-host): Nothing, with DECOSA_LLM_ROUTE=direct.
  - Languages: Speech: ar, de, en, es, fr, hi, it, ja, ko, nl, pt, ru, zh or auto. Targets tested end to end: en, es.
- Samples:
  - 1. Repair visit, Spanish to English (`translate-es-en-repair`): https://decosa.ai/tools/operations/translate?sample=1&autorun=0
  - 2. Tour, English to Spanish (`translate-en-es-tour`): https://decosa.ai/tools/operations/translate?sample=2&autorun=0
- Regulatory note: Machine translation, not a certified interpreter: out of scope for court, legal and medical interpreting. Get consent before you record or transcribe anyone. The hosted demo keeps transcripts in memory only and deletes them when the session ends; for confidential conversations, self-host with DECOSA_LLM_ROUTE=direct so audio and text stay on your machine.

### Make a tamper-evident meeting record (`record`)

Live captions now. Afterwards, a transcript with speakers and a cited summary, sealed in a signed record that anyone can re-check.

- What it is: Live captions now; afterwards a speaker transcript and a cited summary, sealed in a signed record anyone can re-check.
- Who it's for: Teams in public sector and compliance and trust.
- Page: https://decosa.ai/tools/developer/record
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/record):
  - 14/17 council, 9/9 interview: Claim check on the synthetic sessions: sentences supported (synthetic, n = 26)
  - 5 speakers found; one short turn given to the wrong member: Speaker attribution on the synthetic council meeting (5 TTS voices) (synthetic, n = 5)
  - 16.6 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Voxtral 4B · MOSS-TD 0.9B · Qwen3.8-27B. Where: Hosted for public meetings, self-host for investigations.
- Verification: Signed hash chain over every line (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB), or 2× RTX 5090
- Output: A signed session record: every caption, transcript line, summary sentence and model receipt in one hash chain.
- Page: https://decosa.ai/tools/developer/record · JSON: https://decosa.ai/use-cases/record.json
- Prompts: https://decosa.ai/prompts/record-hosted.md · https://decosa.ai/prompts/record-selfhost.md · https://decosa.ai/prompts/record-assemble.md · https://decosa.ai/prompts/record-mac.md
- Tiers:
  - Lite · one 48 GB card, captions only (1x L40S or RTX 6000 Ada 48 GB (not measured)): Captions, actions lane and a cited summary from the live captions, all chained and signed. No speaker labels.
  - Standard · the hosted demo, two passes (1x RTX PRO 6000 Blackwell 96 GB (hosted demo uses two cards on our server)): Live captions, then speakers, roles, cited summary and claim check; the record embeds gateway-signed receipts for every model call.
  - Best · DeepSeek-V4-Flash writes and checks (2x RTX PRO 6000 96 GB for the writer plus the standard card): Standard plus a larger writer and verifier on two more 96 GB cards.
- Buyer facts:
  - Measured latency: Hosted p50 below is the time from the end of the audio to the final `done`, over 3 replays of the sample (2026-09-25, gateway route, one session at a time).
  - Typical run cost: About a cent for the council sample: a couple of dozen model calls at the gateway list price, plus caption receipts signed by the server's key at no model cost.
  - Data retention: The signed record is returned to you in done; the server keeps receipts (hashes), not the transcript.
  - What leaves the box (self-host): Nothing, with DECOSA_LLM_ROUTE=direct. Anyone can verify a record offline or with POST /record/verify.
- Samples:
  - 1. Town council meeting (synthetic) (`record-council-meeting`): https://decosa.ai/tools/developer/record?sample=1&autorun=0
  - 2. Workplace investigation interview (synthetic) (`record-investigation-interview`): https://decosa.ai/tools/developer/record?sample=2&autorun=0
- Regulatory note: Not a certified court record: official transcripts need a certified reporter. Get consent before recording where the law requires it. Receipts attest that the record was not changed after signing; they do not prove that speech was recognised correctly. Captions for ADA Title II web content need accuracy checks; WER on meeting audio is not measured yet.

### Check your provider serves the model you pay for (`auditor`)

Checks whether an OpenAI-compatible endpoint really serves the model it claims, at the quality it claims, and signs the result.

- What it is: Continuous, signed checks that an OpenAI-compatible endpoint serves the model and quality it claims.
- Who it's for: Teams in software and ai ops and compliance and trust.
- Page: https://decosa.ai/tools/developer/auditor
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/auditor):
  - fail, re-check agreed; greedy 0/10, top-5 overlap 0.551, 7 hard divergences: Swap caught: Qwen3.5-4B-Base served as qwen3.8-27b (synthetic)
  - drift, re-check agreed; greedy 5/10, top-5 overlap 0.928 vs band ≥ 0.963: Quantisation drift caught: FP8 re-quant claimed as BF16 (4B) (synthetic)
  - worst greedy repeat 5/10; mean |Δ logprob| ≤ 0.040; top-5 overlap ≥ 0.872: Reference noise, Qwen3.8-27B stack (6 runs) (synthetic, n = 6)
  - 39.8 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Reference runs of Qwen3.8-27B · Qwen3.5-4B. Where: Hosted or self-host.
- Verification: Signed report, receipted probes (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: No GPU needed to audit. References are recorded on 1× RTX PRO 6000 (96 GB).
- Output: A signed audit report.
- Page: https://decosa.ai/tools/developer/auditor · JSON: https://decosa.ai/use-cases/auditor.json
- Prompts: https://decosa.ai/prompts/auditor-hosted.md · https://decosa.ai/prompts/auditor-selfhost.md · https://decosa.ai/prompts/auditor-assemble.md · https://decosa.ai/prompts/auditor-mac.md
- Tiers:
  - Lite · audits only, no GPU (Any Linux or macOS machine with Python 3.11+): Run the auditor against any endpoint with the signed reference fixtures we ship. You cannot record new references.
  - Standard · one 96 GB card (hosted demo) (1x RTX PRO 6000 Blackwell 96 GB): Adds recording your own references on the pinned Qwen3.8-27B stack, with its noise band measured idle and under load.
  - Best · two 96 GB cards (2x RTX PRO 6000 Blackwell 96 GB): Adds a DeepSeek-V4-Flash reference so endpoints selling that model can be audited.
  - Wanted · references for the most-used open models (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). Estimate.): Reference runs of GLM-5.3-Flash and DeepSeek-V4.1-Flash, the two most-used open models, so endpoints that sell them can be audited against the real thing. Neither fits the reference box today. Not served yet.
- Buyer facts:
  - Your provider's API key: Held in memory for one run only: never logged, stored or written into the report. The probes bill your provider (about 23-43 short requests plus the context probe).
  - Data retention: Signed reports are stored on the server and readable by anyone with the report id; a report of your endpoint records its URL and model name, never the key.
  - Hardware: CPU only: comparison data ships as signed reference fixtures. Recording new fixtures needs the model's pinned GPU stack.
  - Typical hosted cost: A fraction of a cent of gateway time for a short audit of our own hosted route (measured). Each run shows its own measured cost.
- Samples:
  - 1. Hosted (`auditor-hosted`): https://decosa.ai/tools/developer/auditor?sample=1&autorun=0
  - 2. Direct (`auditor-direct`): https://decosa.ai/tools/developer/auditor?sample=2&autorun=0
  - 3. Swap (`auditor-swap`): https://decosa.ai/tools/developer/auditor?sample=3&autorun=0
  - 4. Quant (`auditor-quant`): https://decosa.ai/tools/developer/auditor?sample=4&autorun=0
- Regulatory note: Auditing a third-party endpoint uses your own key and credits with that provider: the hosted auditor holds the key in memory for one run and never stores or logs it. Check your provider's terms before publishing results; reports are private by default. Model licences: Apache-2.0 (Qwen3.8-27B, Qwen3.5-4B-Base).

### Make a disclosed UGC ad (`ugc`)

Short UGC-style product videos with an original AI presenter. Labelled as AI ads on screen and in a content credential. Fake reviews and claims without evidence are refused.

- What it is: Short UGC-style product videos with an original AI presenter: disclosed on screen and in a content credential, with no fake reviews and no claims without evidence.
- Who it's for: Teams in sales and marketing and creative and media.
- Page: https://decosa.ai/studio/brand
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/ugc):
  - receipt id recovered in all 6 checks (6/7 to 7/7 frames vote): Watermark survival on the two example ads (metadata stripped / CRF 28 re-encode / 50% resize) (synthetic, n = 6)
  - flagged 1 of 2 clips a person rejected, 0 of 10 approved clips: Brand safety, automated check (RapidOCR rule) on 12 H3 clips (dev (tuned on), n = 12)

- Status: preview. Models: Qwen3.8-27B · MiniMax H3 via fal · Kokoro-82M. Where: Hosted or self-host.
- Verification: Receipted lanes; signed render receipt (receipt coverage: partial). Hosted QA 2026-09-25: partial. Self-host QA 2026-09-25: partial.
- Hardware: Hosted renders on fal (no local GPU); self-host: 1× RTX PRO 6000 (96 GB)
- Output: A credentialed MP4 with a burned-in AI-ad label and end card.
- Page: https://decosa.ai/studio/brand · JSON: https://decosa.ai/use-cases/ugc.json
- Prompts: https://decosa.ai/prompts/ugc-hosted.md · https://decosa.ai/prompts/ugc-selfhost.md · https://decosa.ai/prompts/ugc-assemble.md · https://decosa.ai/prompts/ugc-mac.md
- Tiers:
  - Lite · fully open (Apache-2.0), on your own GPU (1× 48 GB card (lanes and renders take turns), or 1× RTX PRO 6000 96 GB): Wan2.1-T2V-14B shots with the same planning, disclosure and provenance steps. Silent B-roll only and visibly lower quality than H3 or LTX-2.3: fine for drafts and for a stack with no third-party model or service, not for a finished presenter ad.
  - Standard · the hosted default, MiniMax H3 via fal at cost (No GPU for the render (fal); the lane model runs on the gateway): Gateway lanes with receipts, a director pass, then H3 Max on fal: lip-synced presenter, consistent B-roll, 1080x1920. Billed at fal's list price with no markup; renders need an API key.
  - Best for self-host · MiniMax H3 (licence pending) or LTX-2.3 on your own GPU (1× RTX PRO 6000 96 GB): Joint audio-video for lip-synced presenters without a third-party service. Its licence keeps it off the hosted service; the disclosure and provenance layer is the same. Or run the open H3 weights yourself (1× 96 GB GPU + ~115 GB RAM), the same model as the hosted default; its licence doesn't currently cover the US/EU/UK/KR, so check it applies to you.
  - Wanted · MiniMax H3 on two cards, no offload (2x RTX PRO 6000 96 GB, no CPU offload (about 133 GB of BF16 weights; estimate)): Presenter ads take five H3 calls each, so throughput matters. Both halves of MiniMax-H3 in BF16 on two 96 GB cards, so nothing is offloaded to RAM: faster renders (estimate, not measured). Self-host now where H3's current licence covers you; once the pending licence lands we add this compute for the hosted tier ourselves. Not a community-provider model.
- Buyer facts:
  - Data retention: Plans (brief, script, claims, receipt ids) are kept on the server as JSON so a render can refer to them; a rendered ad opens only through an expiring signed link given to the key that made it. No personal data is needed: presenters are invented. On fal (hosted video only): fal keeps request data (prompts and file links) for 30 days by default and serves uploaded and generated files from public CDN links until they are deleted [V: fal docs, Data Retention & Storage, read 29 Sep 2026]; Decosa downloads each result at once, and fal's no-storage and short-expiry options are not switched on yet.
  - What leaves the box: Hosted: the brief goes to the open model through our gateway; hosted renders send the director's shot prompts, two frames of the invented presenter (itself made on fal) and the house-voice voice-over (a stock Kokoro voice, no real person's voice) to fal, a partner GPU provider, under Decosa's account. Self-host: nothing, unless you set the H3 backend with a fal key.
  - Inputs and limits: JSON brief: brand (60 characters), product (80), description (600), up to 6 key points and 8 evidence entries, 2-3 shots, voices af_heart or am_michael, platform tiktok, reels or shorts.
  - Outputs: 720x1248 MP4 (H.264 + AAC) labelled on every frame, ending on a disclosure card, with a C2PA credential, a signed render receipt and a per-frame watermark.
- Samples:
  - 1. Tidepot (`tidepot`): https://decosa.ai/studio/brand?sample=1&autorun=0
  - 2. Lumo loop (`lumo-loop`): https://decosa.ai/studio/brand?sample=2&autorun=0
  - 3. Fernway (`fernway`): https://decosa.ai/studio/brand?sample=3&autorun=0
  - 4. Refused testimonial (`refused-testimonial`): https://decosa.ai/studio/brand?sample=4&autorun=0
- Regulatory note: US: the FTC rule on consumer reviews and testimonials (16 CFR Part 465, in force 21 Oct 2024) bans reviews and testimonials that misrepresent that they come from a real consumer with real experience, including AI-generated ones; the FTC's Q&A says it does not ban virtual influencers or obviously fictional dramatizations. So this tool never writes a customer testimonial, rating or review. The FTC Endorsement Guides (16 CFR Part 255, revised 2023) cover fictitious and virtual endorsers and require clear and conspicuous disclosure. Objective claims need substantiation before they run. Platforms also require AI labels on realistic synthetic content (turn them on when posting). EU AI Act Art. 50(2) machine-readable marking is covered by the C2PA credential and watermark. Credentials use a development certificate (issuer untrusted in public validators). This is a product policy, not legal advice; the advertiser stays responsible for its claims.

### Catch AI answers your sources don't back (`grounding`)

Checks every sentence of a draft against the sources you give it: supported, partial, unsupported or contradicted, with the source text it rests on. A claims gate can stop an email or page before it ships.

- What it is: Is each sentence supported by the sources? Per-sentence verdicts with the source span, a claims gate, and a signed report.
- Who it's for: Teams that ship AI-written answers, emails or FAQs and want each sentence checked against its sources first.
- Page: https://decosa.ai/tools/developer/grounding
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/grounding):
  - 0.593 / 0.556: Unsupported-sentence precision / recall, default gate (held out, n = 2069)
  - 0.966: Recall, strict gate (partial counts too) (held out, n = 2069)
  - 0.691: Response-level F1, default gate (held out, n = 300)
  - 17.2 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Hosted or self-host.
- Verification: Receipt per verdict; signed report (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the judge; the checker itself runs on CPU
- Output: Per-sentence verdicts with source spans, a gate decision and a signed report.
- Page: https://decosa.ai/tools/developer/grounding · JSON: https://decosa.ai/use-cases/grounding.json
- Prompts: https://decosa.ai/prompts/grounding-hosted.md · https://decosa.ai/prompts/grounding-selfhost.md · https://decosa.ai/prompts/grounding-assemble.md · https://decosa.ai/prompts/grounding-mac.md
- Tiers:
  - Lite · CPU only, no GPU (Any Linux or macOS machine with Python 3.11+): The checker plus an NLI cross-encoder. Runs anywhere; far less accurate, and no model receipts.
  - Standard · one GPU for the judge (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B judges each sentence in its own receipted call. This is what the hosted API runs.
  - Wanted · a panel of the largest open judges (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). The 27B stays on one card. Estimate.): DeepSeek-V4.1-Flash and GLM-5.3-Flash judge each sentence beside Qwen3.8-27B. A sentence counts as supported only when the panel agrees, and disagreements are shown. Not served yet.
- Alternates (smaller models for the same job):
  - Fast option-token judge (1x RTX PRO 6000 96 GB (BF16 weights are 49 GB)), not served: A 4B-active judge read through option-token probabilities: cheaper per call, not better. Page 32 suggests the dense Qwen3.5-9B (Apache-2.0) instead. The real blocker is logprob passthrough on the gateway.
- Buyer facts:
  - Data retention: Nothing stored: text and sources live in memory for the request. The signed report carries hashes and character offsets, not your text.
  - Inputs: Text up to 8,000 characters and 40 sentences; up to 10 sources and 60,000 characters in total, as text (files are read in your browser on the console).
  - Typical hosted cost: A fraction of a cent for a short answer against a one-page source at the gateway's list price (measured). Each run shows its own measured cost.
  - What leaves the box (self-host): Nothing: the judge runs on your vLLM (or on CPU with the NLI option) and reports are signed with this box's key.
- Samples:
  - 1. Outreach email (`outreach-email`): https://decosa.ai/tools/developer/grounding?sample=1&autorun=0
  - 2. Bakery faq (`bakery-faq`): https://decosa.ai/tools/developer/grounding?sample=2&autorun=0
  - 3. Support answer (`support-answer`): https://decosa.ai/tools/developer/grounding?sample=3&autorun=0
  - 4. Board summary (`board-summary`): https://decosa.ai/tools/developer/grounding?sample=4&autorun=0
- Regulatory note: A grounding check is a quality control, not a fact check: it reads only the sources you send, and a sentence copied from a wrong source passes. Verdicts come from a language model and can be wrong in both directions (see the RAGTruth numbers below). The hosted demo keeps no text: drafts and sources stay in memory for the request, the report holds hashes only, and logs carry counts. For unpublished or confidential drafts, self-host. Where rules require substantiating marketing claims (for example the FTC's advertising substantiation policy in the US), a pass here is a record that a check ran against the sources you chose, not proof that a claim is substantiated. Not legal advice. Model licences: Apache-2.0 (Qwen3.8-27B; cross-encoder/nli-deberta-v3-base). Checked 24 Sep 2026.

### Digest a deposition (`deposition`)

A page:line digest of deposition or hearing transcripts. Every sentence cites its lines and is checked against them; with two or more transcripts it finds testimony that conflicts, with both cites.

- What it is: A page:line digest where every sentence is checked against the lines it cites, plus the places where witnesses contradict each other.
- Who it's for: Litigation associates, paralegals and litigation-support teams.
- Page: https://decosa.ai/legal/deposition
- Last reviewed: 2026-09-30
- Key numbers (details and caveats: https://decosa.ai/metrics/deposition):
  - 9 / 9 in both runs: Contradiction finder: planted conflicts found (held-out set) (test split, n = 9)
  - 2 / 1 (precision 0.82 / 0.90): Contradiction finder: false-positive flags, run 1 / run 2 (test split, n = 5)
  - 33 / 40: Digest sentences fully supported by their cited lines (hand-checked) (dev (tuned on), n = 40)
  - 40.1 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: live. Models: Qwen3.8-27B. Where: Self-host (transcripts stay with the matter).
- Verification: Every cite checked; receipted calls; signed record (receipt coverage: full). Hosted QA 2026-09-30: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB); no GPU for the parser, flags and export
- Output: A checked digest with page:line cites, exported to Word or Markdown, and a signed record.
- Page: https://decosa.ai/legal/deposition · JSON: https://decosa.ai/use-cases/deposition.json
- Prompts: https://decosa.ai/prompts/deposition-hosted.md · https://decosa.ai/prompts/deposition-selfhost.md · https://decosa.ai/prompts/deposition-assemble.md · https://decosa.ai/prompts/deposition-mac.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; digest and check quality unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B writes the digest, checks every cite and judges contradiction pairs; every call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger writer and checker for many-witness matters.
  - Wanted · GLM-5.3-Flash on your own hardware (Your own hardware: 2x 96 GB cards (NVFP4, about 170-186 GB, unconfirmed) or a Mac with 192 GB or more (MLX 4-bit, 165 GB). Estimate; GLM's SGLang SM120 build hung on our server.): A larger model from a different family for the cross-witness contradiction search and long records. Transcripts are privileged, so it runs on your hardware, never on community providers. Not served yet.
- Buyer facts:
  - Measured latency: Hosted p50 is the whole digest of the fictional two-deposition pair (117 lines), over 3 runs on 2026-09-25, gateway route.
  - Typical run cost: A few cents or less for the fictional pair (a few dozen model calls) at the gateway list price. Each run shows its own measured cost.
  - Data retention: Transcripts and digests live in server memory for one hour (for export) and are never written to disk or logs.
  - What leaves the box (self-host): Nothing: keep DECOSA_LLM_ROUTE=direct.
  - Input: Text or PDF transcripts with page:line numbers (numbered, condensed or plain); up to 4 transcripts and 4,000 lines per digest (8,000 in the self-host compose).
- Samples:
  - 1. Fictional fall (`fictional-fall`): https://decosa.ai/legal/deposition?sample=1&autorun=0
  - 2. Svb hearings (`svb-hearings`): https://decosa.ai/legal/deposition?sample=2&autorun=0
- Regulatory note: Not a certified transcript and not legal advice; the digest is a work aid and a lawyer still reads the transcript. The check shows each sentence is supported by the lines it cites; it does not show the digest covers everything important. ABA Formal Opinion 512 (29 Jul 2024): lawyers must understand how a generative AI tool uses client information, get informed consent before putting it into a self-learning tool (boilerplate in an engagement letter is not enough), and bill only for time actually spent. United States v. Heppner (S.D.N.Y., ruled 10 Feb 2026, opinion 17 Feb 2026) held a defendant's exchanges with a consumer AI service not privileged, partly because the provider was a third party whose terms allowed disclosure; a self-hosted tool used at counsel's direction avoids that third-party problem but does not by itself make anything privileged. Check any protective order before hosted use. The audio path gives an uncertified rough transcript only; whether AI-assisted transcripts of recorded depositions are admissible is being litigated (In re Hughey, Tex., pending as of Sep 2026).

### Check a brief before filing (`filing-preflight`)

One check before a brief is filed: every cited case, statute and rule exists, the cited opinion says what the brief claims, record cites resolve to the excerpts, quotations are verbatim, and no full SSN, birth date, account number or minor's name is left in. Ends in a signed record a partner can review.

- What it is: One check before a brief is filed: authorities exist and say what the brief claims, record cites resolve, quotations are verbatim, and Rule 5.2 identifiers are gone. Ends in a signed record.
- Who it's for: Teams in legal.
- Page: https://decosa.ai/legal/filing-preflight
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/filing-preflight):
  - 89%: Fake citations caught as problem (strict) (test split, n = 19)
  - 92%: Citations for a holding the case does not contain, caught (test split, n = 13)
  - 64%: One-word misquotations caught as problem (strict) (test split, n = 22)
  - 12.1 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for drafts; hosted for public filings.
- Verification: Exact-match quotes; receipt per proposition; signed record (receipt coverage: full). Hosted QA 2026-09-29: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the judge; parsing, lookups, quotes and the privacy check run on CPU
- Output: A checklist with every problem located in the brief, a Markdown review for the partner, and a signed record.
- Page: https://decosa.ai/legal/filing-preflight · JSON: https://decosa.ai/use-cases/filing-preflight.json
- Prompts: https://decosa.ai/prompts/filing-preflight-hosted.md · https://decosa.ai/prompts/filing-preflight-selfhost.md · https://decosa.ai/prompts/filing-preflight-assemble.md · https://decosa.ai/prompts/filing-preflight-mac.md
- Tiers:
  - Lite · CPU only, no GPU (Any Linux or macOS machine with Python 3.11+): Existence, quotations, record pages and the privacy patterns. No proposition check, and minors' names are only flagged for review.
  - Standard · one GPU for the judge (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Adds the proposition and record-support checks with Qwen3.8-27B, one receipted call each. This is what the hosted API runs.
  - Wanted · a GLM-5.3-Flash judge on your own hardware (Your own hardware: 2x 96 GB cards (NVFP4, about 170-186 GB, unconfirmed) or a Mac with 192 GB or more (MLX 4-bit, 165 GB). Estimate; GLM's SGLang SM120 build hung on our server.): A stronger open judge from another family for propositions, where the 27B is least sure. Briefs can be privileged, so it runs on your hardware, never on community providers. Not served yet.
- Alternates (smaller models for the same job):
  - A second judge on one card (1x RTX PRO 6000 96 GB (80 GB of weights)), not served: A cross-family second judge that fits beside the 27B, so we would host it ourselves. sm_120 support unconfirmed.
- Buyer facts:
  - Data retention: Hosted: the brief and excerpts live in memory; an uploaded file's text and a finished run are kept for one hour (for the export links), then dropped. Logs carry counts only.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and its receipt (hashes, token counts, no text) is kept by the gateway and this API. Self-hosted on the direct route: the brief never leaves the box; case lookups send only the citation to CAP, CourtListener, LII and uscode.house.gov; offline mode sends nothing and leaves case items unverified.
  - Input formats: Brief as text, PDF or DOCX (up to 15 MB; 80,000 characters checked); up to 8 record excerpts as text with [[page]] markers or PDFs that keep their pages.
  - Model calls per brief: One receipted judge call per proposition and per record cite; citation lookups use free public case-law data. The measured cost per brief is the cost per task at the top of the page.
- Samples:
  - 1. Demo planted (`demo-planted`): https://decosa.ai/legal/filing-preflight?sample=1&autorun=0
  - 2. Osg case v montana (`osg-case-v-montana`): https://decosa.ai/legal/filing-preflight?sample=2&autorun=0
- Regulatory note: A pre-flight check, not legal advice and not a substitute for reading the authorities: under Fed. R. Civ. P. 11(b)(2) the lawyer who signs still certifies that the legal contentions are warranted by existing law, and courts have sanctioned lawyers for filing citations an AI tool invented (Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023); Park v. Kim, 91 F.4th 610 (2d Cir. 2024); Wadsworth v. Walmart Inc., No. 2:23-cv-00118 (D. Wyo. Feb. 24, 2025)). Many judges' standing orders now require disclosure of AI use or a certification that a person checked every citation; this tool does not write that certification, and its signed record is evidence that a check ran, not the certification. It does not check whether a case is still good law. Privacy: Fed. R. Civ. P. 5.2(a) (and Fed. R. Crim. P. 49.1, Fed. R. Bankr. P. 9037) allow only the last four digits of Social Security, taxpayer and financial-account numbers, the year of birth and a minor's initials; the check finds likely identifiers but cannot know which Rule 5.2(b) exemptions apply. Confidentiality: a draft brief is client information and may be work product; ABA Formal Opinion 512 (29 Jul 2024) asks lawyers to understand how a tool uses what they put in and to protect it, so self-host drafts (offline mode sends nothing out). The hosted demo is for public filings and the fictional sample, and sends only citations to the public sources. Model licence: Apache-2.0 (Qwen3.8-27B). Checked 25 Sep 2026.

### Pre-check promo claims for MLR (`promo-claims-check`)

Checks every claim in a drug, device or supplement promotional piece against its label and references before MLR review: unsupported or off-label claims, missing fair balance, comparisons without head-to-head data, and supplement disease claims. Returns a signed pre-review packet.

- What it is: Every claim in a drug, device or supplement piece checked against its references before MLR review, with the span behind it and what is wrong.
- Who it's for: Teams in healthcare and compliance and trust.
- Page: https://decosa.ai/tools/life-sciences/promo-claims-check
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/promo-claims-check):
  - 20 / 20: Planted problems caught (all categories) (test split, n = 20)
  - 2 / 2: Piece-level problems caught (fair balance, DSHEA) (test split, n = 2)
  - 1 / 63 (1.6%): Unplanted sentences with an issue (false positives) (test split, n = 63)
  - 3.9 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for unpublished data; hosted for public labels.
- Verification: Receipt per model call; signed packet (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for the model; extraction, rules and the packet run on CPU
- Output: An annotated MLR pre-review packet and a signed JSON record of every finding.
- Page: https://decosa.ai/tools/life-sciences/promo-claims-check · JSON: https://decosa.ai/use-cases/promo-claims-check.json
- Prompts: https://decosa.ai/prompts/promo-claims-check-hosted.md · https://decosa.ai/prompts/promo-claims-check-selfhost.md · https://decosa.ai/prompts/promo-claims-check-assemble.md · https://decosa.ai/prompts/promo-claims-check-mac.md
- Tiers:
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B does the grounding, the claim review and the head-to-head and fair-balance checks, each in its own receipted call. This is what the hosted API runs.
- Alternates (smaller models for the same job):
  - Layout-aware check (1x RTX PRO 6000 96 GB (estimate)), not built: Read the rendered page as well as its text with the 27B's own vision input, so fair balance can account for type size and placement. Not built.
- Buyer facts:
  - Data retention: Nothing stored: the piece and references live in memory for the request. The signed packet holds hashes, offsets, finding codes and receipt ids, never the copy.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and its receipt (hashes, token counts, no text) is kept by the gateway and this API. Self-hosted on the direct route: nothing leaves the box.
  - Input formats: Text only: piece up to 8,000 characters and 40 claim sentences; up to 10 references (label, studies, other) and 200,000 characters; up to 60 claims-matrix entries. Extract PDF text yourself.
  - Typical run: The metformin sample: a dozen or so model calls and a few cents at the gateway list price; the supplement sample about the same. Each run shows its own measured cost.
- Samples:
  - 1. Metformin planted (`metformin-planted`): https://decosa.ai/tools/life-sciences/promo-claims-check?sample=1&autorun=0
  - 2. Metformin compliant (`metformin-compliant`): https://decosa.ai/tools/life-sciences/promo-claims-check?sample=2&autorun=0
  - 3. Magnesium planted (`magnesium-planted`): https://decosa.ai/tools/life-sciences/promo-claims-check?sample=3&autorun=0
  - 4. Magnesium compliant (`magnesium-compliant`): https://decosa.ai/tools/life-sciences/promo-claims-check?sample=4&autorun=0
- Regulatory note: Checked 25 Sep 2026. This is a pre-check that helps a medical-legal-regulatory (MLR) committee; it approves nothing, and the company and its reviewers stay responsible for every piece. What it checks against: FDA's prescription drug advertising rules (21 CFR 202.1: not false or misleading, fair balance of benefit and risk, only uses in the approved labeling), which FDA's Office of Prescription Drug Promotion enforces; FDA's 2018 guidance on communications consistent with the FDA-required labeling; the DSHEA rules for supplement claims (21 CFR 101.93: structure/function claims need the disclaimer in 101.93(c), disease claims are not allowed); and the FTC's Health Products Compliance Guidance (December 2022), which expects competent and reliable scientific evidence that fits each claim. It reads text only, so the visual prominence of risk information is not assessed, and a claim the references support can still be misleading in context. Findings come from a language model and can be wrong in both directions (see the eval). Unpublished study data and pre-launch pieces should stay on your own hardware: self-host. The hosted demo keeps nothing (piece and references stay in memory for the request; the packet holds hashes; logs carry counts) and should be used with public labels and synthetic or published copy only. Model licence: Apache-2.0 (Qwen3.8-27B). Not legal or regulatory advice.

### Check if an open model can take over your prompt (`migration-check`)

Paste a production prompt and the outputs your closed model already gave. The same inputs run on an open model; you get agreement per example with an interval, the failure clusters, cost per 1,000 requests, and a go or no-go in a signed record.

- What it is: Can an open model take over this prompt? A shadow run against the outputs you already logged, with a go / no-go in a signed record.
- Who it's for: Teams in software and ai ops.
- Page: https://decosa.ai/tools/developer/migration-check
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/migration-check):
  - 79.3% (74.4-83.5): Not-worse agreement with human experts, both orders (run 1) (test split, n = 300)
  - 0.588: Cohen's kappa vs experts (run 1) (test split, n = 300)
  - 0.755 / 0.839: Precision / recall on "worse" (run 1) (test split, n = 300)
  - 3.6 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B (candidate and judge). Where: Hosted or self-host.
- Verification: Receipt per output and judgment; signed record that recomputes (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the candidate and judge; scoring runs on CPU
- Output: A go / no-go with failure clusters, cost per 1,000 requests and a signed record anyone can re-check.
- Page: https://decosa.ai/tools/developer/migration-check · JSON: https://decosa.ai/use-cases/migration-check.json
- Prompts: https://decosa.ai/prompts/migration-check-hosted.md · https://decosa.ai/prompts/migration-check-selfhost.md · https://decosa.ai/prompts/migration-check-assemble.md · https://decosa.ai/prompts/migration-check-mac.md
- Tiers:
  - Lite · structured prompts on a 24-32 GB card (self-host) (1x RTX 4090 24 GB or RTX 5090 32 GB (estimate)): Exact, label and JSON prompts need no judge: compare a small open model such as Gemma-4-31B against your logs. No free-text judging on this tier.
  - Standard · Qwen3.8-27B as candidate and judge (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): One pinned open model runs your prompt and judges free text in both orders. This is what the hosted API runs.
  - Best · DeepSeek-V4-Flash as the candidate (two 96 GB cards, self-host) (2x RTX PRO 6000 96 GB with community vLLM patches): A 284B-parameter MoE for prompts the 27B cannot carry. Uses both cards, so the judge needs another box or runs between passes.
  - Wanted · the largest open candidates (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). The 27B stays on one card. Estimate.): Check a migration against DeepSeek-V4.1-Flash and GLM-5.3-Flash, the most-used open models, with Qwen3.8-27B as the judge. Not served yet.
- Buyer facts:
  - Data retention: Nothing stored: prompts, inputs and outputs live in memory for the request and come back to you in the sealed record. Strip each entry's text to share the record without your data; it still verifies.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and its receipt (hashes, token counts, no text) is kept by the gateway and this API. Self-hosted on the direct route: nothing leaves the box. It never calls a closed API and never needs your closed-model key.
  - Input formats: JSON: a prompt template and 10 to 200 logged examples (up to 40 with a demo session), each with the current model's output and optionally a human label and logged tokens and latency. Tasks: exact, label, JSON (schema) or free text (judge).
  - Typical run: JSON tickets: a fraction of a cent at the gateway list price. Free-text answers with the judge in both orders: more calls, a few cents. Each run shows its own measured cost.
- Samples:
  - 1. Mtbench gpt4 answers (`mtbench-gpt4-answers`): https://decosa.ai/tools/developer/migration-check?sample=1&autorun=0
  - 2. Mtbench gpt4 judge (`mtbench-gpt4-judge`): https://decosa.ai/tools/developer/migration-check?sample=2&autorun=0
  - 3. Tickets json (`tickets-json`): https://decosa.ai/tools/developer/migration-check?sample=3&autorun=0
- Regulatory note: A migration check compares outputs; it does not certify a model. Agreement with your current model is not correctness, the free-text judge is a language model and can be wrong, and the interval covers only inputs like the ones you sent. Closed-API terms can restrict what you do with outputs: Anthropic's Commercial Terms (section D.4, effective 17 Jun 2025) forbid using the services 'to train competing AI models'. This check trains nothing and only compares, but read your own provider's terms. Production logs often hold personal or customer data, which data-protection law (GDPR, CCPA and others) and your customer contracts govern: the hosted demo keeps nothing, and real logs belong on a self-hosted box. Not legal advice. Model licences: Apache-2.0 (Qwen3.8-27B, Gemma-4-31B-it), MIT (DeepSeek-V4-Flash-0731). Prices and terms checked 25 Sep 2026.

### Classify with a confidence you can act on (`typed-judgment`)

Yes/no, one-of-N, score and label questions that come back as typed answers with a calibrated probability, a short reason and a receipt. It runs on pinned open weights, so a request can be re-run next quarter against the same model, and confident verdicts reproduce. An eval mode grades an output against a rubric or compares two.

- What it is: A calibrated LLM judge API: send a context and up to 16 narrow questions, and each comes back as a typed answer (yes or no, one of N, a score or labels) with a probability you can set a threshold on, a one-line reason and a receipt, on pinned open weights so the same request can be re-run later.
- Who it's for: Teams in software and ai ops.
- Page: https://decosa.ai/tools/developer/typed-judgment
- Last reviewed: 2026-09-25
- Key result: On 1,000 held-out BoolQ questions it scored 90.7% with expected calibration error 0.023 (logprobs, self-host route); the hosted route, which samples, scored 90.5% with ECE 0.032 on 200. Calibration was measured only on public benchmarks and varies by domain, so check it on your own labelled data.
- Key numbers (details and caveats: https://decosa.ai/metrics/typed-judgment):
  - 90.7% / 0.023: BoolQ accuracy / ECE, logprobs (direct route) (held out, n = 1000)
  - 83.4% / 0.026: MMLU accuracy / ECE, logprobs (direct route) (held out, n = 1000)
  - 90.5% / 84.0%: Hosted route, samples k=4: BoolQ / MMLU accuracy (held out, n = 200)
  - 4.1 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Hosted or self-host.
- Verification: Receipt per model call; signed record with the weights, prompt and seed (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; the API runs on CPU
- Output: Typed answers with probabilities and a signed record you can re-run and compare.
- Page: https://decosa.ai/tools/developer/typed-judgment · JSON: https://decosa.ai/use-cases/typed-judgment.json
- Prompts: https://decosa.ai/prompts/typed-judgment-hosted.md · https://decosa.ai/prompts/typed-judgment-selfhost.md · https://decosa.ai/prompts/typed-judgment-assemble.md · https://decosa.ai/prompts/typed-judgment-mac.md
- Tiers:
  - Lite · one call per question (stated confidence) (Hosted; or 1x RTX 5090 32 GB (estimate) / 1x RTX PRO 6000 96 GB (measured) self-hosted): The cheapest hosted setting (method single): the same answers, but the probability is the model's stated confidence mapped on a dev set, which separates right from wrong answers poorly.
  - Standard · hosted, answer plus 4 seeded samples (Hosted (1x RTX PRO 6000 96 GB behind the gateway)): What the hosted API runs by default (method samples, k=4): five calls per question, probabilities from smoothed votes. Better calibrated than lite, coarser than logprobs.
  - Best · self-host with logprobs (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): On your own box the engine reads the answer token's log-probabilities: one call per question and the best-calibrated probabilities in our eval. Hosted too once the gateway passes logprobs through.
  - Wanted · two large judges that must agree (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). The 27B stays on one card. Estimate.): DeepSeek-V4.1-Flash and GLM-5.3-Flash answer beside Qwen3.8-27B; when the families disagree the answer is flagged instead of returned with one model's probability. Not served yet.
- Alternates (smaller models for the same job):
  - Fast option-token judge (1x RTX PRO 6000 96 GB (BF16 weights are 49 GB)), not served: A 4B-active judge read through option-token probabilities: cheaper per call, not better. Page 32 suggests the dense Qwen3.5-9B (Apache-2.0) instead. The real blocker is logprob passthrough on the gateway.
- Buyer facts:
  - Data retention: Nothing stored: contexts and questions live in memory for the request. The signed record holds hashes of the context and questions plus the answers, not your text.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and its receipt (hashes, token counts, no text) is kept by the gateway and this API. Self-hosted on the direct route: nothing leaves the box.
  - Input formats: JSON: a context (text or any JSON, up to 24,000 characters) and 1-16 typed questions (yes/no, one of up to 26 options, a 2-9 level score, or up to 12 labels); or a rubric or a pair of outputs to grade; batches of up to 50 items.
  - Typical run: The support ticket sample: a few dozen calls and a fraction of a cent with repeated samples, or a handful of calls and less with stated confidence, at the gateway list price.
- Samples:
  - 1. Support triage (`support-triage`): https://decosa.ai/tools/developer/typed-judgment?sample=1&autorun=0
  - 2. Agent done check (`agent-done-check`): https://decosa.ai/tools/developer/typed-judgment?sample=2&autorun=0
  - 3. Lead fit (`lead-fit`): https://decosa.ai/tools/developer/typed-judgment?sample=3&autorun=0
  - 4. Eval rubric (`eval-rubric`): https://decosa.ai/tools/developer/typed-judgment?sample=4&autorun=0
  - 5. Eval pairwise (`eval-pairwise`): https://decosa.ai/tools/developer/typed-judgment?sample=5&autorun=0
  - 6. Moderation (`moderation`): https://decosa.ai/tools/developer/typed-judgment?sample=6&autorun=0
- Regulatory note: The answers are model outputs with estimated probabilities, not facts or decisions: keep the decision rule and any human review in your own code. Calibration was measured on public benchmarks (BoolQ, MMLU, SummEval, MT-Bench) and varies by domain, so check it on your own labelled data before a threshold decides anything that matters. Where a judgment feeds a decision about a person (hiring, credit, housing, insurance), rules such as NYC Local Law 144 (in force), Illinois HB 3773 (in force 1 Jan 2026), Colorado's AI Act (effective date moved to 1 Jan 2027) and the EU AI Act's high-risk duties (Annex III uses from 2 Dec 2027) can apply to you as the deployer; the signed records and receipts help with record-keeping but are not a compliance programme. The hosted demo keeps no context: it stays in memory for the request, the record holds hashes only, and logs carry counts. For personal or confidential data, self-host. Not legal advice. Model licence: Apache-2.0 (Qwen3.8-27B). Dataset licences for the eval: BoolQ CC BY-SA 3.0, MMLU MIT, SummEval MIT, MT-Bench human judgments CC BY 4.0. Checked 25 Sep 2026.

### Answer a security questionnaire (`security-questionnaire`)

Fills a buyer's security questionnaire from answers you have already approved. The model chooses and may trim, but anything it changes is checked against the source, stale answers are caught against your current policies, and questions with nothing approved behind them are left for a person.

- What it is: Fills a buyer's security questionnaire from answers you have already approved, checks every edit against its source, and sends the rest to a person.
- Who it's for: Teams in compliance and trust.
- Page: https://decosa.ai/tools/finance/security-questionnaire
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/security-questionnaire):
  - 0.970 (64 of 66): Fill precision (test split, n = 66)
  - 0.984 (63 of 64): Coverage (answerable questions filled correctly) (test split, n = 64)
  - 0.969 (31 of 32): Abstention (unanswerable questions sent to a person) (test split, n = 32)
  - 63.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Self-host (your policies stay on your machine) or hosted.
- Verification: Receipt per selection and check; signed review record (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; parsing, checks and export run on CPU
- Output: The filled sheet (XLSX or CSV) with a source per answer, the questions for a person, and a signed review record.
- Page: https://decosa.ai/tools/finance/security-questionnaire · JSON: https://decosa.ai/use-cases/security-questionnaire.json
- Prompts: https://decosa.ai/prompts/security-questionnaire-hosted.md · https://decosa.ai/prompts/security-questionnaire-selfhost.md · https://decosa.ai/prompts/security-questionnaire-assemble.md · https://decosa.ai/prompts/security-questionnaire-mac.md
- Tiers:
  - Lite · one 32 GB card, no policy check (1x RTX 5090 32 GB (estimate)): The same selector and edit check with the source check off (source_check: false): about half the model calls, but stale approved answers are filled instead of flagged.
  - Standard · one GPU, selection plus both checks (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B chooses the answers and judges every changed sentence and every cited policy, each in its own receipted call. This is what the hosted API runs.
  - Wanted · the largest open models, long context (Network providers: an 8x H200-class node for DeepSeek-V4.1-Flash (476 GB of weights); 2x 96 GB cards or a Mac with 192 GB or more for GLM-5.3-Flash (about 170 GB). The 27B stays on one card. Estimate.): DeepSeek-V4.1-Flash selects and drafts answers with much more of the evidence library in context, and GLM-5.3-Flash checks each one. Both models support 1M tokens (vendor); the proposed network need serves 131k. Not served yet.
- Alternates (smaller models for the same job):
  - Fast option-token selector (1x RTX PRO 6000 96 GB (BF16 weights are 49 GB)), not served: A 4B-active judge read through option-token probabilities: cheaper per call, not better. Page 32 suggests the dense Qwen3.5-9B (Apache-2.0) instead. The real blocker is logprob passthrough on the gateway.
- Buyer facts:
  - Data retention: Nothing stored: the library and the questionnaire live in memory for the request. The review record holds hashes, source ids and statuses, never text.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and its receipt (hashes, token counts, no text) is kept by the gateway and this API. Self-hosted on the direct route: nothing leaves the box. A library describes your security posture: for real policies, self-host.
  - Input formats: Questionnaire as XLSX (first sheet with a Question header), CSV, pasted text or JSON; library as approved answers (CSV, XLSX or JSON) plus policy documents as text. Out: XLSX or CSV. Up to 60 questions per demo run, 200 with a key.
  - Typical run: The sample sheet against its library: a few dozen model calls, a few cents at the gateway list price. Each run shows its own measured cost.
- Samples:
  - 1. Corvid freight (`corvid-freight`): https://decosa.ai/tools/finance/security-questionnaire?sample=1&autorun=0
- Regulatory note: Answers to a security questionnaire can become representations in a contract, so a person should read the sheet before it is sent; this tool fills only what your library already says and marks what needs sign-off, but the checks are model judgements and can be wrong. A trimmed answer can leave out a qualifying sentence, and the checks do not catch omissions. Your policies describe your security posture: self-host to keep them on your own machine. The hosted API keeps nothing (library and questionnaire stay in memory for the request; the review record holds hashes only; logs carry counts). Questionnaire templates carry their own licences: the CSA CAIQ is a free download from the Cloud Security Alliance after sign-in, but we found no licence that allows redistributing it, so the demo sheet is written from scratch in its shape. Not legal advice. Model licence: Apache-2.0 (Qwen3.8-27B). Checked 25 Sep 2026.

### Prove what your agent did (`flight-recorder`)

A signed, step-by-step record of what a browser or computer-use agent saw, decided and did. Replay it step by step; change one screenshot or one decision and verification names the step.

- What it is: A signed, step-by-step record of what a browser or computer-use agent saw, decided and did, that anyone can re-check.
- Who it's for: Teams in software and ai ops and compliance and trust.
- Page: https://decosa.ai/tools/developer/flight-recorder
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/flight-recorder):
  - 8/8: Held-out agent runs reaching the expected outcome (held out, n = 8)
  - 25.9%: MiniWoB++ success, production agent with guards (held out, n = 625)
  - 44.7% / 40.7%: Mind2Web element accuracy / step success (held out, n = 300)
  - 5.3 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B for decisions; any agent through the SDK. Where: Hosted or self-host.
- Verification: Receipted decisions; signed, hash-chained run record (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for the decision model; the recorder and browser run on CPU
- Output: A signed run record with thumbnails, receipts and guard events that anyone can re-verify.
- Page: https://decosa.ai/tools/developer/flight-recorder · JSON: https://decosa.ai/use-cases/flight-recorder.json
- Prompts: https://decosa.ai/prompts/flight-recorder-hosted.md · https://decosa.ai/prompts/flight-recorder-selfhost.md · https://decosa.ai/prompts/flight-recorder-assemble.md · https://decosa.ai/prompts/flight-recorder-mac.md
- Tiers:
  - Lite · record only, any CPU (Any Linux or macOS machine with Python 3.11+): Your agent and your model; the recorder chains, seals and verifies. Decisions from another model are recorded with their hash but carry no receipt.
  - Standard · receipted decisions, one 96 GB card (hosted demo) (1× RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B makes each decision through /decide, with a gateway receipt inside the step and guards before anything runs.
- Alternates (smaller models for the same job):
  - Typed decisions on Apple silicon (Jev-compatible) (Apple silicon with 32 GB or more), self-host only: DiffusionGemma as a local decision server for the macOS harness: fast single choices, weak at ordered sequences. Not a hosted model, so no receipts.
- Buyer facts:
  - Data retention: Hosted: runs and screenshots are kept 24 hours (DECOSA_FLIGHT_TTL_S), then deleted. Self-host: you set the TTL; sealed records verify offline, so archive them yourself.
  - What leaves the box: Hosted: whatever the agent posts (screenshots, page text, decisions), unless the SDK runs with send_images=False, which sends only screenshot hashes. Self-host: nothing, unless you point DECOSA_LLM_ROUTE at a gateway.
  - Inputs and limits: Screenshots as JPEG, PNG or WebP up to 3 MB each (4 MB of images per run; the server keeps a 320 px thumbnail and the hash), page text up to 16,000 characters per step, up to 300 steps per run.
  - Logs: The service log carries method, path, status and timing only; task text and screenshots are not logged (checked on the self-host box).
- Samples:
  - 1. Cheapest tool (`cheapest-tool`): https://decosa.ai/tools/developer/flight-recorder?sample=1&autorun=0
  - 2. Swap cart (`swap-cart`): https://decosa.ai/tools/developer/flight-recorder?sample=2&autorun=0
  - 3. Place order (`place-order`): https://decosa.ai/tools/developer/flight-recorder?sample=3&autorun=0
- Regulatory note: Checked 2026-09-25. EU AI Act: Article 12 requires high-risk AI systems to be able to record events automatically (logs) over their lifetime, and Articles 19 and 26(6) require providers and deployers to keep those logs for at least six months unless other law says otherwise. That duty applies only to high-risk systems (Annex I products, and Annex III uses such as hiring, credit scoring or access to essential services); after the AI Omnibus (Council approval 29 June 2026) Annex III duties apply from 2 December 2027 and Annex I from 2 August 2028. Most browser agents doing back-office work are not high-risk systems, and for them this record is useful evidence, not a legal requirement. Article 12 does not require signatures, hash chains or screenshots, and the harmonised logging standards (prEN 18229-1, ISO/IEC DIS 24970) are still drafts, so no product can claim conformity with them yet; this one does not. Screenshots of back-office screens usually contain personal data: under the GDPR, minimisation and storage limits apply, so use hash-only mode (images stay with you) or self-host. The hosted demo keeps runs for 24 hours and only visits a fictional shop. In a dispute a signature shows the record was not changed after signing and which key signed it; it does not show that what an agent reported about itself was true, and its weight as evidence is for the court or arbitrator. Model licence: Apache-2.0 (Qwen3.8-27B). Not legal advice.

### Run an end-to-end browser test (`test-runs`)

Write a test as plain-language steps with assertions. A browser agent carries out the steps; pass or fail comes only from the assertions, checked in code. Each run ends in a signed certificate you can keep as release evidence, and a failed assertion fails the build.

- What it is: Plain-language browser tests judged only by assertions in code, each run sealed into a signed certificate you can keep as release evidence.
- Who it's for: Teams in software and ai ops and compliance and trust.
- Page: https://decosa.ai/tools/developer/test-runs
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/test-runs):
  - 52/52: Verdict agrees with the scripted Playwright test, first run of each case (test split, n = 52)
  - 120/120: Verdict agrees, all runs (test split, n = 120)
  - 24/24: Seeded bugs and faulty users caught (test split, n = 24)
  - 86.1 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B for actions; assertions in code. Where: Hosted (verified domains) or self-host.
- Verification: Receipted actions; signed run certificate whose verdict recomputes (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for the action model; the runner, browser and verifier run on CPU
- Output: A signed run certificate: spec, target, build, every assertion's value and a verdict anyone can re-derive.
- Page: https://decosa.ai/tools/developer/test-runs · JSON: https://decosa.ai/use-cases/test-runs.json
- Prompts: https://decosa.ai/prompts/test-runs-hosted.md · https://decosa.ai/prompts/test-runs-selfhost.md · https://decosa.ai/prompts/test-runs-assemble.md · https://decosa.ai/prompts/test-runs-mac.md
- Tiers:
  - Lite · navigation-only specs, any CPU (Any Linux machine with 2 GB RAM per browser): Every step is a goto with assertions: no model, no GPU, still a signed certificate. For smoke checks of pages, not flows.
  - Standard · agent steps, one 96 GB card (hosted demo) (1× RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B carries out the plain-language steps through the flight recorder's /decide, a receipt per action; assertions judge.
- Alternates (smaller models for the same job):
  - Vision actions for canvas and iframe apps (1x 96 GB card (estimate)), not served: Try the standard decider's own vision tower first (Qwen3.8-27B, same weights; screenshots took our MiniWoB++ run from 25.9% to 55.8%), then a dedicated grounder such as Holo-3.1 when pixel grounding is needed. Needs image-input receipts on the gateway.
- Buyer facts:
  - What decides pass or fail: Only the assertions in your spec, judged in code on the live page. The model picks actions and never sees them.
  - Typical run cost: A handful of model actions, a fraction of a cent at the gateway list price (measured over many runs).
  - Data retention: Hosted: certificates and screenshots-as-thumbnails for 24 hours, then deleted. Self-host: as long as you set.
  - What leaves the box (hosted): Each step's element table and page text go to the gateway model; screenshots stay on the runner as hashes and 320 px thumbnails in the certificate. Secrets are replaced by {{NAME}} before anything is recorded or sent.
  - Where it may run: The fixture app, saucedemo.com, and https hosts whose domain your API key verified (DNS TXT or /.well-known file, valid 30 days). Self-host: any host you list.
  - Spec format: YAML or JSON: do or goto steps; assertions text_present, text_absent, url_contains, url_equals, title_contains, element_count, value_equals, value_contains.
  - CI: decosa_testrun.py (standard library): exit 0 pass, 1 fail, 2 error; writes and verifies the certificate; --repeat for a signed flake report.
- Samples:
  - 1. Kiln cart total@good (`kiln-cart-total@good`): https://decosa.ai/tools/developer/test-runs?sample=1&autorun=0
  - 2. Kiln cart total@total ignores qty (`kiln-cart-total@total-ignores-qty`): https://decosa.ai/tools/developer/test-runs?sample=2&autorun=0
  - 3. Kiln cart total@redesign (`kiln-cart-total@redesign`): https://decosa.ai/tools/developer/test-runs?sample=3&autorun=0
  - 4. Sauce checkout@standard user (`sauce-checkout@standard_user`): https://decosa.ai/tools/developer/test-runs?sample=4&autorun=0
  - 5. Kiln signup@signup broken (`kiln-signup@signup-broken`): https://decosa.ai/tools/developer/test-runs?sample=5&autorun=0
  - 6. Kiln promo@promo wrong (`kiln-promo@promo-wrong`): https://decosa.ai/tools/developer/test-runs?sample=6&autorun=0
  - 7. Kiln remove@remove wrong (`kiln-remove@remove-wrong`): https://decosa.ai/tools/developer/test-runs?sample=7&autorun=0
  - 8. Kiln checkout@checkout disabled (`kiln-checkout@checkout-disabled`): https://decosa.ai/tools/developer/test-runs?sample=8&autorun=0
- Regulatory note: Checked 2026-09-25. SOC 2 (AICPA Trust Services Criteria, CC8.1) and ISO/IEC 27001:2022 (Annex A 8.29 security testing in development and acceptance, 8.32 change management) expect evidence that changes were tested and approved before release; neither requires a particular tool, a signature or a hash chain. A run certificate is one piece of evidence an auditor can sample and re-check; it does not make a change process compliant, does not record who approved the release, and covers only what the spec asserts. Testing a website without the owner's permission can be unlawful (for example under the US Computer Fraud and Abuse Act or the UK Computer Misuse Act); the hosted service therefore only visits its own fixture app, Sauce Labs' public test site and domains whose owner proved control with a DNS TXT record or a /.well-known file. Use test accounts and test data: screenshots of a staging site can contain personal data, and under the GDPR minimisation and storage limits apply (hosted certificates are kept 24 hours; self-host keeps everything on your box). Passwords go in run secrets and never reach the model or the certificate. Model licence: Apache-2.0 (Qwen3.8-27B). Not legal advice.

### Build a model-risk evidence pack (`model-risk-pack`)

Signed evidence that an LLM a bank, insurer or lender relies on still behaves as validated: identity probes, graded test cases, stability against the validated run, counterfactual fairness pairs and a model card, with a re-check before any alert. Validators can recompute every number from the record.

- What it is: Signed evidence that an LLM your bank, insurer or lending team relies on still behaves as validated, with a record your validators can recompute.
- Who it's for: Teams in finance and insurance and compliance and trust.
- Page: https://decosa.ai/tools/finance/model-risk-pack
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/model-risk-pack):
  - 0 of 8 / 0 of 8: False alarms on unchanged runs (ALERT / WATCH) (test split, n = 8)
  - 8 of 12 runs: Injected changes detected (ALERT) (test split, n = 12)
  - 0 of 4 runs: Sampling drift detected (temperature 0.3 / 0.7) (test split, n = 4)
  - 62.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B (hosted system and fixed grader); any OpenAI-compatible endpoint under test. Where: Hosted or self-host.
- Verification: Receipt per call; signed pack and hash-chained record that recompute (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the grader; the pack runner is CPU; the system under test runs wherever it runs
- Output: A signed evidence pack with a Markdown binder, and a hash-chained record a validator can recompute.
- Page: https://decosa.ai/tools/finance/model-risk-pack · JSON: https://decosa.ai/use-cases/model-risk-pack.json
- Prompts: https://decosa.ai/prompts/model-risk-pack-hosted.md · https://decosa.ai/prompts/model-risk-pack-selfhost.md · https://decosa.ai/prompts/model-risk-pack-assemble.md · https://decosa.ai/prompts/model-risk-pack-mac.md
- Tiers:
  - Lite · grader on one 32 GB card (self-host) (1x RTX 5090 32 GB (estimate)): The same pinned grader at 4-bit on an RTX 5090; the runner on the same box. Test your own endpoints on your network.
  - Standard · hosted grader and demo systems (Qwen3.8-27B) (1x RTX PRO 6000 96 GB (measured, shared)): What the hosted API runs: every probe of the hosted system and every grader call is receipted by the gateway.
- Buyer facts:
  - Typical run cost: A few cents or less per pack at the gateway list price; more with a re-check. Each run shows its own measured cost.
  - Data retention: Hosted: packs for the demo systems (synthetic) are stored; packs for your endpoint or suite are never stored, they go back to you. Self-host: nothing leaves the box.
  - What leaves the box: Hosted: your suite's case texts go to the system under test and, for drafts, to the hosted grader. Self-host: only calls to the endpoint you test.
  - Inputs: A suite of up to 40 classify cases, 10 fairness bases with up to 8 variants, 10 drafts (200 KB); any OpenAI-compatible endpoint, https and public on the hosted API
  - Evidence: Signed pack (Ed25519) + hash-chained record with every output and receipt; /mrm/verify recomputes every number; Markdown binder
  - Not: A validation opinion, legal advice, or a disparate-impact study on real applicants
- Samples:
  - 1. Mrm vendor (`mrm-vendor`): https://decosa.ai/tools/finance/model-risk-pack?sample=1&autorun=0
  - 2. Mrm vendor updated (`mrm-vendor-updated`): https://decosa.ai/tools/finance/model-risk-pack?sample=2&autorun=0
  - 3. Mrm vendor biased (`mrm-vendor-biased`): https://decosa.ai/tools/finance/model-risk-pack?sample=3&autorun=0
  - 4. Mrm vendor drift (`mrm-vendor-drift`): https://decosa.ai/tools/finance/model-risk-pack?sample=4&autorun=0
  - 5. Mrm vendor swapped (`mrm-vendor-swapped`): https://decosa.ai/tools/finance/model-risk-pack?sample=5&autorun=0
  - 6. Mrm lender (`mrm-lender`): https://decosa.ai/tools/finance/model-risk-pack?sample=6&autorun=0
- Regulatory note: Checked 25 Sep 2026; not legal advice. US banks: SR 26-2 (Fed, OCC, FDIC, 17 Apr 2026) replaced SR 11-7 and OCC 2011-12, and its footnote 3 puts generative and agentic AI models outside its scope, leaving their controls to each bank's own risk management and governance. This pack is evidence for that governance; it is not a validation opinion and does not make anyone compliant. Lenders: ECOA / Regulation B (12 CFR 1002.9) requires specific principal reasons for adverse action; the CFPB withdrew Circulars 2022-03 and 2023-03 on 12 May 2025, but the rule stands. Insurers: the NAIC AI model bulletin is adopted in 25 jurisdictions (NAIC map, 1 Apr 2026), insurers stay responsible for vendor AI systems, and a 12-state AI Systems Evaluation Tool pilot ran March to September 2026; NYDFS Circular Letter 7 (2024) asks for testing for unfair discrimination before and after deployment. EU: credit scoring of natural persons is high-risk under the AI Act (Annex III 5(b)); after Regulation (EU) 2026/1744 those obligations apply from 2 Dec 2027. The fairness probes change one attribute at a time among the values in the suite; they do not measure disparate impact on real applicants. Test cases often hold customer data: keep real cases on a self-hosted box. Model licences: Apache-2.0 (Qwen3.8-27B, Qwen3-1.7B).

### Check an AI note (`clinical-ai-monitor`)

Checks any commercial AI scribe's note against the visit transcript: every sentence typed as not in the visit, a changed detail, the wrong speaker or an invented exam finding, with the transcript line, plus the medicines, allergies and plan items the note left out. Tightens the note without adding a word. Buyer mode scores vendors into a signed scorecard.

- What it is: Checks any AI scribe's note against the visit transcript, sentence by sentence: invented facts, changed doses and details, exam findings nobody made, the wrong speaker and what was left out, each with the transcript line. Tightens the note without adding anything, and scores vendors in bulk.
- Who it's for: Teams in healthcare and compliance and trust.
- Page: https://decosa.ai/clinics/clinical-ai-monitor
- Last reviewed: 2026-09-28
- Key numbers (details and caveats: https://decosa.ai/metrics/clinical-ai-monitor):
  - 49 (98%): Invented fact caught as an error (test split, n = 50)
  - 47 (94%): Changed detail caught as an error (test split, n = 50)
  - 45 (90%): Key item left out caught (test split, n = 50)
  - 28.0 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Qwen3.8-27B · MOSS diarizer (audio). Where: Self-host (patient data stays on site).
- Verification: Receipt per verdict; signed reports and summary (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the judge; the diarizer (optional, for audio) needs a second GPU slot
- Output: A signed per-encounter report and a signed summary of rates per vendor, specialty and week.
- Page: https://decosa.ai/clinics/clinical-ai-monitor · JSON: https://decosa.ai/use-cases/clinical-ai-monitor.json
- Prompts: https://decosa.ai/prompts/clinical-ai-monitor-hosted.md · https://decosa.ai/prompts/clinical-ai-monitor-selfhost.md · https://decosa.ai/prompts/clinical-ai-monitor-assemble.md · https://decosa.ai/prompts/clinical-ai-monitor-mac.md
- Tiers:
  - Lite · text transcripts, one 32 GB card (1x RTX 5090 32 GB (estimate)): The judge only: send transcripts as text (most scribe products export one). No audio input.
  - Standard · judge plus diarizer (hosted demo) (1x RTX PRO 6000 96 GB (measured), diarizer on a second slot): Adds audio input through the MOSS diarizer. This is what the hosted API runs.
  - Wanted · a second judge from another family (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash beside the standard card, or a Mac with 192 GB or more (MXFP4 MLX build, 156 GB on our Mac; not measured). Estimate.): DeepSeek-V4-Flash re-judges the transcripts to check the first judge's rates, where wrong-speaker and changed-detail errors are hardest. Patient data stays on your hardware, never on community providers. Not served yet.
- Alternates (smaller models for the same job):
  - A second judge on one card (1x RTX PRO 6000 96 GB (80 GB of weights)), not served: A cheaper cross-family check that fits beside the 27B: smaller than DeepSeek-V4-Flash, so we would host it ourselves. sm_120 support unconfirmed.
- Buyer facts:
  - What leaves the box: Nothing when self-hosted (DECOSA_LLM_ROUTE=direct). The hosted demo takes public mock visits only.
  - Retention: None: transcripts, notes and audio stay in memory for one request; logs carry counts.
  - What you keep: A signed report per visit (hashes, line numbers, counts; no text, the ISO week not the date) and signed summaries.
  - Inputs: Transcript as text with speaker labels, WebVTT or SRT captions, a scribe export with timestamps, or segments (up to 40,000 characters), or WAV audio up to 15 minutes through the API; the note as text (up to 12,000 characters). Buyer mode: up to 25 visits per call hosted, no limit self-hosted.
  - Typical cost: A few cents or less per note at the gateway list price (measured on the demo notes), rising with note and transcript length, including tighten; the detail checker runs on CPU at no token cost. Self-hosted: your own hardware.
  - Hardware: One GPU with 32 GB or more for the judge; the diarizer (audio only) needs about 4 GB more; the detail checker runs on CPU (8 threads, about 2.5 GB of RAM).
  - Output: Per sentence: the flag type, the note sentence and the transcript lines (or 'not in the transcript'); key items left out; a tightened note; a signed per-visit report; in buyer mode, a signed scorecard per vendor and version.
- Samples:
  - 1. Video errors (`video-errors`): https://decosa.ai/clinics/clinical-ai-monitor?sample=1&autorun=0
  - 2. Video clean (`video-clean`): https://decosa.ai/clinics/clinical-ai-monitor?sample=2&autorun=0
  - 3. Gp changed (`gp-changed`): https://decosa.ai/clinics/clinical-ai-monitor?sample=3&autorun=0
  - 4. Gp frequency (`gp-frequency`): https://decosa.ai/clinics/clinical-ai-monitor?sample=4&autorun=0
  - 5. Scorecard (`scorecard`): https://decosa.ai/clinics/clinical-ai-monitor?sample=5&autorun=0
- Regulatory note: A documentation-quality measure for the health system's own oversight, not a medical device function and not software that recommends care: it does not analyse images or signals, does not diagnose, and makes no recommendation about any patient's care; it compares a note with the conversation it came from. The closest statutory exclusion is software for administrative support of a health care facility (FD&C Act 520(o)(1)(A)); FDA's clinical decision support guidance (reissued 29 Jan 2026) covers software that recommends care, which this does not. Keep it that way: do not wire its flags into real-time changes to a patient's care. HIPAA: run it on your own hardware; measuring your own notes is a quality assessment activity, part of health care operations (45 CFR 164.501). A vendor that hosts it or can reach the PHI needs a business associate agreement. The hosted demo takes only PriMock57 mock consultations and synthetic notes. Governance: the Joint Commission and CHAI guidance on the Responsible Use of AI in Healthcare (17 Sep 2025) names clinical documentation in scope and asks for local validation, ongoing quality monitoring and error reporting to leadership and vendors; the Joint Commission's voluntary certification on it was announced 1 Jun 2026. ONC HTI-1's decision support interventions criterion (45 CFR 170.315(b)(11)) binds certified EHR developers, not hospitals, and HTI-5, proposed 29 Dec 2025 and not final when checked, would remove its source-attribute and risk-management duties; neither changes what a hospital may measure itself. State laws bind the providers that use scribes, not this monitor: California AB 3030 (disclaimers on AI-generated patient communications, from 1 Jan 2025), Texas SB 1188 (practitioners review AI-created records, from 1 Sep 2025) and TRAIGA (disclose AI use in care, from 1 Jan 2026), and Illinois Public Act 104-0054 (consent before recording therapy sessions for AI, 1 Aug 2025). Its reports can document the review these laws expect. Some scribe contracts restrict benchmarking or publishing results; check yours. Not legal advice. Model licences: Apache-2.0 (Qwen3.8-27B, MOSS-Transcribe-Diarize). Checked 25 Sep 2026.

### Build a privilege log (`privilege-log`)

Typed privilege calls for each email or memo in a production set: attorney-client, work product, not privileged, or needs attorney review, with the lawyer involved, the reason checked against the document and a calibrated probability. It drafts log entries that describe without disclosing, catches the ones that leak, gives duplicates the same call, flags outsiders who could waive, and seals it all in a signed record.

- What it is: A privilege log drafted on your own GPU: a typed privilege call for every email or memo, with the lawyer involved, a grounded reason and a calibrated probability; Rule 26(b)(5)(A) log entries that describe without disclosing; the same call for every copy; and one signed record. It drafts; a lawyer decides.
- Who it's for: Teams in legal.
- Page: https://decosa.ai/legal/privilege-log
- Last reviewed: 2026-09-25
- Key result: On 149 held-out emails from Enron lawyers' mailboxes the privileged-or-not call was 83.9% accurate (AUROC 0.908), with 38% sent to attorney review. The labels are one AI reviewer's, not a lawyer's, and on hard calls the model alone is barely better than chance, so the review queue is the safeguard.
- Key numbers (details and caveats: https://decosa.ai/metrics/privilege-log):
  - 83.9% / 82.3% / 79.7%: Privileged vs not, model call: accuracy / precision / recall (test split, n = 149)
  - 0.908 / 0.058: AUROC of p(privileged) / ECE (test split, n = 149)
  - 4 (6.2% of privileged): Privileged emails the tool would have produced (waiver risk) (test split, n = 64)
  - 121.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real matters; hosted for the demo sets only.
- Verification: Receipt per call; leak check; signed record and reviewer ledger (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; the people map, grouping and leak rules run on CPU
- Output: A draft privilege log (CSV), a review memo for the supervising lawyer, and a signed record and ledger of every call and decision.
- Page: https://decosa.ai/legal/privilege-log · JSON: https://decosa.ai/use-cases/privilege-log.json
- Prompts: https://decosa.ai/prompts/privilege-log-hosted.md · https://decosa.ai/prompts/privilege-log-selfhost.md · https://decosa.ai/prompts/privilege-log-assemble.md · https://decosa.ai/prompts/privilege-log-mac.md
- Tiers:
  - Lite · one 32 GB card, self-hosted (1x RTX 5090 32 GB): The same model and prompts on a single RTX 5090, one call per question with log-probabilities. The grounding check of the reason can be switched off (DECOSA_PRIVILEGE_GROUNDING=0) to save one call per document.
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB): Qwen3.8-27B on an RTX PRO 6000. Self-hosted it uses log-probabilities (about 5 calls per document); the hosted demo, whose gateway does not pass log-probabilities through yet, samples the privilege call 5 times instead.
  - Wanted · a larger second judge on your own hardware (Your own hardware: 2x 96 GB cards (NVFP4, about 170-186 GB, unconfirmed) or a Mac with 192 GB or more (MLX 4-bit, 165 GB). Estimate; GLM's SGLang SM120 build hung on our server.): GLM-5.3-Flash re-reads the calls Qwen3.8-27B is least sure of, from a different model family. Privileged documents stay on your hardware, never on community providers. Not served yet.
- Alternates (smaller models for the same job):
  - Fast option-token judge for volume (1x RTX PRO 6000 96 GB (BF16 weights are 49 GB)), not served: A 4B-active judge read through option-token probabilities: cheaper per call, not better. Page 32 suggests the dense Qwen3.5-9B (Apache-2.0) instead. The real blocker is logprob passthrough on the gateway.
- Buyer facts:
  - Data retention: Documents are held in memory for the request; the run (calls, log, no document text) for one hour, for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: nothing. Hosted demo: the documents go to the model through our gateway, which is why the hosted demo is for the sample sets only.
  - Model cost per 1,000 documents: A dollar or two per thousand documents on the direct route at the gateway's list price (measured); GPU time only when self-hosted.
  - Input: Emails or memos as JSON: id, date, from, to, cc, bcc, subject, body, thread_id, family_id, parent_id, bates. Up to 25 per request hosted, 12,000 characters each; set DECOSA_PRIVILEGE_MAX_DOCS on your own box.
  - Output: Per document: withhold / produce / needs review, p(privileged), the lawyers, a grounded reason and flags. A Rule 26(b)(5)(A) log as CSV, a review memo, a signed record and a signed ledger of each lawyer's decision.
- Samples:
  - 1. Harborline dispute (`harborline-dispute`): https://decosa.ai/legal/privilege-log?sample=1&autorun=0
  - 2. Enron legal (`enron-legal`): https://decosa.ai/legal/privilege-log?sample=2&autorun=0
- Regulatory note: A review aid for lawyers, not legal advice and not a privilege determination. Fed. R. Civ. P. 26(b)(5)(A) requires a party withholding documents as privileged or work product to expressly claim it and describe each document in a way that lets others assess the claim without revealing the protected content; the attorney who signs discovery responses certifies them under Rule 26(g), so every call and log entry here is a draft for that attorney. A wrong 'not privileged' can waive privilege; Fed. R. Evid. 502(b) protects inadvertent disclosure only where reasonable steps were taken, and a Rule 502(d) order (non-waiver regardless of care) is the safer footing for any AI-assisted review. Courts have judged generative-AI review by the same reasonableness and proportionality standard as earlier technology-assisted review (Schulte v. LinkedIn, N.D. Cal., 30 Jun 2026); the signed record and ledger document which model, prompt and lawyer produced each call. Confidentiality: ABA Formal Opinion 512 (29 Jul 2024) asks lawyers to understand how an AI tool uses what they put in, to protect client information, and to get informed consent before putting it into a self-learning tool; United States v. Heppner (S.D.N.Y., Feb 2026) held a defendant's chats with a consumer AI service not privileged, partly because its terms allowed disclosure to third parties. Run real matters on your own machine; the hosted demo is for the fictional and public sample sets only. Model licence: Apache-2.0 (Qwen3.8-27B). The Enron emails are public records released by FERC (2003), used here as a small research sample. Checked 25 Sep 2026.

### Log human edits and editor sign-off (`editorial-ledger`)

Keeps the AI first draft, every human edit as a diff with edit metrics, and a named editor's sign-off in one signed record, with a public page per published piece. It is the evidence behind the EU AI Act's exception for AI-assisted text that had human review and editorial responsibility (Art. 50(4)).

- What it is: A record for the EU AI Act Article 50 human-review exception: the AI draft, every human edit and a named editor's sign-off on one exact version, sealed in one signed ledger, with a public check for each published piece. It is evidence that a person reviewed the text, not a safe harbour.
- Who it's for: Teams in creative and media and compliance and trust.
- Page: https://decosa.ai/tools/media/editorial-ledger
- Last reviewed: 2026-09-25
- Key result: On the IteraTeR test split the claim check flagged 31 of 35 human edits that changed meaning (precision 0.49 against a proxy label), and 120 of 120 tampered ledgers failed verification (synthetic). No real newsroom editors on real copy were measured.
- Key numbers (details and caveats: https://decosa.ai/metrics/editorial-ledger):
  - 24 / 24: Wording-only paraphrases with no claim change flagged (held out) (held out, n = 24)
  - 31 of 35 (0.89): IteraTeR test, meaning-changed edits flagged (recall) (test split, n = 35)
  - 80 / 80: Edit metrics exact against scripted diffs (synthetic, n = 80)

- Status: live. Models: Qwen3.8-27B (draft and claim check). Where: Hosted for published pieces; self-host for embargoed copy.
- Verification: Receipted draft; signed, hash-chained ledger; optional C2PA credential (receipt coverage: full). Hosted QA 2026-09-25: pass (pre-release server). Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; metrics, ledger and verification run on CPU
- Output: A signed editorial ledger with a public verify page, and a DOCX of the published text with a C2PA credential.
- Page: https://decosa.ai/tools/media/editorial-ledger · JSON: https://decosa.ai/use-cases/editorial-ledger.json
- Prompts: https://decosa.ai/prompts/editorial-ledger-hosted.md · https://decosa.ai/prompts/editorial-ledger-selfhost.md · https://decosa.ai/prompts/editorial-ledger-assemble.md · https://decosa.ai/prompts/editorial-ledger-mac.md
- Tiers:
  - Lite · ledger only, on CPU (self-host) (Any CPU): Your CMS's own AI tool writes the draft; the ledger keeps it (marked as not receipted), diffs every save, computes the metrics and seals the sign-off. No model here, so no claim check.
  - Standard · Qwen3.8-27B drafts and checks claims (hosted demo) (1× RTX PRO 6000 96 GB (measured) or 1× RTX 5090 32 GB (estimate)): The model writes the first draft from the sources with a signed receipt, and lists the claims each edit added or removed. This is what the hosted API runs.
- Buyer facts:
  - What you send: Sources (up to 6, 20,000 characters each) or a draft your own tool wrote, then each saved version as text or HTML, the editor's name and role, and the sign-off.
  - What you get: A signed ledger per piece, a public page at /ledger/<id> that checks it in the reader's browser, a copyable article badge, and a DOCX of the published text with a C2PA credential.
  - Typical cost: Two model calls per piece: a fraction of a cent at the gateway list price. The ledger work is CPU.
  - Retention (hosted): Open pieces 7 days, published ledgers 30 days, with the texts. Choose whether the public page shows the AI draft; its hash is always there. Self-host keeps everything on your box.
  - What leaves the box (self-host): Nothing: the model, the ledger and the public page can all run on your machine. On the hosted route the sources and the texts reach our server and the gateway.
  - Identity: The ledger records the name your key-holder sends. For stronger evidence an editor can sign the sign-off with their own Ed25519 key; there is no SSO binding yet.
- Samples:
  - 1. Ferry terminal (`ferry-terminal`): https://decosa.ai/tools/media/editorial-ledger?sample=1&autorun=0
  - 2. Library hours (`library-hours`): https://decosa.ai/tools/media/editorial-ledger?sample=2&autorun=0
- Regulatory note: Checked 25 Sep 2026. EU AI Act (Regulation (EU) 2024/1689) Art. 50(4), second subparagraph: deployers of an AI system that generates or manipulates text published to inform the public on matters of public interest must disclose that it was artificially generated or manipulated, unless the content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for the publication. It applies from 2 August 2026 (Art. 113); transparency breaches can be fined up to EUR 15 million or 3% of worldwide turnover (Art. 99(4)). The Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) gave generative systems already on the market until 2 December 2026 for the provider marking duty; published summaries report no change to this deployer duty. The Commission's voluntary Code of Practice on marking and labelling AI-generated content (about 190 signatories by July 2026) and its Art. 50 guidelines restate the exception. This ledger is evidence that a review happened and who holds responsibility; whether a given review meets the exception is for the publisher, and ultimately regulators and courts, to judge, and the metrics cannot tell a careful read from a careless one. US: the Copyright Office's registration guidance (16 March 2023, 88 FR 16190) and its Part 2 report on copyrightability (29 January 2025) protect only human authorship and ask applicants to disclose AI-generated material and describe the human contribution; the ledger documents what a person changed, it does not make text protectable. Model licence: Apache-2.0 (Qwen3.8-27B). Not legal advice.

### Check a police report against bodycam (`report-integrity`)

Checks an incident report (police, security, EMS, workplace) against the transcript of its recording, sentence by sentence, and lists key events the report leaves out, with times you can cue up. Or drafts the report from the recording, keeps a signed first draft, and records which sentences a person changed before signing.

- What it is: An AI police report check: every sentence of an incident report, written by an officer or drafted by AI, is compared with the transcript of the body cam or other recording it describes, with the times, and the key events the report leaves out are listed. It runs on your own GPU.
- Who it's for: Teams in public sector and legal.
- Page: https://decosa.ai/legal/report-integrity
- Last reviewed: 2026-09-25
- Key result: On 8 held-out synthetic incidents (test split, one run) it found 16 of 16 planted additions, 16 of 16 contradictions and 16 of 16 omissions, and flagged 0 of 64 faithful sentences. The building agent wrote the scenarios, so they are clear-cut; it has not been measured on real body-worn-camera audio.
- Key numbers (details and caveats: https://decosa.ai/metrics/report-integrity):
  - 16/16: Planted additions found (test split, n = 16)
  - 16/16 (16/16): Planted contradictions found (exact) (test split, n = 16)
  - 16/16 (13/16): Planted omissions found, missing or partly (missing only) (test split, n = 16)
  - 97.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: MOSS-Transcribe-Diarize · Qwen3.8-27B. Where: Self-host (hosted demo: synthetic material only).
- Verification: Receipt per model call; signed check; signed first draft and hash-chained edit record (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Qwen3.8-27B and the diarizer; the check itself needs no GPU
- Output: Every sentence marked against the recording with its times, the key events the report leaves out, and a signed statement; in draft mode a signed first draft and an edit record.
- Page: https://decosa.ai/legal/report-integrity · JSON: https://decosa.ai/use-cases/report-integrity.json
- Prompts: https://decosa.ai/prompts/report-integrity-hosted.md · https://decosa.ai/prompts/report-integrity-selfhost.md · https://decosa.ai/prompts/report-integrity-assemble.md · https://decosa.ai/prompts/report-integrity-mac.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B lists key events, drafts, judges every sentence and every event; MOSS-Transcribe-Diarize turns recordings into transcripts. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger judge for long, many-speaker incidents.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Reports and transcripts stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: A written report against the transcript of its recording: typed, WebVTT or SRT, or made from the audio by the diarizer on your box.
  - Data kept: Nothing. Transcripts, reports and records live in memory for the request; logs carry counts and timings only. You keep the signed statement and record.
  - What leaves the box (hosted demo): The transcript and report go to Qwen3.8-27B through our gateway; the gateway's receipts hold hashes, not text. Self-hosted: nothing leaves.
  - Model calls per check: One per report sentence, one for the key-event list, one per key event: about 9 for a 10-sentence report (284 calls over 32 eval runs).
  - Typical run cost: A fraction of a cent per check at the open-market median price for Qwen3.8-27B. Each run shows its own measured cost.
  - Disclosure wording: California SB 524 and Utah SB 180 text built in, plus a neutral option; the signed record holds the disclosure and the author's certification.
- Samples:
  - 1. Traffic stop audio (`traffic-stop-audio`): https://decosa.ai/legal/report-integrity?sample=1&autorun=0
  - 2. Bar fight audio (`bar-fight-audio`): https://decosa.ai/legal/report-integrity?sample=2&autorun=0
  - 3. Retail detention (`retail-detention`): https://decosa.ai/legal/report-integrity?sample=3&autorun=0
  - 4. Ems fall (`ems-fall`): https://decosa.ai/legal/report-integrity?sample=4&autorun=0
  - 5. Forklift (`forklift`): https://decosa.ai/legal/report-integrity?sample=5&autorun=0
  - 6. Traffic stop faithful (`traffic-stop-faithful`): https://decosa.ai/legal/report-integrity?sample=6&autorun=0
- Regulatory note: Checked 25 Sep 2026 against the enacted texts. California SB 524 (Stats. 2025, ch. 587; Penal Code 13663, in force 1 Jan 2026): an official report written fully or partly with AI must say so on each page or in its body and name the AI program, carry the officer's signature verifying the facts are true and correct, keep the first draft as long as the official report, and keep an audit trail of who used AI and the footage used; drafts other than the official report are not the officer's statement. Utah SB 180 (2025; Utah Code 53-25-601 and 53-25-602, in force 7 May 2025): a report created wholly or partly with generative AI must carry a disclaimer and the author's certification that they read and reviewed it for accuracy, and each agency needs a written AI policy. The disclosure and certification text here follows those statutes but is not legal advice. 'Not in the recording' means the transcript does not contain it, not that it is false: a camera can miss what a person saw. Criminal justice information held by agencies falls under the FBI CJIS Security Policy; this stack has no CJIS attestation, which is one more reason it is self-host first.

### Redline from firm precedents (`legal-drafting-editor`)

Applies a firm's playbook to a contract and returns a Word file with real tracked changes and a margin comment on each, naming the rule, the position and the precedent. New wording comes only from the firm's own precedent clauses: code traces every inserted word back to one, a grounding check and a second read of the rule confirm each edit, and reject-all gives back the original word for word.

- What it is: A firm's playbook applied to a contract: a Word redline with real tracked changes, a margin comment on each, and new wording only from the firm's own precedents, traced in code.
- Who it's for: Teams in legal.
- Page: https://decosa.ai/legal/legal-drafting-editor
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/legal-drafting-editor):
  - 0.77 / 0.79: Clause detection vs CUAD labels, v1 mapping (like for like): precision / recall (held out, n = 30)
  - 0.82 / 0.85: Clause detection vs CUAD labels, v2 mapping: precision / recall (held out, n = 30)
  - 0.98 / 1.00: Planted deviations flagged: precision / recall (held out, n = 102)
  - 28.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for client contracts; hosted for public and fictional samples.
- Verification: Receipt per call; every insertion traced to a precedent; validated redline; signed record and decision ledger (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; the DOCX writer, search and checks run on CPU
- Output: A .docx with tracked changes and margin comments to accept or reject in Word, a review memo, a signed record, and a signed ledger of the lawyer's decisions read back from the returned file.
- Page: https://decosa.ai/legal/legal-drafting-editor · JSON: https://decosa.ai/use-cases/legal-drafting-editor.json
- Prompts: https://decosa.ai/prompts/legal-drafting-editor-hosted.md · https://decosa.ai/prompts/legal-drafting-editor-selfhost.md · https://decosa.ai/prompts/legal-drafting-editor-assemble.md · https://decosa.ai/prompts/legal-drafting-editor-mac.md
- Tiers:
  - Lite · one 32 GB card, self-hosted (1x RTX 5090 32 GB): The same model and prompts on a single RTX 5090. The grounding check of each edit can be switched off (DECOSA_DRAFTING_GROUNDING=0) to save a call or two per change; the code trace still guarantees no invented words.
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB): Qwen3.8-27B on an RTX PRO 6000: one call per playbook rule (with the closest CUAD train-split examples of the rule's clause), one per targeted edit, plus the grounding check and the re-read of each edit.
  - Wanted · a larger second reader on your own hardware (Your own hardware: 2x 96 GB cards (NVFP4, about 170-186 GB, unconfirmed) or a Mac with 192 GB or more (MLX 4-bit, 165 GB). Estimate; GLM's SGLang SM120 build hung on our server.): GLM-5.3-Flash reads each clause beside Qwen3.8-27B, and a rule fires only when both agree or the reviewer decides. Privileged drafts stay on your hardware, never on community providers. Not served yet.
- Alternates (smaller models for the same job):
  - Fast reader for long contracts (1x RTX PRO 6000 96 GB (BF16 weights are 49 GB)), not served: Gemma-4-26B-A4B for the clause-by-clause reading at volume, keeping Qwen3.8-27B for edits (page 28's plan; not run). Page 32 suggests the dense Qwen3.5-9B (Apache-2.0) instead.
- Buyer facts:
  - Data retention: The contract is held in memory for the request; the run (findings, changes, the redline) for one hour, for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: nothing. Hosted demo: the contract's candidate paragraphs go to the model through our gateway, which is why the hosted demo is for public and fictional contracts only.
  - Where new wording comes from: Only the firm's precedent clauses, with the parties' defined terms filled in. Every inserted chunk is traced to its precedent in code; if a check fails the change falls back to the precedent's own sentence or to a comment.
  - Input: A .docx up to 2 MB or text up to 200,000 characters; a playbook of up to 20 rules and a library of up to 80 precedent clauses as JSON (the default is a synthetic one).
  - Output: A .docx with Word tracked changes and a comment per change; a review memo; a signed record; and, from the file returned from Word, a signed ledger of which changes the lawyer accepted or rejected.
  - Model cost per contract: About a cent or less per contract at the gateway's list price (median, measured on CUAD runs); GPU time only when self-hosted.
- Samples:
  - 1. Northwind msa (`northwind-msa`): https://decosa.ai/legal/legal-drafting-editor?sample=1&autorun=0
  - 2. Cuad dragon outsourcing (`cuad-dragon-outsourcing`): https://decosa.ai/legal/legal-drafting-editor?sample=2&autorun=0
  - 3. Cuad garman software (`cuad-garman-software`): https://decosa.ai/legal/legal-drafting-editor?sample=3&autorun=0
- Regulatory note: A drafting aid for lawyers, not legal advice; the lawyer reviews and adopts every change. Under ABA Model Rule 1.1 (comment 8) a lawyer using it must understand what it does and check its output, and Rules 5.1 and 5.3 make supervision of the tool the lawyer's duty: the signed record and the ledger of accepted and rejected changes are evidence of that review. Rule 1.6(c) requires reasonable efforts to prevent unauthorised disclosure of client information, and ABA Formal Opinion 512 (29 Jul 2024) asks lawyers to understand how a generative-AI tool uses what they put in and to get informed consent before putting client information into a self-learning tool (boilerplate in an engagement letter is not enough); United States v. Heppner (S.D.N.Y., Feb 2026) held chats with a consumer AI service not privileged, partly because its terms allowed disclosure. Self-hosted, the contract never leaves the firm's machine and nothing is trained on it; the hosted demo is for public (CUAD) and fictional contracts only. Bar guidance varies by state (California's 2026 practical guidance, for example, requires AI costs to be billed at cost without markup unless agreed in writing); check the rules where you practise. The playbook and precedents in the demo are synthetic. CUAD v1 (The Atticus Project) is CC BY 4.0; its contracts are public SEC EDGAR filings. Model licence: Apache-2.0 (Qwen3.8-27B). Checked 25 Sep 2026.

### Redact a public-records release (`foia-desk`)

For a records officer answering a FOIA or state public-records request: turns the request into a search scope, finds the responsive records, and proposes each redaction with its exemption and a reason that names the harm without repeating the withheld words. The release PDF is built without the withheld text and read back to prove nothing under a box can be copied out. It drafts the index and the response letter, and seals every call and every officer decision in a signed record.

- What it is: FOIA software for records officers that runs on the agency's own machine: it scopes a FOIA or public-records request, finds the responsive records, proposes each redaction with its exemption and a reason, and releases a PDF with the withheld text really removed, checked, indexed and signed. It proposes; the officer decides.
- Who it's for: Teams in public sector and legal.
- Page: https://decosa.ai/legal/foia-desk
- Last reviewed: 2026-09-25
- Key result: On the current pipeline it redacted 59 of 60 planted personal details (98.3%) at 83.1% redaction precision, and 0 of 95 withheld spans were recoverable from the release PDF. The test sets were run five times with fixes in between and labelled by an AI reviewer, so these are not clean held-out numbers.
- Key numbers (details and caveats: https://decosa.ai/metrics/foia-desk):
  - 5/8: PII recall on a set new to the pipeline (bus-contract, run 3) (held out, n = 8)
  - 59 of 60 (98.3%): Personal-data spans redacted (PII recall), final pipeline (test split, n = 60)
  - 70 of 71 (98.6%): Planted spans redacted, final pipeline (test split, n = 71)
  - 40.0 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real requests; hosted for the demo sets only.
- Verification: Receipt per call; release check; signed record and decision ledger (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; pattern finders, the release PDF and its check run on CPU
- Output: A release PDF with a passed integrity check, an index of withheld information (CSV), a response letter, and a signed record and decision ledger.
- Page: https://decosa.ai/legal/foia-desk · JSON: https://decosa.ai/use-cases/foia-desk.json
- Prompts: https://decosa.ai/prompts/foia-desk-hosted.md · https://decosa.ai/prompts/foia-desk-selfhost.md · https://decosa.ai/prompts/foia-desk-assemble.md · https://decosa.ai/prompts/foia-desk-mac.md
- Tiers:
  - Lite · one 32 GB card, self-hosted (1x RTX 5090 32 GB): The same model and prompts on a single RTX 5090, one call per question with log-probabilities. The privilege engine's reason check can be switched off (DECOSA_FOIA_GROUNDING=0) to save a call per record with counsel on it.
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB): Qwen3.8-27B on an RTX PRO 6000, through our gateway with one call per question on the hosted demo (log-probabilities when self-hosted).
- Buyer facts:
  - Data retention: Records are held in memory for the request and the run (one hour, so the release can be rebuilt after the officer's decisions), for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: nothing. Hosted demo: the records go to the model through our gateway, which is why the hosted demo is for the sample sets only.
  - Model cost per 1,000 records: Measured on the held-out test run at the gateway's list price (see Cost at list price in the measured results); GPU time only when self-hosted.
  - Exemption lists: Federal FOIA (5 U.S.C. § 552(b)(1)-(9)) and the California Public Records Act, checked against the statutes on 25 Sep 2026; any other state as a custom list the agency checks.
  - Output: Per record: responsive or not, disposition, each redaction with its exemption, category and reason. A release PDF with a passed integrity check, the index as CSV, a response letter draft, a review memo, a signed record and a signed ledger of each officer decision.
  - Input: The request text and the records as JSON text (id, date, from, to, cc, subject, custodian, body). Up to 25 records per request hosted; set DECOSA_FOIA_MAX_DOCS on your own box. No PDF, scan or email-archive intake in this version.
- Samples:
  - 1. Coastal permits (`coastal-permits`): https://decosa.ai/legal/foia-desk?sample=1&autorun=0
  - 2. Alder point (`alder-point`): https://decosa.ai/legal/foia-desk?sample=2&autorun=0
  - 3. Ferc enron (`ferc-enron`): https://decosa.ai/legal/foia-desk?sample=3&autorun=0
- Regulatory note: A drafting aid for records officers, not legal advice and not a determination under any public-records law. Federal: 5 U.S.C. § 552(b)(1)-(9) lists the exemptions; an agency may withhold only where it reasonably foresees harm to an interest an exemption protects or where law prohibits disclosure (§ 552(a)(8)(A), FOIA Improvement Act of 2016); it must release reasonably segregable portions and mark the exemption at the place of each deletion where technically feasible (§ 552(b)); the deliberative process privilege does not apply to records created 25 years or more before the request (§ 552(b)(5)); an adverse determination must give at least 90 days to appeal and point to the FOIA Public Liaison and OGIS (§ 552(a)(6)(A)). California: the Public Records Act, Gov. Code § 7920.000 et seq. (recodified by AB 473, operative 1 Jan 2023), with § 7927.700 (personnel, medical or similar files), § 7927.705 (records exempt under other law, including Evidence Code privileges), § 7922.000 (the public-interest balancing test, under which California courts recognise the deliberative process privilege), § 7922.525(b) (segregable portions) and § 7922.540 (a written denial naming each person responsible). A custom list for another state is the agency's to check. Vaughn indexes (Vaughn v. Rosen, 484 F.2d 820 (D.C. Cir. 1973)) are filed in litigation; the index here is a draft for that use. Real requests involve personal data: run them on the agency's own machine; the hosted demo is for the fictional and public sample sets only. Model licence: Apache-2.0 (Qwen3.8-27B). The Enron emails are public records released by FERC (2003), used as a small research sample. Statutes checked against the official texts on 25 Sep 2026.

### Score an oral exam against a rubric (`oral-assessment`)

Vivas and structured interviews scored against a rubric. While it runs, the examiner sees which criteria still lack evidence and neutral follow-ups. Afterwards each criterion gets a draft score with the candidate's words cited and checked, the examiner confirms or overrides every score with a reason, and both steps are sealed in signed records that survive an appeal. Examiner side only.

- What it is: An AI oral exam assistant for vivas and structured interviews: during the session it shows the examiner which rubric criteria still lack evidence; afterwards it drafts a score for each criterion that cites the candidate's words. The examiner decides every score, and both steps are sealed in signed records that survive an appeal.
- Who it's for: Teams in education and hr and recruiting.
- Page: https://decosa.ai/tools/operations/oral-assessment
- Last reviewed: 2026-09-25
- Key result: On 120 held-out synthetic test criteria the draft level matched the label 90.8% of the time and was always within one level (quadratic weighted kappa 0.971). One AI author wrote the answers and the labels, so treat this as an upper bound on real vivas.
- Key numbers (details and caveats: https://decosa.ai/metrics/oral-assessment):
  - 90.8%: Draft level equals the label (exact) (test split, n = 120)
  - 0.971: Quadratic weighted kappa vs labels (test split, n = 120)
  - 70.8%: Misconception or mixed answers scored exactly (test split)
  - 17.6 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Voxtral 4B · MOSS-TD 0.9B · Qwen3.8-27B. Where: Self-host for real students and candidates; hosted for pilots on synthetic or consented data.
- Verification: Receipt per caption, line and model call; signed draft and review records, chained (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Voxtral, the diarizer and Qwen3.8-27B; text-only scoring needs only the Qwen server
- Output: Draft scores with cited evidence, then the examiner's decisions: two signed, chained records (the bundle) that an appeals panel can re-check.
- Page: https://decosa.ai/tools/operations/oral-assessment · JSON: https://decosa.ai/use-cases/oral-assessment.json
- Prompts: https://decosa.ai/prompts/oral-assessment-hosted.md · https://decosa.ai/prompts/oral-assessment-selfhost.md · https://decosa.ai/prompts/oral-assessment-assemble.md · https://decosa.ai/prompts/oral-assessment-mac.md
- Tiers:
  - Lite · one 48 GB card, captions only (1x L40S or RTX 6000 Ada 48 GB (not measured)): Live captions and the examiner prompt, then scores from the captions with roles guessed from question marks. No speaker labels and no uncertainty marks.
  - Standard · the hosted demo, two recognisers (1x RTX PRO 6000 Blackwell 96 GB (hosted demo uses two cards on our server)): Captions and the examiner prompt live; then speaker lines, uncertainty marks, blinding, cited scores with a grounding check, and the signed records. Every model call carries a gateway-signed receipt.
  - Best · DeepSeek-V4-Flash scores and checks (2x RTX PRO 6000 96 GB for the scorer plus the standard card): Standard with a larger scorer and grounding judge on two more 96 GB cards.
- Buyer facts:
  - Measured latency: The signed draft arrives seconds after the audio ends (median of hosted replays, gateway route, one session at a time); longer when the shared gateway is loaded.
  - Typical run cost: A fraction of a cent per short viva on the hosted route (a couple of dozen model calls at the gateway list price), plus a speech receipt per caption and per transcript line signed by the server's key at no model cost. A pasted transcript costs less. Each run shows its own measured cost.
  - What the examiner gets: Per criterion: a draft level, the candidate's lines it rests on, a grounding check of that evidence and flags (re-listen, leading question, low confidence). Then a signed bundle: the draft record and the examiner's decisions with reasons.
  - Fairness guards: The scorer never sees names, pronouns or stated personal details (0 of 32 planted ones survived). It scores content, not fluency: strong answers in plain, non-native English scored the same as fluent ones (24 of 24).
  - Data retention: Nothing stored: transcripts and records live in memory for the request or session and come back to you. The server keeps receipts (hashes), not text. For students under 18 the records keep hashes only.
  - What leaves the box (self-host): Nothing, with DECOSA_LLM_ROUTE=direct. Anyone can re-check a bundle with POST /oral/verify or a single record in the browser.
- Samples:
  - 1. Statistics viva: partly correct, with a leading question (synthetic) (`oral-stats-partial`): https://decosa.ai/tools/operations/oral-assessment?sample=1&autorun=0
  - 2. Statistics viva: strong answers, non-native speaker (synthetic) (`oral-stats-strong`): https://decosa.ai/tools/operations/oral-assessment?sample=2&autorun=0
  - 3. Statistics viva: weak answers (synthetic) (`oral-stats-weak`): https://decosa.ai/tools/operations/oral-assessment?sample=3&autorun=0
  - 4. Support specialist interview: mixed answers (synthetic) (`oral-support-interview`): https://decosa.ai/tools/operations/oral-assessment?sample=4&autorun=0
- Regulatory note: As of 25 Sep 2026, not legal advice. Education: exam recordings, transcripts and scores are education records under FERPA (34 CFR Part 99); a vendor gets them only as a school official under the institution's control, so run real exams self-hosted or under that agreement. Students under 18: set minor (hashes-only records) and self-host; COPPA applies to services directed at children under 13. EU AI Act Annex III lists evaluating learning outcomes (point 3) and evaluating candidates (point 4) as high-risk; those duties apply from 2 Dec 2027 after the AI Omnibus. Hiring in New York City: Local Law 144 requires a bias audit and candidate notice before an automated tool substantially assists the decision; this tool does not collect demographics, so it cannot run that audit itself. A person decides every score (GDPR Art. 22). Tell everyone the session is recorded and get consent where the law requires it.

### Chart claim support in a patent spec (`patent-claim-support`)

Maps every element of every claim to the specification paragraphs that describe it, with a support strength and the model's reason, one receipted call per element. Code checks antecedent basis, claim references and claim terms the specification never uses. The output is a claim chart, an issue list and a signed record, for a registered practitioner to review before filing or in a response.

- What it is: Every claim element mapped to the specification paragraphs that describe it, with antecedent-basis, claim-reference and claim-term checks, one receipt per element and a signed record.
- Who it's for: Teams in legal.
- Page: https://decosa.ai/legal/patent-claim-support
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/patent-claim-support):
  - 55% (36 of 66): Antecedent-basis flags that are genuine, random granted claims (held out, n = 66)
  - 93.4%: Supported-or-not call agrees with the labels (test split, n = 166)
  - 100%: Supported elements where a cited paragraph is one the labels list (test split, n = 150)
  - 7.8 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for unfiled inventions; hosted for published patents only.
- Verification: Receipt per element; signed record and practitioner ledger (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; parsing and the code checks run on CPU
- Output: A claim chart (CSV or Markdown) mapping each element to its supporting paragraphs, an issue list, and a signed record and practitioner ledger.
- Page: https://decosa.ai/legal/patent-claim-support · JSON: https://decosa.ai/use-cases/patent-claim-support.json
- Prompts: https://decosa.ai/prompts/patent-claim-support-hosted.md · https://decosa.ai/prompts/patent-claim-support-selfhost.md · https://decosa.ai/prompts/patent-claim-support-assemble.md · https://decosa.ai/prompts/patent-claim-support-mac.md
- Tiers:
  - Lite · code checks only, any CPU (Any CPU): Antecedent basis, claim references and claim terms, with no model and no GPU (support: false, or POST /patent/parse). No support map.
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB): Qwen3.8-27B on an RTX PRO 6000 reads every element against the specification; the code checks run alongside. Also fits a 32 GB card (estimate).
  - Wanted · a GLM-5.3-Flash judge on your own hardware (Your own hardware: 2x 96 GB cards (NVFP4, about 170-186 GB, unconfirmed) or a Mac with 192 GB or more (MLX 4-bit, 165 GB). Estimate; GLM's SGLang SM120 build hung on our server.): A stronger legal-reasoning model from another family for the partial-support calls the standard model misses (5 of 13 found). Specifications before filing stay on your hardware, never on community providers. Not served yet.
- Alternates (smaller models for the same job):
  - A second judge on one card (1x RTX PRO 6000 96 GB (80 GB of weights)), not served: A cross-family second judge that fits beside the 27B, so we would host it ourselves. sm_120 support unconfirmed.
- Buyer facts:
  - Data retention: The application is held in memory for the request; the run (chart, issues, no specification) for one hour, for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: nothing. Hosted demo: the text goes to the model through our gateway, which is why the hosted demo is for published patents and applications only.
  - Model cost per application: A few cents per application on average at the gateway's list price (measured on granted patents); GPU time only when self-hosted.
  - Input: Specification and claims as text (or one pasted application cut at "What is claimed is:"). Paragraphs numbered like [0012] are cited by number. Hosted limits: 150,000 characters of specification, 60 claims, 120 elements per request; raise them on your own box.
  - Output: A claim chart (CSV or Markdown) with the supporting paragraphs and a strength per element, an issue list (support, antecedent basis, claim references, terms), a signed record and a signed ledger of the practitioner's decisions.
- Samples:
  - 1. Ladar planted (`ladar-planted`): https://decosa.ai/legal/patent-claim-support?sample=1&autorun=0
  - 2. Ladar (`ladar`): https://decosa.ai/legal/patent-claim-support?sample=2&autorun=0
  - 3. Filter (`filter`): https://decosa.ai/legal/patent-claim-support?sample=3&autorun=0
- Regulatory note: A review aid for registered patent practitioners, not legal advice or a patentability opinion. The written-description and enablement requirements are in 35 U.S.C. 112(a) and MPEP 2163-2164; lack of antecedent basis is an indefiniteness question under 112(b) (MPEP 2173.05(e)); dependent and multiple dependent claims follow 112(d)-(e) and 37 CFR 1.75(c); claim terms must find clear support or antecedent basis in the description (37 CFR 1.75(d)(1)). A supported element here means the model found paragraphs that describe it, not that the claim is allowable. The USPTO's guidance on AI-based tools in practice (89 FR 25609, 11 Apr 2024; still the current practitioner guidance in the Federal Register as of 25 Sep 2026) applies the duty of candor and the signature rule (37 CFR 11.18) to AI-assisted papers: the signer must review them and relying on an AI tool alone is not a reasonable inquiry; it reminds practitioners to prevent disclosure of client information through AI tools (37 CFR 11.106) and warns that releasing controlled technical data to a foreign person can be a deemed export (15 CFR 734.13(b)); it says there is no general duty to tell the Office an AI tool was used unless asked, but information on AI use that is material to patentability or inventorship must be disclosed (the inventorship guidance it cites was revised on 28 Nov 2025, 90 FR 54636). Run unfiled applications on your own machine; the hosted demo takes published patents and applications only (US patent text is public domain). Model licence: Apache-2.0 (Qwen3.8-27B). Checked 25 Sep 2026.

### Make a cleared song (`music-gen-cleared`)

Music for ads, games and creators with a paper trail. A prompt guard refuses or strips named artists, songs, labels and voice imitation (code rules, then one receipted typed judgment), the track renders on MiniMax-Music3 or ACE-Step, and a CPU check scores it against a catalogue of openly licensed recordings. Each track ships with a C2PA credential and a signed certificate that states the model, seed, checks, similarity score and the exact licence.

- What it is: Music for ads, games and creators, with a prompt guard, a similarity check and a signed licence certificate for every track.
- Who it's for: Teams in music and creative and media.
- Page: https://decosa.ai/apps/music-gen-cleared
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/music-gen-cleared):
  - 100% (48/48): Prompt guard precision (refuse) (test split, n = 79)
  - 98.0% (48/49): Prompt guard recall (refuse) (test split, n = 79)
  - 0 of 30: Safe prompts refused (test split, n = 30)
  - 42.7 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: MiniMax-Music3 · ACE-Step 1.5 · Qwen3.8-27B · LAION CLAP. Where: Hosted or self-host.
- Verification: Receipted prompt check; C2PA credential and signed licence certificate (receipt coverage: partial). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: About 50 GB free on one 96 GB card for MiniMax-Music3 (ACE-Step needs about 15 GB); the similarity check runs on CPU
- Output: An MP3 with a C2PA credential, a similarity report, and a signed licence certificate (JSON, verifiable at /record/verify, and a readable Markdown copy).
- Page: https://decosa.ai/apps/music-gen-cleared · JSON: https://decosa.ai/use-cases/music-gen-cleared.json
- Prompts: https://decosa.ai/prompts/music-gen-cleared-hosted.md · https://decosa.ai/prompts/music-gen-cleared-selfhost.md · https://decosa.ai/prompts/music-gen-cleared-assemble.md · https://decosa.ai/prompts/music-gen-cleared-mac.md
- Tiers:
  - Lite · MIT engine, about 15 GB of GPU (1× 24-32 GB card for ACE-Step, plus the guard's model (or a remote one); CPU for the similarity check): ACE-Step 1.5 renders with the full guard, similarity check and certificate. No community-licence terms, but instrumentals and drafts below Music3's quality.
  - Standard · the hosted demo, MiniMax-Music3 on a 96 GB card (1× RTX PRO 6000 Blackwell 96 GB (about 50 GB free for Music3), CPU for the similarity check): MiniMax-Music3 songs with vocals (community licence, stated on every certificate) and ACE-Step, with the same guard, check and certificate.
- Buyer facts:
  - Data retention: The prompt and lyrics are kept only while the job is queued; runs, certificates and the refusal log hold hashes, rules and settings. Rendered audio stays in the studio's media folder. Compared reference files are not kept.
  - What leaves the box: Hosted: the brief goes to the text model through our gateway and its render service. Self-host: nothing.
  - Licence per track: Each certificate names the model's licence: the MiniMax-Music3 Community License (credit "Powered by MiniMax-Music3", USD 20M revenue threshold, use policy) or MIT for ACE-Step 1.5.
  - Cost per track: A fraction of a cent of model time for the prompt check (measured); GPU time on your own card when self-hosted.
  - Typical time: Under a minute from brief to a certified track on a quiet queue (measured); the similarity check takes seconds on CPU.
  - Outputs: MP3 with a C2PA credential, a similarity report naming the closest references and their licences, and a signed certificate (JSON and Markdown) that verifies at /record/verify.
- Samples:
  - 1. Strip demo (`strip-demo`): https://decosa.ai/apps/music-gen-cleared?sample=1&autorun=0
  - 2. Refuse demo (`refuse-demo`): https://decosa.ai/apps/music-gen-cleared?sample=2&autorun=0
  - 3. Sneaky (`sneaky`): https://decosa.ai/apps/music-gen-cleared?sample=3&autorun=0
  - 4. Ad bed (`ad-bed`): https://decosa.ai/apps/music-gen-cleared?sample=4&autorun=0
  - 5. Game loop (`game-loop`): https://decosa.ai/apps/music-gen-cleared?sample=5&autorun=0
  - 6. Public domain (`public-domain`): https://decosa.ai/apps/music-gen-cleared?sample=6&autorun=0
  - 7. Lyrics (`lyrics`): https://decosa.ai/apps/music-gen-cleared?sample=7&autorun=0
- Regulatory note: Not legal advice and not a rights clearance. Licences, read from the licence files on 25 Sep 2026: MiniMax-Music3 Community License (commercial use allowed; show "MiniMax-Music3" prominently in a commercial UI, §3.1; prior written authorization from MiniMax above USD 20M yearly revenue, §3.2; a hosted service must implement, test and review safeguards against infringing uses and outputs and is responsible for its downstream users, §4; the Acceptable Use Policy in Exhibit A applies, including disclosure of machine-generated content published in public places, item 11, and no impersonation without consent, item 12). ACE-Step 1.5: MIT. Qwen3.8-27B and LAION CLAP: Apache-2.0. Copyright: the US Copyright Office's Part 2 report on copyrightability (29 Jan 2025) says AI outputs are protected only where a human author determined sufficient expressive elements and that prompts alone are not enough; the certificate claims no copyright. Other countries differ. Reference catalogue: 240 Free Music Archive tracks under CC BY, CC BY-SA, CC0 or public domain (per-track licence in the catalogue), stored as vectors only. Style is not protected by copyright in the US, but a soundalike can still infringe a composition or a recording, and imitating a singer's voice can raise right-of-publicity claims: the guard refuses both kinds of request rather than judging them.

### Pre-check samples and lyrics (`sample-clearance`)

Checks a track before release against your own catalog: samples (even pitched, stretched, filtered or looped), re-played melodies and lifted lyric lines, each with its time in the track and in the reference. It compares what it finds with what the artist declared and marks each item clear, replace or remove, with a signed record. Triage for clearance staff, not legal advice.

- What it is: Samples, re-played melodies and lifted lyric lines found before release, each with its time in the track, declared or not, and a clear, replace or remove call, in one signed record.
- Who it's for: Teams in music and legal.
- Page: https://decosa.ai/tools/media/sample-clearance
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/sample-clearance):
  - 62 of 108 (57%): Audio plants found, medium and high confidence (test split, n = 108)
  - 0.87 (9 false items): Audio precision, medium and high confidence (test split)
  - 35 of 108 (32%), precision 1.00: Audio plants found, high confidence only (precision) (test split, n = 108)
  - 6.7 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Audio matching in code (no model); Qwen3.8-27B labels near-duplicate lyric lines. Where: Self-host with your own catalog; the hosted demo uses a CC BY demo catalog.
- Verification: Receipt per lyric-label call; signed record of what was checked, against which catalog (receipt coverage: partial). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: Any CPU for the audio matching and lyric spans; Qwen3.8-27B (1× RTX 5090 32 GB or larger) for the lyric labels
- Output: A clearance checklist (CSV or Markdown) with timestamps, confidence, declared status and clear, replace or remove per item, and a signed record.
- Page: https://decosa.ai/tools/media/sample-clearance · JSON: https://decosa.ai/use-cases/sample-clearance.json
- Prompts: https://decosa.ai/prompts/sample-clearance-hosted.md · https://decosa.ai/prompts/sample-clearance-selfhost.md · https://decosa.ai/prompts/sample-clearance-assemble.md · https://decosa.ai/prompts/sample-clearance-mac.md
- Tiers:
  - Lite · code only, any CPU (Any CPU): Audio matching, exact lyric spans and trigram near-duplicates with no model and no GPU (judge_lyrics: false). Re-worded lyric lines are not labelled; stock phrases are not dismissed by a model.
  - Standard · adds lyric labels by Qwen3.8-27B (hosted demo) (Any CPU plus 1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate) for the model): Everything in Lite, plus MiniLM embeddings on CPU and the model's lift, variant or stock label for each near-duplicate lyric line, with a receipt per call.
- Buyer facts:
  - Data retention: The audio and lyrics are held in memory for the request; the run (checklist and record, no audio) for one hour, for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: nothing (the lyric labels go to your own model server). Hosted demo: the audio is matched on Decosa's server; only near-duplicate lyric line pairs go to the model through our gateway.
  - Your catalog: A folder of audio files and a catalog.json (title, artist, rights per recording); the index builds on first start. Melody references (a rendering of each composition) enable the interpolation check. Bring your own lyric set, or use the public-domain one.
  - Cost per track: A fraction of a cent in model time at the gateway's list price for the planted demo (one label call); the audio matching is CPU time only.
  - Output: A checklist (CSV or Markdown): each item's kind, reference and rights, times in the track and in the reference, confidence, declared status, and clear, replace or remove with the reason and how to clear it; declared samples not in the catalog; and a signed record.
- Samples:
  - 1. Planted (`planted`): https://decosa.ai/tools/media/sample-clearance?sample=1&autorun=0
  - 2. Clean (`clean`): https://decosa.ai/tools/media/sample-clearance?sample=2&autorun=0
- Regulatory note: Triage, not legal advice; checked 25 Sep 2026. In the US a sample usually needs two licences: the sound recording (the master, 17 U.S.C. 114, usually from the label) and the musical work in it (from its publishers). An interpolation re-records the music, so it needs the composition licence only; the compulsory mechanical licence does not cover changing a song's melody or fundamental character (17 U.S.C. 115(a)(2)), so interpolations are negotiated. Whether a very short sample needs a licence is unsettled: Bridgeport Music v. Dimension Films, 410 F.3d 792 (6th Cir. 2005) held there is no de minimis defence for sampling a sound recording ("get a license or do not sample"); VMG Salsoul v. Ciccone, 824 F.3d 871 (9th Cir. 2016) held there is (a 0.23-second horn hit in "Vogue"). The Supreme Court has not resolved the split as far as we found, so the checklist never treats a short sample as safe. In the EU, Pelham v Hütter (CJEU C-476/17, 29 Jul 2019) held that even a very short audio fragment is a reproduction unless it is changed so it is unrecognisable to the ear. Public domain in the US: sound recordings published before 1923 since 1 Jan 2022 (Music Modernization Act, 17 U.S.C. 1401), later ones 100 years after publication (recordings from 1925 on 1 Jan 2026); compositions and lyrics published in 1930 or earlier as of 1 Jan 2026 (the demo lyric set stops at 1928). Creative Commons: BY allows sampling with credit; NC forbids a commercial release and ND forbids adaptations. AcoustID (the public Chromaprint service) is free for non-commercial use and paid for commercial use; this product does not call it. Model licences: Apache-2.0 (Qwen3.8-27B, all-MiniLM-L6-v2).

### Check split sheets and metadata (`split-sheet-check`)

Reconciles split sheets, release metadata (DDEX ERN or a distributor CSV), contract excerpts and PRO or MLC registration exports. Code checks every ISRC, ISWC, IPI and UPC check digit and adds up every split; an open model reads the messy documents, and every value it reads must be found in the document. You get each discrepancy with the lines it came from, a proposed credit sheet to confirm, and a signed record.

- What it is: A split sheet checker for a release: it reconciles split sheets with the DDEX or distributor metadata, contracts and PRO or MLC registrations, lists every discrepancy with the lines it came from, and proposes a credit sheet a person confirms before code re-checks and signs it.
- Who it's for: Teams in music.
- Page: https://decosa.ai/tools/media/split-sheet-check
- Last reviewed: 2026-09-25
- Key result: On 30 synthetic test catalogs, run once after dev was frozen, it found all 43 planted errors. Precision was 19 of 19 on typed split sheets but 27 of 52 on catalogs with a scanned sheet, where every false alarm traced to OCR. The catalogs come from our own generator.
- Key numbers (details and caveats: https://decosa.ai/metrics/split-sheet-check):
  - 43 of 43: Planted errors found (test split, n = 43)
  - 64.8%: Precision, all catalogs (test split)
  - 19 of 19: Precision, typed split sheets only (test split, n = 19)
  - 6.7 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Hosted or self-host; unreleased catalogs and contracts on your own machine.
- Verification: Receipt per model call; signed check record and signed confirmed credit sheet (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; parsing, OCR and every check run on CPU
- Output: A reconciliation report with every discrepancy and its sources, a confirmed credit sheet (CSV), and signed records of the check and the confirmation.
- Page: https://decosa.ai/tools/media/split-sheet-check · JSON: https://decosa.ai/use-cases/split-sheet-check.json
- Prompts: https://decosa.ai/prompts/split-sheet-check-hosted.md · https://decosa.ai/prompts/split-sheet-check-selfhost.md · https://decosa.ai/prompts/split-sheet-check-assemble.md · https://decosa.ai/prompts/split-sheet-check-mac.md
- Tiers:
  - Lite · code only, any CPU (Any CPU): DDEX, distributor and registration files parsed; every identifier and share sum checked; sources compared. Split sheets and contracts are not read (POST /splits/parse or "extract": false).
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB): Qwen3.8-27B reads the split sheets and contracts and answers the typed field questions; everything else is the lite tier's code. Also fits a 32 GB card (estimate).
- Alternates (smaller models for the same job):
  - A vision-language reader for scans (1x 96 GB card), not served: Read scanned and handwritten split sheets as images instead of through OCR. Page 32 suggests the standard reader's own vision tower (Qwen3.8-27B, same weights) plus a layout model instead of this Jan 2025 7B.
- Buyer facts:
  - Data retention: Documents are held in memory for the request; the run (report and proposal) for one hour, for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: nothing. Hosted: the documents go to the model through our gateway, which is why the hosted demo is for released metadata and test catalogs.
  - Model cost per release: A fraction of a cent per release on average at the gateway's list price (measured on synthetic releases); GPU time only when self-hosted.
  - Input: Split sheets and contracts as text, PDF, DOCX or scans (PNG, JPEG, image-only PDF); release metadata as DDEX ERN XML or CSV; registrations as CSV. Hosted limits: 12 documents, 8 of them read by the model, 4 MB per file.
  - Output: Discrepancies with the source lines, a proposed credit sheet (CSV) that a person confirms, a Markdown report, and signed records of the check and the confirmed sheet.
- Samples:
  - 1. Harbor lights planted (`harbor-lights-planted`): https://decosa.ai/tools/media/split-sheet-check?sample=1&autorun=0
  - 2. Harbor lights clean (`harbor-lights-clean`): https://decosa.ai/tools/media/split-sheet-check?sample=2&autorun=0
  - 3. Copper moon scan (`copper-moon-scan`): https://decosa.ai/tools/media/split-sheet-check?sample=3&autorun=0
- Regulatory note: A checking aid for songwriters, managers, publishers and distributors, not legal or accounting advice. Background, checked 25 Sep 2026: a recording carries two copyrights, the sound recording (identified by ISRC, ISO 3901) and the musical work (ISWC, ISO 15707); writers and publishers are identified by CISAC IPI numbers and releases by GS1 UPC/EAN. In the US the performing-rights societies (ASCAP, BMI, SESAC, GMR) pay the writer's and the publisher's share separately. The Orrin G. Hatch-Bob Goodlatte Music Modernization Act (signed 11 Oct 2018) created a blanket mechanical licence for digital services administered by the Mechanical Licensing Collective (The MLC, designated by the Copyright Office in July 2019; the blanket licence has been available since 1 Jan 2021); royalties The MLC cannot match to a registered work are held and, after a statutory holding period, can be distributed by market share to other copyright owners (17 U.S.C. 115(d)(3)(H)-(J)). A valid check digit means an identifier is well formed, not that it is registered to the person named; the writers agree the splits, and a society pays on its own registration. Rules and forms differ by society and territory. Writer names and IPI numbers are personal data and deal terms are confidential: use the hosted demo for released metadata and test catalogs, and self-host for unreleased catalogs and contracts. Model licence: Apache-2.0 (Qwen3.8-27B); OCR: Tesseract (Apache-2.0).

### Check a paper's citations (`paper-claim-check`)

Checks every citing sentence of a manuscript against the cited paper's open full text and quotes the passage: supported, overstated, not supported or contradicted, one receipted call per citation. Each reference is looked up in Crossref and OpenAlex for retractions, corrections, wrong DOIs and years, and self-citation stacks are flagged. The output is an issue list and a signed report for an editor, reviewer or integrity office to act on.

- What it is: Every citing sentence checked against the cited paper's open full text, with the passage quoted; retracted and corrected sources, wrong DOIs and years, and self-citation stacks flagged; one receipt per check and a signed report.
- Who it's for: Teams in science and research.
- Page: https://decosa.ai/tools/life-sciences/paper-claim-check
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/paper-claim-check):
  - 68 of 68 / 68 of 68: Retraction flag on known retracted works, with DOI / without (held out, n = 68)
  - 0 of 60 / 0 of 60: Retraction flag on random control articles, with DOI / without (held out, n = 60)
  - 12 of 12: Planted wrong-paper citations flagged (test split, n = 12)
  - 9.6 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Hosted for published papers and your own drafts; self-host for manuscripts under review.
- Verification: Receipt per citation checked; signed report (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; parsing and the reference checks run on CPU
- Output: An issue list, every citation with its verdict and quoted passage, a Markdown or JSON report and a signed record.
- Page: https://decosa.ai/tools/life-sciences/paper-claim-check · JSON: https://decosa.ai/use-cases/paper-claim-check.json
- Prompts: https://decosa.ai/prompts/paper-claim-check-hosted.md · https://decosa.ai/prompts/paper-claim-check-selfhost.md · https://decosa.ai/prompts/paper-claim-check-assemble.md · https://decosa.ai/prompts/paper-claim-check-mac.md
- Tiers:
  - Lite · references only, any CPU (Any CPU): Retractions, corrections, wrong DOIs, years and first authors, missing and uncited references, and self-citation stacks, with no model and no GPU (check_claims: false). No claim check.
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB): Qwen3.8-27B on an RTX PRO 6000 reads every citing sentence against the cited paper's open text and reviews its own flags; the reference checks run alongside. Also fits a 32 GB card (estimate).
  - Wanted · GLM-5.3-Flash for the claim check (Network providers: 2x 96 GB cards or a Mac with 192 GB or more (about 170 GB of weights). Estimate.): A stronger reasoning model from another family for the background citations the standard model flags too eagerly. It needs two 96 GB cards or a large Mac, so community providers are welcome. Not served yet.
- Buyer facts:
  - Data retention: The manuscript is held in memory for the request; the run (verdicts, issues, the text for exports) for one hour, for the token or key that made it. Nothing is written to disk; logs carry counts only.
  - What leaves the box: Self-hosted: only DOIs, identifiers and reference strings, to Crossref, OpenAlex, Europe PMC, arXiv and doi.org. Hosted demo: also the text, to the model through our gateway, so use it for published papers or your own drafts.
  - Model cost per paper: A few cents to about ten cents per paper at the gateway's list price, rising with the number of checked citations (measured); the metadata services are free.
  - Input: Text with a reference list under a heading such as "References", or a PDF, DOCX or LaTeX file; numbered [3] or author-year (Smith et al., 2020) citations. Hosted limits: 250,000 characters, 250 references, 60 checked citations per request.
  - Output: An issue list, every citation with a verdict and the quoted passage, a Markdown or JSON report, and a signed record verifiable at /record/verify.
  - Politeness: Identifies itself with a contact address to every service, stays under Crossref's published polite-pool limits, uses only OpenAlex lookups that cost no credits, and fetches arXiv at most once every 3 seconds.
- Samples:
  - 1. Telecommuting planted (`telecommuting-planted`): https://decosa.ai/tools/life-sciences/paper-claim-check?sample=1&autorun=0
  - 2. Telecommuting (`telecommuting`): https://decosa.ai/tools/life-sciences/paper-claim-check?sample=2&autorun=0
- Regulatory note: A screening aid, not a finding of research misconduct. Citation manipulation, including self-citation stacking, is treated by journals under COPE's guidance, and an allegation is for the journal or institution to investigate. A manuscript under peer review is confidential: NIH prohibits its peer reviewers from using generative AI tools to analyse or critique applications (notice NOT-OD-23-149, June 2023), and many publishers tell reviewers not to upload manuscripts to AI tools, so reviewers and editors should run the self-hosted version on their own machine. Only DOIs, identifiers and reference strings are sent to Crossref, OpenAlex, Europe PMC, arXiv and doi.org; the manuscript's text is not. Open-access full text is read under each article's own licence and only short passages are quoted back. Retraction and correction data come from Crossref (which has made the Retraction Watch database openly available since September 2023) and OpenAlex, and can lag the publishers. Model licence: Apache-2.0 (Qwen3.8-27B). Checked 25 Sep 2026.

### Keep a signed lab notebook (`signed-lab-notebook`)

A tamper-evident lab notebook with an AI-analysis log. Entries hold text and instrument files by SHA-256, amendments point at the original and give a reason, and witnesses and approvers sign with their own keys. When a model analyses data, the prompt, the exact data files, the model and its signed receipt go on the record, so a reviewer sees which conclusions were AI-assisted. RFC 3161 timestamps from an independent authority, an audit-trail export and a verify page anyone can run.

- What it is: A tamper-evident lab notebook that shows which conclusions an AI model helped with, and lets an inspector check it without trusting you.
- Who it's for: Teams in science and research and compliance and trust.
- Page: https://decosa.ai/tools/life-sciences/signed-lab-notebook
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/signed-lab-notebook):
  - 300 / 300: Genuine synthetic exports that verify (synthetic, n = 300)
  - 1,400 / 1,400: Outsider alterations caught (editing the export without keys) (synthetic, n = 1400)
  - 515 of 520: Insider alterations caught, personal keys + TSA, export alone (synthetic, n = 520)

- Status: live. Models: Qwen3.8-27B (AI analyses). Where: Hosted for synthetic or public data; self-host for unpublished results.
- Verification: Receipt per AI analysis; signed hash chain, personal e-signatures, RFC 3161 timestamps (receipt coverage: full). Hosted QA 2026-09-26: pass (pre-release server). Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the analysis model; the notebook, signatures and verification run on CPU
- Output: A signed notebook export (JSON) with an audit-trail CSV for an inspector, checkable in the browser, with RFC 3161 timestamps.
- Page: https://decosa.ai/tools/life-sciences/signed-lab-notebook · JSON: https://decosa.ai/use-cases/signed-lab-notebook.json
- Prompts: https://decosa.ai/prompts/signed-lab-notebook-hosted.md · https://decosa.ai/prompts/signed-lab-notebook-selfhost.md · https://decosa.ai/prompts/signed-lab-notebook-assemble.md · https://decosa.ai/prompts/signed-lab-notebook-mac.md
- Tiers:
  - Lite · notebook, signatures and timestamps on CPU (self-host) (Any CPU): Everything except AI analyses: entries, amendments, personal e-signatures, RFC 3161 timestamps, the signed export and the verify page. Analyses you run elsewhere can be recorded as ordinary entries, without a receipt.
  - Standard · Qwen3.8-27B writes receipted AI analyses (hosted demo) (1× RTX PRO 6000 96 GB (measured) or 1× RTX 5090 32 GB (estimate)): Adds the AI-analysis log: the model writes a note from the selected entries and attached data, with a signed receipt, the exact prompt and the data hashes on the record. This is what the hosted API runs.
- Buyer facts:
  - What you send: Entry text, file hashes (files are hashed in your browser), members and their public keys, signatures, and the data text for an AI analysis (up to 60,000 characters).
  - What you get: A signed export (JSON) with every entry, amendment, AI analysis, e-signature and RFC 3161 token; an audit-trail CSV; a verify page an inspector runs in their browser.
  - Typical cost: One model call per AI analysis: a fraction of a cent at the gateway list price. Entries, signatures, timestamps and exports are CPU.
  - Retention (hosted): Demo notebooks expire after 14 days. Self-host keeps notebooks as append-only files in your data volume for as long as you set.
  - What leaves the box (self-host): Only the timestamp request: a 32-byte digest of the chain head, sent to the TSA. On the hosted route the entries and analysis data reach our server and the gateway.
  - Part 11: Supports the audit-trail, signature-manifestation and signature/record-linking requirements. Validation, SOPs, SSO and two-component sign-in are yours to provide.
- Samples:
  - 1. Enzyme kinetics (`enzyme-kinetics`): https://decosa.ai/tools/life-sciences/signed-lab-notebook?sample=1&autorun=0
- Regulatory note: Checked 26 Sep 2026. 21 CFR Part 11 sets when FDA treats electronic records and signatures as equivalent to paper for records other FDA rules require (the predicate rules). The notebook supports the audit-trail requirement of 11.10(e) (secure, computer-generated, time-stamped records of who created or changed what, without obscuring earlier entries), the signature manifestation of 11.50 (printed name, date and time, and meaning) and the signature/record linking of 11.70. Compliance also needs the lab's validation, written procedures, training, access control and signature controls under 11.100 to 11.300 (for example two distinct identification components for non-biometric signatures); software alone does not make anyone compliant. FDA's guidance Part 11, Electronic Records; Electronic Signatures: Scope and Application (August 2003) narrowed enforcement (validation, audit trails, record retention and copying) while the predicate rules still apply; the final guidance Electronic Systems, Electronic Records, and Electronic Signatures in Clinical Investigations: Questions and Answers (October 2024, docket FDA-2017-D-1105) covers clinical investigations. The trusted timestamps follow RFC 3161; their weight depends on the Time-Stamp Authority (FreeTSA, the default, is free with no SLA; an eIDAS-qualified or commercial TSA can be configured). AI analyses are recorded and receipted, not validated: a person must still check them. Model licence: Apache-2.0 (Qwen3.8-27B). Not legal or regulatory advice.

### Check green claims in copy (`green-claims-check`)

Checks every environmental claim in pack copy, a product page or an ad script against the EU bans that apply from 27 September 2026 (Directive (EU) 2024/825) or the UK CMA Green Claims Code: generic claims, offset-based climate neutral, own-brand labels, whole-product claims, legal requirements sold as features and future targets without a plan. Each claim gets a verdict with the rule cited, the evidence-file line that backs it, and a rewrite; you get a CSV claim table and a signed report. A triage, not a compliance sign-off.

- What it is: A green claims checker for marketing copy: each environmental claim is judged against the EU bans that apply from 27 September 2026 (Directive (EU) 2024/825) or the UK CMA Green Claims Code, with the rule cited, the evidence line that backs it, and a rewrite that says only what the evidence supports. A triage for a reviewer, not a sign-off.
- Who it's for: Teams in compliance and trust.
- Page: https://decosa.ai/tools/finance/green-claims-check
- Last reviewed: 2026-09-25
- Key result: On 6 held-out synthetic pieces (44 sentences) the first run got 43 of 44 verdicts right and flagged 19 of 19 planted violations with no false alarms; later reruns gave 41 of 44. The pieces are small, and the same author wrote them and the prompts.
- Key numbers (details and caveats: https://decosa.ai/metrics/green-claims-check):
  - 43/44 (98%): Verdict accuracy, v2 on test2 (held out, first run) (held out, n = 44)
  - 19/19: Planted violations flagged, v2 on test2 (held out) (held out, n = 19)
  - 0/25: False alarms on clean claims and non-claims, v2 on test2 (held out) (held out, n = 25)
  - 8.6 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B. Where: Hosted for published or synthetic copy; self-host for unreleased copy and supplier contracts.
- Verification: Receipt per model call; signed report (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for the model; the rulepack, grounding spans and the report run on CPU
- Output: A claim table (CSV) with verdict, rule, quoted evidence and rewrite per claim, a Markdown report and a signed JSON record.
- Page: https://decosa.ai/tools/finance/green-claims-check · JSON: https://decosa.ai/use-cases/green-claims-check.json
- Prompts: https://decosa.ai/prompts/green-claims-check-hosted.md · https://decosa.ai/prompts/green-claims-check-selfhost.md · https://decosa.ai/prompts/green-claims-check-assemble.md · https://decosa.ai/prompts/green-claims-check-mac.md
- Tiers:
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B types each sentence, judges support against the evidence and reads the evidence file, each in its own receipted call; the rulepack decides the verdict. This is what the hosted API runs.
- Alternates (smaller models for the same job):
  - Pack artwork and other languages (1x RTX PRO 6000 96 GB (estimate)), not built: Read claims in pack artwork, badges and colours from images with the 27B's own vision input, and check copy in other EU languages. Not built.
- Buyer facts:
  - Data retention: Nothing stored: copy and evidence live in memory for the request. The signed report holds hashes, offsets, verdicts, rule ids and receipt ids, never the copy.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and its receipt (hashes, token counts, no text) is kept by the gateway and this API. Self-hosted on the direct route: nothing leaves the box.
  - Input formats: Text only: copy up to 8,000 characters and 40 claim sentences; up to 10 evidence files (LCA, certificate, offset contract, plan, policy, other) and 200,000 characters. Extract PDF text yourself. Evidence is optional, but without it nothing can be substantiated.
  - What it will not say: Never 'compliant'. A clean result reads 'nothing flagged (not a compliance sign-off)'. UK mode never says 'banned', because the CMA code is guidance.
  - Typical run: The Fernhollow sample: a few dozen model calls and about a cent at the gateway list price; the eval average per piece is lower. Each run shows its own measured cost.
- Samples:
  - 1. Fernhollow planted (`fernhollow-planted`): https://decosa.ai/tools/finance/green-claims-check?sample=1&autorun=0
  - 2. Quillbrook compliant (`quillbrook-compliant`): https://decosa.ai/tools/finance/green-claims-check?sample=2&autorun=0
  - 3. Northwick uk (`northwick-uk`): https://decosa.ai/tools/finance/green-claims-check?sample=3&autorun=0
- Regulatory note: Checked 25 Sep 2026 against the primary sources. EU: Directive (EU) 2024/825 (OJ L, 6.3.2024; EUR-Lex link under Tools) amends the Unfair Commercial Practices Directive 2005/29/EC. Member States had to adopt their measures by 27 March 2026 and apply them from 27 September 2026 (Article 4(1)); national transposing laws are what apply, and this tool does not check national differences. It encodes the new Annex I bans (2a sustainability labels without a certification scheme or public authority, 4a generic claims without recognised excellent environmental performance, 4b whole-product or whole-business claims when only a part qualifies, 4c offset-based neutral, reduced or positive greenhouse-gas claims, 10a legal requirements presented as a distinctive feature) and Article 6(2)(d) (future performance claims without a verified implementation plan), which is case by case, so it can only raise 'needs substantiation'. The durability, repair and software bans (23d to 23j) are not checked. The proposed Green Claims Directive (2023/0085(COD)) is not law and is not encoded: on 20 June 2025 the Commission announced its intent to withdraw it and the last trilogue was cancelled; the 2026 work programme lists it as pending (European Parliament Legislative Train, updated 1 Aug 2026). UK: the CMA Green Claims Code (20 Sep 2021; GOV.UK link under Tools) is guidance, not a list of per se bans, so UK mode never says 'banned'; since April 2025 the CMA can itself fine up to 10% of global turnover for consumer-law breaches under the DMCC Act 2024 (CMA press release, 7 Apr 2025). Text only: artwork, colours and label images are not read. Verdicts come from a language model plus fixed rules and can be wrong in both directions (see the eval). Not legal advice, and a clean result is not a compliance sign-off. The hosted demo keeps nothing; unreleased copy and supplier contracts belong on a self-hosted box. Model licence: Apache-2.0 (Qwen3.8-27B).

### Record F&I add-on disclosures (`fi-disclosure-record`)

Checks a recorded F&I-office conversation against its deal jacket under California's CARS Act (SB 766, in force 1 Oct 2026). For each add-on it asks whether the add-on was said to be optional and whether it was presented as required, with a quote and a time. It flags add-ons that were declined or never discussed but are on the contract, and prices or payments that differ. It reads the written disclosures, the 3-day cancel right and banned add-ons from the jacket, and seals it all in a signed record to keep for two years.

- What it is: What the F&I manager said about each add-on, quoted with its time, checked against the deal jacket under California's CARS Act.
- Who it's for: Teams in sales and marketing and finance and insurance.
- Page: https://decosa.ai/tools/finance/fi-disclosure-record
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/fi-disclosure-record):
  - 31/32: Per-check accuracy, test B (held out), run 1 (held out, n = 32)
  - 8/8: Planted answers found, test B (held out), run 1 (held out, n = 8)
  - 1 of 14: False alarms on clean calls, test B (held out), run 1 (held out, n = 14)
  - 3.3 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: MOSS-Transcribe-Diarize · Qwen3.8-27B. Where: Self-host for real deals (hosted demo: synthetic role-plays only).
- Verification: Receipt per model call; every yes backed by a quote found in the transcript; signed hash-chained record with a retention date (receipt coverage: full). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Qwen3.8-27B and the diarizer; the jacket rules and the record need no GPU
- Output: A checklist of typed answers with quotes and times, the jacket mismatches, the jacket-only rules, and a signed record with its retention date (creation plus two years).
- Page: https://decosa.ai/tools/finance/fi-disclosure-record · JSON: https://decosa.ai/use-cases/fi-disclosure-record.json
- Prompts: https://decosa.ai/prompts/fi-disclosure-record-hosted.md · https://decosa.ai/prompts/fi-disclosure-record-selfhost.md · https://decosa.ai/prompts/fi-disclosure-record-assemble.md · https://decosa.ai/prompts/fi-disclosure-record-mac.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same checks on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B extracts and answers every typed check; MOSS-Transcribe-Diarize turns recordings into transcripts. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, many-speaker F&I sessions.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Call recordings stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: What was said in the F&I office about each add-on, consent, payments and the 3-day cancel right, against the deal jacket. It also checks the jacket's written disclosures and banned add-ons under California SB 766.
  - What it does not do: It gives no compliance verdict. It does not check advertising or the first written price quote (1784.41(a)), whether a GAP loan-to-value is lawful, add-on payment timing, or whether a written disclosure was clear, conspicuous and in the negotiation language.
  - Consent: Every run needs a statement that all parties agreed to the recording (Penal Code 632). The statement goes into the signed record, and a check looks for the consent on the recording itself.
  - Data kept: Nothing on the server. Transcripts, jackets and records live in memory for the request, and logs carry counts and timings only. You keep the signed record for two years (1784.44(a)).
  - What leaves the box (hosted demo): The transcript and the jacket's add-on list go to Qwen3.8-27B through our gateway. The gateway's receipts hold hashes, not text. Self-hosted, nothing leaves.
  - Model calls per deal: One extraction, plus one typed check per question: 2 per add-on, 3 general, and 2 more on used cars at $50,000 or less. That is 10 for a used car with two add-ons.
  - Typical run cost: A fraction of a cent at the gateway list price for a used-car deal with add-ons (measured on the smoke sample). Each run shows its own measured cost.
- Samples:
  - 1. T7 declined and price (`t7-declined-and-price`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=1&autorun=0
  - 2. T5 gap required (`t5-gap-required`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=2&autorun=0
  - 3. Tb4 no consent no cooling off (`tb4-no-consent-no-cooling-off`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=3&autorun=0
  - 4. T6 no cancel high fee (`t6-no-cancel-high-fee`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=4&autorun=0
  - 5. T1 clean used (`t1-clean-used`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=5&autorun=0
  - 6. T8 no optional wrong payment (`t8-no-optional-wrong-payment`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=6&autorun=0
  - 7. T10 term ev (`t10-term-ev`): https://decosa.ai/tools/finance/fi-disclosure-record?sample=7&autorun=0
- Regulatory note: Checked 25 Sep 2026 against the chaptered text. California SB 766, the California Combating Auto Retail Scams (CARS) Act (Stats. 2025, ch. 354; approved and filed 6 Oct 2025), adds Civil Code 1784.20-1784.44, operative 1 Oct 2026. It bans misrepresenting costs, terms or any aspect of an add-on (1784.40). It requires, in writing, that an add-on is not required and the car can be bought without it, the total paid at a quoted monthly payment, and a warning that lower payments often increase the total (1784.41(b)-(d)). It bans charging for add-ons the buyer cannot benefit from, with seven examples such as oil changes on an EV (1784.42(a)). It gives a 3-day right to cancel on used vehicles sold or leased at $50,000 or less, with a separate disclosure and a restocking fee of 1.5% of the price ($200 minimum, $600 maximum) plus $1 a mile over 250 (up to $150) (1784.31(g), 1784.43). It requires dealers to keep records showing compliance for two years (1784.44). The written disclosures cannot be met by speech: this tool reads them from the jacket and checks whether what was said agrees. Recording an F&I conversation needs the consent of every party (Penal Code 632(a)). The federal FTC CARS Rule was vacated by the Fifth Circuit on 27 Jan 2025 (NADA v. FTC, No. 24-60013) and is context only. Not legal advice.

### Dub a video in your own voice (`consented-dubbing`)

A Spanish track for a creator's video in the creator's own consented voice, or in a consented dubber's. The consent ledger is checked when the job starts, before the voice is rendered and again at approval. The track is cut to the video's exact length, the original track is never replaced, and subtitles get a QA pass. Nothing is published until a person approves. Approval seals a signed record and a C2PA label: "AI-dubbed, voice consented by" the named person.

- What it is: Your voice, your approval: a Spanish track cut to the exact video length, released only after you approve it.
- Who it's for: Teams in film, tv and games and creative and media.
- Page: https://decosa.ai/apps/consented-dubbing
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/consented-dubbing):
  - 11 of 11: Dub tracks exactly the video's length (sample count) (synthetic, n = 11)
  - 9 of 9: Consent gate decisions as expected (synthetic, n = 9)
  - 3.9% (own-voice), 1.0% (dubber-voice): ASR word error rate against the narration script (synthetic)
  - 74.0 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Voxtral Mini 4B Realtime · Qwen3.8-27B · Chatterbox Multilingual. Where: Hosted with the demo's fictional voices; self-host with your own consented voices.
- Verification: Receipts for every ASR segment and model call; signed consent decisions; signed record and C2PA manifest at approval (receipt coverage: partial). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Qwen3.8-27B and Voxtral; the voice model needs about 4 GB of GPU or runs on CPU at about 4× real time
- Output: A C2PA-labelled WAV cut to the exact video length, SRT and WebVTT subtitles with a QA report, a hand-off script for a human dubber, and a signed record.
- Page: https://decosa.ai/apps/consented-dubbing · JSON: https://decosa.ai/use-cases/consented-dubbing.json
- Prompts: https://decosa.ai/prompts/consented-dubbing-hosted.md · https://decosa.ai/prompts/consented-dubbing-selfhost.md · https://decosa.ai/prompts/consented-dubbing-assemble.md
- Tiers:
  - Lite · voice on CPU (A 96 GB card (or remote endpoints) for Voxtral and Qwen3.8-27B; the voice on a 16-thread CPU): The same pipeline and gates with the voice model on CPU: no GPU needed for the voice, about five times slower end to end.
  - Standard · the hosted demo, voice on a shared GPU (1x RTX PRO 6000 96 GB for the text models plus about 4 GB of GPU for the voice while a job runs): Voxtral, Qwen3.8-27B through our gateway, and Chatterbox Multilingual on GPU when 8 GB are free (CPU otherwise).
- Alternates (smaller models for the same job):
  - Better voice and lip-sync (One 96 GB card (not measured)), not served: Chatterbox Multilingual V3 or its LatAm Spanish pack for the voice, and InfiniteTalk (Apache-2.0, Wan2.1-14B based) for lip-sync on talking heads. Neither has been run here.
- Buyer facts:
  - Consent: Every voice render checks the consent ledger three times (submit, render, approval), and each decision is signed. Uploaded audio is only transcribed, never used as a voice.
  - Length: The dub WAV has exactly round(video length x 48,000) samples: 11 of 11 real runs and 6 of 6 awkward test files.
  - Approval: Nothing is released until a person approves the exact draft they reviewed. The release gets a signed record and a C2PA label: "AI-dubbed, voice consented by <name> (consent entry <id>)". The original track is never replaced.
  - Cost per video: A fraction of a cent of model time for a short video (a few receipted translation calls, measured); the voice runs on your own GPU or CPU.
  - Typical time: About a minute from upload to review for a short video with the voice on a GPU; several minutes with the voice on CPU (measured).
  - Data retention: Jobs, uploads and files are deleted after 24 hours. Logs hold ids, counts and timings, never text.
  - What leaves the box: Hosted: the video, and the translation calls through our gateway. Self-host: nothing, unless you point translation at the gateway.
  - For human dubbers: A hand-off export (JSON and CSV) with timecodes, slot lengths, character budgets, the glossary and the machine draft, plus a subtitle QA that works on human-made files.
- Samples:
  - 1. Own voice (`own-voice`): https://decosa.ai/apps/consented-dubbing?sample=1&autorun=0
  - 2. Dubber voice (`dubber-voice`): https://decosa.ai/apps/consented-dubbing?sample=2&autorun=0
  - 3. Revoked (`revoked`): https://decosa.ai/apps/consented-dubbing?sample=3&autorun=0
  - 4. Strike (`strike`): https://decosa.ai/apps/consented-dubbing?sample=4&autorun=0
  - 5. Wrong territory (`wrong-territory`): https://decosa.ai/apps/consented-dubbing?sample=5&autorun=0
- Regulatory note: Not legal advice. Laws read on 25 Sep 2026 (the consent ledger's sources, linked there): Tennessee's ELVIS Act (in force 1 Jul 2024) makes whoever distributes a tool whose primary purpose is producing someone's voice without authorization liable. That is why every render here goes through the consent ledger. California AB 2602 (signed 17 Sep 2024) and New York S7676B (contracts from 1 Jan 2025) make a replica clause that replaces in-person work unenforceable without a reasonably specific use description and counsel or union representation; the ledger stores both. EU AI Act Art. 50(2) (applies from 2 Aug 2026) requires machine-readable marking of synthetic audio; the C2PA manifest and the Perth watermark cover it. Credentials use a development certificate, so public validators show the issuer as untrusted. YouTube's disclosure rules (Help 14328491, read 25 Sep 2026) exempt cloning your own voice for dubs; a dubber's clone of someone else is not in that exemption, so label it. Guild terms (SAG-AFTRA 2023 and 2026) were not verified against the contract text.

### Pre-flight a synthetic-performer ad (`disclosure-preflight`)

Checks an ad or short before it ships. It finds synthetic performers from the file's own provenance (its C2PA credential, Decosa's render receipt or watermark) and what you declare, and reads the frames to see whether an AI label is on screen. It burns in the label and signs a C2PA marking when the New York or EU rules need them, checks each real performer's consent by ledger id, and flags real people, other brands, profanity, rating triggers and music cues in the script. A named reviewer signs off, and the sign-off is sealed into an AI-use attestation. Advisory, not a legal clearance.

- What it is: Before an ad ships: who in it is synthetic and how we know, a visible AI label and a C2PA marking when the rules need them, consent by ledger id, script flags with their lines, and a named reviewer's sign-off in a signed AI-use attestation.
- Who it's for: Teams in film, tv and games and sales and marketing.
- Page: https://decosa.ai/tools/media/disclosure-preflight
- Last reviewed: 2026-09-25
- Key numbers (details and caveats: https://decosa.ai/metrics/disclosure-preflight):
  - 27 of 29 (93%): Planted script issues found (full config) (test split, n = 29)
  - 1.00 (27 flags, 0 false): Precision of script flags (full config) (test split, n = 27)
  - 0 of 8: Clean scripts with a flag (test split, n = 8)
  - 3.6 s: Median end-to-end run, hosted (QA sweep 2026-09-25)

- Status: live. Models: Qwen3.8-27B reads the script (typed yes/no per candidate); code reads provenance, OCRs the label and adds it. Where: Hosted or self-host; unreleased ads on your own machine.
- Verification: Receipt per model call; receipted consent decisions; C2PA marking; signed AI-use attestation with the reviewer's sign-off (receipt coverage: partial). Hosted QA 2026-09-25: pass. Self-host QA 2026-09-25: pass.
- Hardware: Any CPU for the video steps (ffmpeg, Tesseract, C2PA); Qwen3.8-27B (1× RTX 5090 32 GB or larger) for the script checks
- Output: The ad with a visible AI label and a C2PA marking (when needed), a report and flag list (Markdown, CSV), and a signed AI-use attestation naming the reviewer.
- Page: https://decosa.ai/tools/media/disclosure-preflight · JSON: https://decosa.ai/use-cases/disclosure-preflight.json
- Prompts: https://decosa.ai/prompts/disclosure-preflight-hosted.md · https://decosa.ai/prompts/disclosure-preflight-selfhost.md · https://decosa.ai/prompts/disclosure-preflight-assemble.md
- Tiers:
  - Lite · code only, any CPU (Any CPU): Provenance, the label check and burn-in, the C2PA marking, consent lookups and the attestation, with the script read by word lists only ("use_model": false): profanity and music-cue markup, no names, brands or rating triggers.
  - Standard · adds script checks by Qwen3.8-27B (hosted demo) (Any CPU plus 1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate) for the model): Everything in Lite, plus the model's candidate flags and a typed yes/no for each, with a receipt per call, and the watermark decoder for stripped Decosa renders.
- Buyer facts:
  - Data retention: The upload and intermediate files are deleted when the run ends; the disclosed video and the report are kept for one hour, for the token or key that made the run. Logs carry counts only, never script text, names or the reviewer.
  - What leaves the box: Self-hosted: nothing (the script checks go to your own model server). Hosted demo: the video is checked on Decosa's server; only the script goes to the model through our gateway.
  - How 'synthetic' is decided: From the file's C2PA credential, Decosa's render receipt (by file hash, credential or invisible watermark) and your declaration. No detector model; no provenance and no declaration means 'needs a declaration'.
  - Consent: One receipted consent-ledger check per enrolled performer and kind (face, voice) for the project, purpose 'advertising', territory and date. Nobody is identified from their face or voice.
  - Cost per ad: A fraction of a cent in model time at the gateway's list price for the planted demo script (a few calls); the video steps are CPU time.
  - Output: The ad with a visible label and a C2PA marking when a rule needs them; a report and flag list (Markdown, CSV); and a signed AI-use attestation naming the reviewer, verifiable at /record/verify.
- Samples:
  - 1. Ad lumo loop (`ad-lumo-loop`): https://decosa.ai/tools/media/disclosure-preflight?sample=1&autorun=0
  - 2. Ad fernway h3 (`ad-fernway-h3`): https://decosa.ai/tools/media/disclosure-preflight?sample=2&autorun=0
  - 3. Raw undeclared (`raw-undeclared`): https://decosa.ai/tools/media/disclosure-preflight?sample=3&autorun=0
  - 4. Raw declared (`raw-declared`): https://decosa.ai/tools/media/disclosure-preflight?sample=4&autorun=0
  - 5. Stripped credential (`stripped-credential`): https://decosa.ai/tools/media/disclosure-preflight?sample=5&autorun=0
  - 6. Script planted (`script-planted`): https://decosa.ai/tools/media/disclosure-preflight?sample=6&autorun=0
  - 7. Replica wrong purpose (`replica-wrong-purpose`): https://decosa.ai/tools/media/disclosure-preflight?sample=7&autorun=0
- Regulatory note: Advisory, not legal advice; checked against the primary texts on 25 Sep 2026. New York General Business Law § 396-b as amended by S.8420-A / A.8887-B (signed 11 Dec 2025, chapter 617; in force on the 180th day, 9 Jun 2026): whoever produces or creates an advertisement must conspicuously disclose that it includes a synthetic performer, where they have actual knowledge of it. A synthetic performer is a digitally created asset, made or changed with generative AI or a software algorithm, meant to give the impression of a human performer who is not recognisable as any identifiable natural performer; a replica of a real person is not one (that is consent and right-of-publicity territory). Audio-only ads, AI used only to translate a human performer, and ads for expressive works (films, shows, games) consistent with the work are exempt; $1,000 for a first violation and $5,000 for each after. Whether a given label is conspicuous is a legal judgement: this tool reports that OCR read it on every sampled frame, nothing more. EU AI Act, Regulation (EU) 2024/1689, Art. 50(2): providers of systems that generate synthetic audio, image or video must mark the outputs in a machine-readable format, detectable as AI-generated; Art. 50(4) with Art. 3(60): deployers must disclose a deep fake (content resembling existing persons that would falsely appear authentic), limited for evidently artistic, creative, satirical or fictional work to a disclosure that does not hamper it. Art. 50 applies from 2 Aug 2026 (Art. 113); Regulation (EU) 2026/1744 (the Digital Omnibus on AI, OJ 24 Jul 2026) gives generators already on the market before that date until 2 Dec 2026 to comply with Art. 50(2). A C2PA credential is one machine-readable marking; the Act does not name a standard. SAG-AFTRA's 2023 TV/Theatrical agreement also requires notice to the union before a synthetic performer is used (from a secondary summary, not verified against the contract text); it binds signatory producers, not this tool. Consent is checked only for identities enrolled in the consent ledger; nobody is identified from their face or voice. Model licences: Apache-2.0 (Qwen3.8-27B).

### Turn a script into an animatic (`animatic-studio`)

Paste a short script or ad brief and get a timed animatic. Qwen3.8-27B breaks it into shots that each cite their script lines; code checks that every line is covered and inserts the dialogue word for word, and the grounding check flags shots that add events the script does not have. Frames come from Wan2.2-VACE-Fun-A14B, optional motion from the same model, temp dialogue from stock voices the consent ledger allows. Edit any shot and re-render just that one. The MP4 carries shot numbers, timecode, a C2PA credential and a signed record of which steps were AI and which were your edits. Previsualisation that saves a storyboard artist's first pass, not a replacement for one.

- What it is: Paste a short script or ad brief, get a timed animatic with a checked shot list, consented temp voices and a signed record of what was AI and what was you.
- Who it's for: Teams in film, tv and games and creative and media.
- Page: https://decosa.ai/apps/animatic-studio
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/animatic-studio):
  - 100% in 10 of 10 runs: Shot-list coverage of lines needing a shot, before any code repair (synthetic, n = 10)
  - 0: Invented dialogue (synthetic, n = 10)
  - 128 / 17 / 3 / 3: Grounding verdicts on shot actions: supported / partial / unsupported / contradicted (synthetic, n = 151)
  - 531.0 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B · Wan2.2-VACE-Fun-A14B · Kokoro-82M. Where: Hosted for sample and original scripts; self-host keeps unreleased scripts in-house.
- Verification: Receipt per model call (shot list and grounding); render receipt and C2PA credential on the MP4; signed consent decisions; a signed record of AI steps and human edits (receipt coverage: partial). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB): Qwen3.8-27B for the shot list, and about 34 GB for Wan2.2-VACE-Fun-A14B (fp8) in ComfyUI; the voices run on CPU
- Output: An MP4 animatic with shot numbers, timecode and an AI label (C2PA credential embedded), a shot-list CSV, a storyboard PDF and a signed record of AI steps and human edits.
- Page: https://decosa.ai/apps/animatic-studio · JSON: https://decosa.ai/use-cases/animatic-studio.json
- Prompts: https://decosa.ai/prompts/animatic-studio-hosted.md · https://decosa.ai/prompts/animatic-studio-selfhost.md · https://decosa.ai/prompts/animatic-studio-assemble.md
- Tiers:
  - Lite · breakdown only, no GPU renderer (The text model (or a remote endpoint); no render GPU): The checked shot list, character bible, grounding verdicts and CSV, without frames or voices.
  - Standard · the hosted demo (1x RTX PRO 6000 96 GB for the text model plus about 34 GB of GPU for the renderer while a job runs): Shot list and grounding on Qwen3.8-27B, frames and motion on Wan2.2-VACE-Fun-A14B in the shared ComfyUI, Kokoro voices, C2PA and the signed record.
  - Best · self-host with LTX-2.3 or MiniMax H3 (licence pending) motion (not measured): Standard plus LTX-2.3 or MiniMax H3 (licence pending) for motion shots on your own GPU, if the licence suits you. Not wired in or measured here.
- Alternates (smaller models for the same job):
  - Reference-based consistency (1x 96 GB card (a 20B edit model; not measured)), not built: An edit model conditioned on each character sheet (Qwen-Image-Edit, Apache-2.0) to hold faces and costumes across shots. Not installed here.
- Buyer facts:
  - What it replaces: The first pass: breakdown, rough frames and timing. A storyboard artist still stages, draws and cuts the real thing.
  - Grounded shot list: Every shot cites its script lines. In 10 of 10 eval runs the model's own list covered every line; dialogue is inserted word for word by code, and 10 of 10 planted invented events were flagged.
  - Consent: Every temp voice is checked in the consent ledger before the render and the decision is signed; a refused voice is subtitled instead. Stock voices only, nothing cloned.
  - AI and human record: A signed record lists the model's shot list, each of your edits by shot and field, the consent checks, frames, voices and the cut, with hashes. No script text in it.
  - Cost per animatic: A fraction of a cent of model time for the shot list and grounding (measured); rendering takes GPU-minutes on the shared card (measured), not billed in the demo.
  - Data retention: Hosted: the script, shot list, frames, voices and MP4 are kept 7 days so you can edit and re-render, then deleted. Frames and the MP4 open only through signed links that expire within hours and are given to the session or key that made the run. The signed record is public by design (anyone with its link can verify it): it holds hashes only and does not lead to your frames. Logs hold ids and counts, never script text.
  - What leaves the box: Hosted: the script goes to Decosa's API and its model server through our gateway. Self-host: nothing.
- Samples:
  - 1. Last lamp (`last-lamp`): https://decosa.ai/apps/animatic-studio?sample=1&autorun=0
  - 2. Night shift ad (`night-shift-ad`): https://decosa.ai/apps/animatic-studio?sample=2&autorun=0
  - 3. Salvage cutscene (`salvage-cutscene`): https://decosa.ai/apps/animatic-studio?sample=3&autorun=0
  - 4. Earnest handbag (`earnest-handbag`): https://decosa.ai/apps/animatic-studio?sample=4&autorun=0
- Regulatory note: Not legal advice. The cut carries a burned-in "AI-generated animatic (previs)" label and a C2PA credential, and every voice is checked in the consent ledger (tool 47) first. Laws below were read on 25 Sep 2026 by the disclosure pre-flight tool (49) and not re-read for this page: EU AI Act, Regulation (EU) 2024/1689 (https://eur-lex.europa.eu/eli/reg/2024/1689/oj), Art. 50(2) machine-readable marking of synthetic video applies from 2 Aug 2026 (later for systems already on the market) and Art. 50(4) asks deployers to disclose deepfakes; New York General Business Law 396-b (https://www.nysenate.gov/legislation/bills/2025/S8420/amendment/A), in force 9 Jun 2026, requires a conspicuous disclosure when an advertisement includes a synthetic performer, which matters only if an animatic is itself run as an ad. Copyright: the US Copyright Office's report on copyrightability (29 Jan 2025, https://www.copyright.gov/ai/) says AI output is protectable only where a person determined enough of the expression; the record shows which steps were yours. Script rights are yours to have; the samples are original (CC0) or public domain.

### Produce an audio drama (`audio-drama-studio`)

Paste a radio script or a prose chapter and review the parse: every speaker, line, direction and sound cue, before anything is voiced. Cast each role from consented house voices; every line passes the consent ledger right before it is spoken. Music comes from cleared cues with licence certificates, and sound from CC0 or public-domain recordings or code. The mix is mastered to a podcast or ACX-style audiobook spec, with chapters, captions, show notes with an AI disclosure, a C2PA credential, a signed record and sides for human actors.

- What it is: Your script, produced tonight, with every voice consented and every track cleared.
- Who it's for: Teams in film, tv and games and creative and media.
- Page: https://decosa.ai/apps/audio-drama-studio
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/audio-drama-studio):
  - 100% (389/389): Radio scripts: speaker right (code alone) (test split, n = 389)
  - 99.0% (96/97): Sound cue mapped to the right library tag (model) (test split, n = 97)
  - 87.5% (7/8): Cue with nothing in the library left as "no sound" (test split, n = 8)
  - 28.0 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B (parse) · Kokoro-82M (voices, CPU) · ACE-Step 1.5 (music, via music-gen-cleared). Where: Hosted with public-domain and original scripts; self-host for unpublished work.
- Verification: Receipt per model call; a signed consent decision per spoken line; C2PA credential linking every role's consent entry; signed production record (receipt coverage: partial). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: Qwen3.8-27B (1× RTX 5090 32 GB or larger) for the parse; the voices, sound and mix run on CPU; ACE-Step only if you compose new music (a GPU with about 10 GB free)
- Output: An MP3 episode with chapters (or ACX-style chapter files with credits), captions, a Podcasting 2.0 transcript and chapters, show notes, a sides pack for human actors, a C2PA credential and a signed record.
- Page: https://decosa.ai/apps/audio-drama-studio · JSON: https://decosa.ai/use-cases/audio-drama-studio.json
- Prompts: https://decosa.ai/prompts/audio-drama-studio-hosted.md · https://decosa.ai/prompts/audio-drama-studio-selfhost.md · https://decosa.ai/prompts/audio-drama-studio-assemble.md
- Tiers:
  - Lite · CPU only, radio scripts (Any 8-core CPU): No language model: speaker lines and cues are read by code, cues mapped by keywords; prose chapters are not supported.
  - Standard · the hosted demo (1x 96 GB card (or 32 GB for Qwen alone) plus an 8-core CPU): Qwen3.8-27B through our gateway for the parse, Kokoro on CPU, library music, full mix and credentials.
  - Best · compose new music per episode (Standard plus about 15 GB free on a GPU for ACE-Step): Standard plus a theme and a sting composed for the episode from your brief, through the music-gen-cleared guard, similarity check and certificate.
- Alternates (smaller models for the same job):
  - Voices that can act (About 4 GB of GPU beside the standard tier), not served: An MIT voice model with emotion control (Chatterbox, exaggeration setting) on stock reference voices.
- Buyer facts:
  - Consent: Every role and every line is checked against the consent ledger before it is spoken, and each decision is signed; a revoked or out-of-scope voice stops the render. Only stock voices with entries can be cast; nothing is cloned.
  - What you check: The parse is shown before anything is voiced: speakers, lines, sound and music cues, all editable. Spoken words are always the script's own.
  - Loudness: Measured on the delivered file: podcast -16 LUFS +/- 1 and true peak at most -1 dBTP, or ACX RMS, peak, noise floor and room tone per chapter file. All sample episodes passed.
  - Licences: Voices Apache-2.0 (Kokoro-82M), music MIT (ACE-Step) with a signed certificate per cue, sounds CC0 or public domain with source links, or generated by code. The show notes list them.
  - Cost per episode: A fraction of a cent of model time for the parse (one receipted call for a script); voices and mixing run on CPU, a few seconds per finished minute (measured).
  - Data retention: Episode files are deleted after 7 days; until then they open only through signed links that expire within hours, given to the token or key that made the episode. The server keeps ids, hashes, settings and measurements, and the signed record; never script text in logs.
  - What leaves the box: Hosted: the script, sent to Decosa's API; its parse calls go through the Decosa API. Self-host: nothing.
  - Not for ACX: ACX prohibits unauthorised AI narration (requirements dated 15 Apr 2026). The audiobook mode meets its technical spec only; use it for other platforms, classrooms and pitches.
- Samples:
  - 1. Scandal in bohemia (`scandal-in-bohemia`): https://decosa.ai/apps/audio-drama-studio?sample=1&autorun=0
  - 2. Cask of amontillado (`cask-of-amontillado`): https://decosa.ai/apps/audio-drama-studio?sample=2&autorun=0
  - 3. Night shift (`night-shift`): https://decosa.ai/apps/audio-drama-studio?sample=3&autorun=0
  - 4. Town mouse (`town-mouse`): https://decosa.ai/apps/audio-drama-studio?sample=4&autorun=0
- Regulatory note: Not legal advice. Checked 26 Sep 2026: ACX's Audio Submission Requirements (help.acx.com, page dated 15 Apr 2026) prohibit unauthorised text-to-speech or AI narration, so this studio's audiobook mode meets ACX's technical spec (RMS -23 to -18 dB, peak -3 dB, noise floor -60 dB, 1-5 s room tone, credits files) but does not make a title eligible for ACX. Apple Podcasts recommends -16 LKFS +/- 1 dB and true peak at most -1 dBFS (podcasters.apple.com/support/893-audio-requirements). EU AI Act Art. 50(2) (Regulation (EU) 2024/1689; applies from 2 Aug 2026) requires machine-readable marking of synthetic audio: the C2PA credential covers it, signed with a development certificate, so public validators show the issuer as untrusted. Voice-replica laws (Tennessee ELVIS Act, California AB 2602, New York S7676B, as summarised by the consent ledger, tool 47) are why every line goes through the ledger; this build uses stock synthetic voices only and does not clone. Platform AI-disclosure rules for podcasts (Apple, Spotify) were not verified.

### Make a music video for your track (`music-video-studio`)

A first music video for an independent artist's own track. It finds the beats, bars and sections, times the artist's lyrics to the audio, runs the sample-clearance pre-check, and has an open model write scenes that each cite the lines they illustrate; code and a receipted yes/no check test every citation. Clips render with an open video model, cut on the bar lines, with the lyrics as karaoke captions, in 16:9 and 9:16. Each file carries a C2PA credential and a burned-in AI label, and a signed record holds the rights statement, the checks and the measured cut timing. The visuals are 480p upscaled: a stylised first video, not a studio shoot.

- What it is: A first music video from an artist's own track: cuts on the beat, the lyrics as timed captions, scenes written from the lyrics, labelled and credentialed.
- Who it's for: Teams in music and film, tv and games.
- Page: https://decosa.ai/apps/music-video-studio
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/music-video-studio):
  - 81.8% / 87.7%: Lyric lines within 0.3 s / 1 s of human timing (test split, n = 6)
  - 0.868: Beat F-measure (±70 ms), artist's BPM given (test split, n = 40)
  - 67.3%: Typed check says no to a swapped (mismatched) scene (test split, n = 52)
  - 15.1 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B writes and checks the treatment; Wan2.2-VACE-Fun-A14B renders; wav2vec2 aligns the lyrics; librosa finds the beat. Where: Hosted (the artist's own or openly licensed tracks) or self-host.
- Verification: Receipt per model call; render receipt and C2PA credential per file; signed record with the rights statement and checks (receipt coverage: partial). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) shared with the text model, or a 48 GB card for the video model plus a remote text model; the analysis runs on CPU
- Output: Two MP4s (16:9 and 9:16) with captions, an AI label and a C2PA credential, SRT and LRC captions, and a signed record of the rights statement, checks and cut timing.
- Page: https://decosa.ai/apps/music-video-studio · JSON: https://decosa.ai/use-cases/music-video-studio.json
- Prompts: https://decosa.ai/prompts/music-video-studio-hosted.md · https://decosa.ai/prompts/music-video-studio-selfhost.md · https://decosa.ai/prompts/music-video-studio-assemble.md
- Tiers:
  - Lite · timed captions and a cited treatment, no GPU for video (CPU (8+ cores) plus the text model (about 20 GB of GPU, or a remote server)): Beat grid, lyric timing (SRT, LRC, ASS), the clearance pre-check and a checked treatment with an edit plan. Bring your own footage.
  - Standard · the hosted demo, clips on one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB (shared), CPU for the analyzer and the edit): Everything in Lite plus a clip per scene on Wan2.2-VACE (Apache-2.0), cut on the bar lines in 16:9 and 9:16, labelled and credentialed.
  - Best · MiniMax H3 clips (licence pending), self-host (1x RTX PRO 6000 96 GB plus about 115 GB of RAM for offload): Everything in Standard with clips from MiniMax-H3 instead of Wan2.2. Self-host available where H3's current licence covers you; the pending licence would bring it to the hosted demo. Not wired in yet.
  - Wanted · MiniMax H3 on two cards, no offload (2x RTX PRO 6000 96 GB, no CPU offload (about 133 GB of BF16 weights; estimate)): A clip per scene adds up over a song. Both halves of MiniMax-H3 in BF16 on two 96 GB cards, so nothing is offloaded to RAM: faster renders (estimate, not measured). Self-host now where H3's current licence covers you; once the pending licence lands we add this compute for the hosted tier ourselves. Not a community-provider model.
- Alternates (smaller models for the same job):
  - Sharper clips with LTX-2.3 (1x 96 GB card), self-host only: LTX-2.3 is on the box but not wired in, and self-host only: its community licence has a competing-service clause, so we can't serve it hosted.
- Buyer facts:
  - Data retention: The track is kept 24 hours after upload so it can be rendered (longer only while a render is queued or running), then deleted. The run (hashes, analysis, lyric timings, captions, plan, record) is kept 7 days; the rendered videos stay in the studio's media folder and open only through signed links that expire within hours, given to the run's owner or its watch link. Logs hold ids, counts and hashes, never lyrics.
  - What leaves the box: On the hosted route, the lyrics and plan go to the text model through our gateway and the render runs on Decosa's hosted service. Self-hosted, nothing leaves your machine.
  - Rights: You confirm you own the track or hold the rights; the statement is signed into the record, not verified. The clearance pre-check covers only a small open catalogue.
  - Cost per video: A fraction of a cent of model calls per analysis (measured). Rendering a video in one format takes several GPU-minutes, tens of cents at a GPU rental list price; both formats about twice that. Hosted demo renders are not billed.
  - Typical time: Seconds for the analysis and treatment of an excerpt; many minutes to render it in one format and about twice that in both, on a shared GPU (measured).
  - Outputs: MP4s in 16:9 (1280x720) and 9:16 (720x1280) at 30 fps with captions, an AI label on every frame and a C2PA credential; SRT, LRC and ASS captions; a signed record with the rights statement, checks, receipts and measured cut timing.
- Samples:
  - 1. Rxbyn bad side (`rxbyn-bad-side`): https://decosa.ai/apps/music-video-studio?sample=1&autorun=0
  - 2. Cortez feel (`cortez-feel`): https://decosa.ai/apps/music-video-studio?sample=2&autorun=0
  - 3. Lanterns music3 (`lanterns-music3`): https://decosa.ai/apps/music-video-studio?sample=3&autorun=0
- Regulatory note: Not legal advice; checked on 26 Sep 2026 unless stated. Rights: the artist confirms they own the track or have the rights to make and publish a video with it; the statement goes into the signed record and is not verified. The sample-clearance pre-check compares the upload with a small open catalogue only, and a clean result says nothing about commercial music (US courts disagree on short samples: Bridgeport v. Dimension Films, 6th Cir. 2005, against VMG Salsoul v. Ciccone, 9th Cir. 2016). Demo tracks are CC BY excerpts from JamendoLyrics, credited on screen, and one song made with MiniMax-Music3 through the music-gen-cleared tool. Disclosure: every file has a visible "AI-generated visuals / Made with AI" label on every frame and a C2PA credential that declares composite AI media. EU AI Act (Regulation (EU) 2024/1689) Art. 50(2) asks providers to mark generated video in a machine-readable way (from 2 Aug 2026; checked by the disclosure pre-flight on 25 Sep 2026; EUR-Lex link under Tools). New York General Business Law § 396-b requires a conspicuous disclosure of synthetic performers in advertisements (in force 9 Jun 2026; bill text linked under Tools): it applies if the video is used as an ad, and the console runs that rule on the file. YouTube asks creators to disclose realistic altered or synthetic content when they upload and labels it (help page read 26 Sep 2026, linked under Tools); TikTok and Meta have their own AI labels (not checked here): use each platform's toggle as well. Copyright: the US Copyright Office's Part 2 report (29 Jan 2025) says AI output is protected only where a human determined sufficient expressive elements; the record claims no copyright in the generated visuals. Scenes never depict real people, brands or readable text by prompt rule and screen.

### Write a SAR narrative (`sar-narrative-desk`)

Drafts a suspicious-activity narrative from an alert's case file (a transaction ledger, the customer profile and the investigator's notes) in the structure FinCEN asks for: who, what, when, where, why and how. Every sentence cites the rows, fields or notes it rests on. Every amount, date, count and total is recomputed in code from the cited rows, and a wrong one is held back with the right figure beside it. It also writes a signed no-file rationale for alerts closed without a SAR, or checks a draft you wrote. A drafting aid: an investigator decides and files.

- What it is: A suspicious-activity narrative drafted from the case file, with every sentence citing its rows and every number recomputed in code.
- Who it's for: Teams in finance and insurance and compliance and trust.
- Page: https://decosa.ai/tools/finance/sar-narrative-desk
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/sar-narrative-desk):
  - 1,598 of 1,600 (99.9%): Planted wrong numbers caught in reference narratives, first run (no model) (test split, n = 1600)
  - 340 of 356 (95.5%): Wrong numbers planted in the model's own sentences caught (v3) (test split, n = 356)
  - 18 of 20: Invented sentences held in investigator drafts (test split, n = 20)
  - 24.7 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B drafts and judges; the numbers and the red-flag screen are plain code. Where: Self-host for real cases (SAR confidentiality); the hosted demo takes synthetic case files only.
- Verification: Receipt per model call; every number recomputed from the cited rows; signed report over hashes and verdicts (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for Qwen3.8-27B; the numeric check, the screen and the report run on CPU
- Output: A filing copy of the narrative (no citations, held sentences left out), a cited workpaper in Markdown for the case file, and a signed JSON report.
- Page: https://decosa.ai/tools/finance/sar-narrative-desk · JSON: https://decosa.ai/use-cases/sar-narrative-desk.json
- Prompts: https://decosa.ai/prompts/sar-narrative-desk-hosted.md · https://decosa.ai/prompts/sar-narrative-desk-selfhost.md · https://decosa.ai/prompts/sar-narrative-desk-assemble.md · https://decosa.ai/prompts/sar-narrative-desk-mac.md
- Tiers:
  - Lite · numbers and citations only, no GPU (Any CPU): POST /sar/numbers checks a narrative you wrote with any tool: every citation must exist and every number must match or be computed from the cited rows. No drafting, no judge for the other facts.
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B drafts each section and judges each sentence against what it cites; the screen and the numbers are code. This is what the hosted demo runs, on synthetic case files only.
- Alternates (smaller models for the same job):
  - Statements as PDFs and scans (1x RTX PRO 6000 96 GB (estimate)), not built: Read the ledger from bank statements and wire confirmations instead of CSV, with each row tied to its page. The OCR model is not chosen yet; it must be licence-clean.
- Buyer facts:
  - Data retention: Nothing stored: the case file lives in memory for the request. The signed report holds hashes, citations, verdicts and receipt ids, never case text; logs carry counts only.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and only synthetic case files are accepted. Self-hosted on the direct route: nothing leaves the box.
  - What it will not do: Decide whether to file, or file. It never says 'compliant'. A clean result reads 'every sentence traced to the case file (review still required)'.
  - Input formats: Transactions as CSV (date and amount required; type, direction, channel, location, counterparty, memo optional), up to 400 rows; profile and alert fields as name: value pairs; notes up to 6,000 characters; your own draft up to 12,000 characters with [T3, K.name] citations.
  - Typical run: A structuring draft: a few dozen model calls, a few cents or less at the gateway list price. Checking a short draft: a few calls, a fraction of a cent. Each run shows its own measured cost.
- Samples:
  - 1. Structuring (`structuring`): https://decosa.ai/tools/finance/sar-narrative-desk?sample=1&autorun=0
  - 2. Rapid movement (`rapid-movement`): https://decosa.ai/tools/finance/sar-narrative-desk?sample=2&autorun=0
  - 3. Funnel (`funnel`): https://decosa.ai/tools/finance/sar-narrative-desk?sample=3&autorun=0
  - 4. Elder exploitation (`elder-exploitation`): https://decosa.ai/tools/finance/sar-narrative-desk?sample=4&autorun=0
  - 5. Clean (`clean`): https://decosa.ai/tools/finance/sar-narrative-desk?sample=5&autorun=0
  - 6. Planted errors (`planted-errors`): https://decosa.ai/tools/finance/sar-narrative-desk?sample=6&autorun=0
- Regulatory note: Checked 26 Sep 2026 against primary sources (links under Tools). SAR information is confidential: 31 U.S.C. 5318(g)(2) bars notifying anyone involved that a transaction was reported, and 31 CFR 1020.320(e) makes a SAR, and anything that would reveal one, confidential. So this is self-host first, and the hosted demo refuses any case file not marked synthetic. The draft follows FinCEN's Guidance on Preparing a Complete & Sufficient SAR Narrative (November 2003: the five W's plus how; introduction, body, conclusion; chronological; individual dates and amounts, not only totals; no tables or 'see attached') and the FinCEN SAR Filing Instructions (version 1.0, April 2026, Part V: do not file supporting documentation with the SAR), which is why the filing copy drops the citations and the cited workpaper stays in the case file (31 CFR 1020.320(d): keep supporting documentation five years). A bank files within 30 calendar days of initial detection, up to 60 if no suspect is identified (31 CFR 1020.320(b)(3)). The no-file rationale is optional: FinCEN's SAR FAQs of 9 October 2025 (Q4) say there is no requirement to document a decision not to file, FinCEN encourages it, and a short statement will usually suffice. Typologies follow FinCEN Advisories FIN-2014-A005 (funnel accounts, 28 May 2014) and FIN-2022-A002 (elder financial exploitation, 15 June 2022: an older adult is 60 or over); structuring follows 31 CFR 1010.100(xx) and FAQ Q1 (amounts near $10,000 alone do not require a SAR). The rapid-movement red flag cites the FFIEC BSA/AML manual appendix, which we could not re-fetch (unverified). The screen's thresholds are our own, not regulatory. This is not legal advice and never a compliance determination: the investigator decides whether to file, edits and files. Model licence: Apache-2.0 (Qwen3.8-27B).

### Review a claim file for conduct (`claims-conduct-pack`)

Reads a claim file and its policy the way a market-conduct examiner would, under the NAIC model, California (10 CCR 2695) or Texas (Ins. Code 542). It computes the acknowledgement, decision, status-letter and payment deadlines from the file's own dates, checks each denial reason for a cited provision and grounds it in the policy text, and looks for the Department of Insurance review notice. It gives typed judgments with quotes on misrepresentation, low offers and missing investigation, records any AI use, and seals it all in a signed file review a reviewer can sign off.

- What it is: A claims file review for unfair claims practices rules: it reads a claim file and the policy the way a market-conduct examiner does, works out every deadline from the file's own dates, checks each denial reason against the policy, and seals the result in a signed file review for a person to confirm.
- Who it's for: Teams in finance and insurance and compliance and trust.
- Page: https://decosa.ai/tools/finance/claims-conduct-pack
- Last reviewed: 2026-09-26
- Key result: On 48 synthetic test files, run once after the prompts were frozen, it found 52 of 54 planted problems with one false flag, and all 26 in the files with held-out phrasing. The files are templated and written by the same author as the prompts, so the numbers do not predict accuracy on a carrier's files.
- Key numbers (details and caveats: https://decosa.ai/metrics/claims-conduct-pack):
  - 26 / 0 / 0: Held-out phrasing files: found / false / missed (held out, n = 24)
  - 52 of 54 (precision 0.98, recall 0.96): Planted problems found, all checks (test split, n = 54)
  - 1: False flags (test split, n = 48)
  - 10.1 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real claim files (hosted demo: synthetic files only).
- Verification: Receipt per model call; every quote and date checked against the file; signed hash-chained file review with each step marked rule, model or human (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the deadline rules and the record need no GPU
- Output: A conduct checklist (flag, review, ok, n/a) with rules, dates and quotes, a Markdown file review, and a signed file review that a reviewer's sign-off points at.
- Page: https://decosa.ai/tools/finance/claims-conduct-pack · JSON: https://decosa.ai/use-cases/claims-conduct-pack.json
- Prompts: https://decosa.ai/prompts/claims-conduct-pack-hosted.md · https://decosa.ai/prompts/claims-conduct-pack-selfhost.md · https://decosa.ai/prompts/claims-conduct-pack-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B reads the file, grounds the denial and answers the conduct questions; the deadlines are plain code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, multi-claimant files.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Claim files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: Acknowledgement, investigation, decision, status-letter, payment and reply deadlines from the file's own dates; denial reasons (a cited provision that exists and says what the letter says); the Department of Insurance review notice; misrepresentation, low offers and missing investigation; AI use. Rulepacks: NAIC model, California, Texas.
  - What it does not do: It never says a file is compliant and does not decide coverage. Not checked: total-loss valuation, subrogation, statute-of-limitations notices, fraud-investigation extensions, surplus lines, 'final payment' wording, other states. Scanned letters need OCR first.
  - Data retention: Nothing kept on the server. Files, policies and records live in memory for the request; logs carry counts only. You keep the signed file review with the claim file.
  - What leaves the box (hosted demo): The claim file and policy go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for synthetic files only.
  - Model calls per file: One extraction, one grounding call per denial reason, and up to four typed judgments: 2 to 6 calls.
  - Typical run cost: A fraction of a cent at the gateway list price for the planted California file (a handful of model calls). Each run shows its own measured cost.
  - Who did what: The signed record marks each step as rule, model or human; a reviewer's confirm, dismiss or escalate decisions become a second signed record pointing at the first.
- Samples:
  - 1. Ca auto planted (`ca-auto-planted`): https://decosa.ai/tools/finance/claims-conduct-pack?sample=1&autorun=0
  - 2. Ca home clean (`ca-home-clean`): https://decosa.ai/tools/finance/claims-conduct-pack?sample=2&autorun=0
  - 3. Tx home lowball (`tx-home-lowball`): https://decosa.ai/tools/finance/claims-conduct-pack?sample=3&autorun=0
  - 4. Tx auto clean (`tx-auto-clean`): https://decosa.ai/tools/finance/claims-conduct-pack?sample=4&autorun=0
  - 5. Naic auto noclause (`naic-auto-noclause`): https://decosa.ai/tools/finance/claims-conduct-pack?sample=5&autorun=0
  - 6. Ca home notices (`ca-home-notices`): https://decosa.ai/tools/finance/claims-conduct-pack?sample=6&autorun=0
- Regulatory note: Rules read from primary texts on 26 Sep 2026. NAIC Unfair Claims Settlement Practices Act (Model #900, adopted 1990, amended 1991) and Unfair Property/Casualty Claims Settlement Practices Model Regulation (Model #902, 1990, amended 1991): model laws that bind nobody until a state enacts them; #902 asks for acknowledgement within 15 days, a decision within 21 days of proof of loss or a written reason for more time, then a letter every 45 days, payment within 30 days, and a denial that references the provision it relies on. California Fair Claims Settlement Practices Regulations, 10 CCR 2695.5 and 2695.7 (operative 1993; 2695.7 last amended 14 Jul 2021): acknowledge, send forms and begin investigating within 15 calendar days, accept or deny within 40 calendar days of proof of claim or give written notice of the need for more time every 30 days, pay within 30 days, list every basis and the provision in a written denial, and tell the claimant they may have the matter reviewed by the California Department of Insurance, with its address and phone number. Texas Insurance Code 542.055 to 542.058 (eff. 1 Apr 2005): acknowledge within 15 days, accept or reject within 15 business days of receiving all items (or give reasons and decide within 45 days), state the reasons for a rejection, and pay within 5 business days. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (adopted 4 Dec 2023) covers claim administration and says the Department may ask for documentation of AI systems used in decisions; it applies in a state once that state issues it. Business days skip weekends and US federal holidays, an assumption. Not legal advice.

### Review a collections call (`collections-call-qa`)

QA for recorded debt-collection and mortgage-servicing calls under Regulation F and Regulation X. On the call it answers yes, no or unclear, with a quote and a time: debt-collector disclosure, identity before the debt, third-party disclosure, threats, a stop request or dispute brushed aside, and loss-mitigation options. From the call log, in code: the 7-in-7 call count, calling hours in every time zone on file, calls after a stop request, and live contact by day 36. Signed record kept with the recording for three years.

- What it is: A recorded collection or servicing call and its call log, checked against Regulation F and X: each answer quoted with its time, the call counts computed in code.
- Who it's for: Teams in finance and insurance and compliance and trust.
- Page: https://decosa.ai/tools/finance/collections-call-qa
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/collections-call-qa):
  - 32/33: Per-rule accuracy, test B run 1 (written after test run 1) (held out, n = 33)
  - 81/83: Per-rule accuracy, test run 1 (test split, n = 83)
  - 11/12: Planted answers found, test run 1 (test split, n = 12)
  - 4.2 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: MOSS-Transcribe-Diarize · Qwen3.8-27B. Where: Self-host for real calls (hosted demo: synthetic role-plays only).
- Verification: Receipt per model call; every yes backed by a quote found in the transcript; call-log findings computed in code; signed hash-chained record with a retention date (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Qwen3.8-27B and the diarizer; the call-log maths and the record need no GPU
- Output: A QA checklist of typed answers with quotes and times, the call-log findings with the calls behind each, and a signed record with its retention date (the call date plus three years).
- Page: https://decosa.ai/tools/finance/collections-call-qa · JSON: https://decosa.ai/use-cases/collections-call-qa.json
- Prompts: https://decosa.ai/prompts/collections-call-qa-hosted.md · https://decosa.ai/prompts/collections-call-qa-selfhost.md · https://decosa.ai/prompts/collections-call-qa-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same checks on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B extracts and answers every typed check; MOSS-Transcribe-Diarize turns recordings into transcripts; the call-log maths is code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long calls and hard cases.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Call recordings stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: On each recorded call: the debt-collector disclosure, identity before the debt, third-party disclosure, false or threatening statements, a stop request or dispute brushed aside, and loss-mitigation options on a delinquent mortgage. From the call log: the 7-in-7 and 7-days-after-a-conversation presumptions, calling hours in every time zone on file, calls after a stop request or a lawyer, and live contact by day 36.
  - What it does not do: It gives no compliance verdict. It does not check validation notices, letters, emails or texts, credit reporting, time-barred-debt disclosures, Regulation X written notices or loss-mitigation application handling beyond what is said on the call. The log findings are presumptions from the log you give.
  - Consent: Every run needs a statement that the call was recorded with the consent the law requires (some states need every party's). The statement goes into the signed record, and a check looks for the recording notice on the call.
  - Data retention: Nothing on the server. Transcripts, call logs and records live in memory for the request, and logs carry counts and timings only. You keep the signed record with the recording for three years (12 CFR 1006.100(b)).
  - What leaves the box (hosted demo): The transcript goes to Qwen3.8-27B through our gateway, which is Decosa-operated. The call log is processed in code and not sent to the model. The gateway's receipts hold hashes, not text. Self-hosted, nothing leaves.
  - Model calls per call: One extraction plus 8 typed checks (10 for a mortgage). The call-log findings need no model: POST /collections/log returns them on their own.
  - Typical run cost: A fraction of a cent at the gateway list price for a short call (measured on the smoke sample). Each run shows its own measured cost.
- Samples:
  - 1. Tb1 showcase (`tb1-showcase`): https://decosa.ai/tools/finance/collections-call-qa?sample=1&autorun=0
  - 2. T5 early hours (`t5-early-hours`): https://decosa.ai/tools/finance/collections-call-qa?sample=2&autorun=0
  - 3. T6 threat dispute (`t6-threat-dispute`): https://decosa.ai/tools/finance/collections-call-qa?sample=3&autorun=0
  - 4. T7 stop refused (`t7-stop-refused`): https://decosa.ai/tools/finance/collections-call-qa?sample=4&autorun=0
  - 5. T8 coworker (`t8-coworker`): https://decosa.ai/tools/finance/collections-call-qa?sample=5&autorun=0
  - 6. T9 mortgage no options (`t9-mortgage-no-options`): https://decosa.ai/tools/finance/collections-call-qa?sample=6&autorun=0
  - 7. T1 clean callback (`t1-clean-callback`): https://decosa.ai/tools/finance/collections-call-qa?sample=7&autorun=0
  - 8. T2 clean mortgage collector (`t2-clean-mortgage-collector`): https://decosa.ai/tools/finance/collections-call-qa?sample=8&autorun=0
- Regulatory note: Checked 26 Sep 2026 against eCFR (text current to 24 Sep 2026). Regulation F, 12 CFR part 1006 (in force since 30 Nov 2021), binds debt collectors as defined in 1006.2(i): third-party collectors and debt buyers, and a mortgage servicer only for loans it obtained in default. Encoded: the debt-collector disclosure in the initial and later communications (1006.18(e)); no communication about the debt with third parties (1006.6(d)(1); location calls must not say a debt is owed, 1006.10(b)); no false, deceptive or misleading representations or threats (1006.18); no calls through a medium the person asked not to use, e.g. "stop calling" (1006.14(h), comment 14(h)(1)-3.i); the presumption of a violation for more than seven calls in seven consecutive days, or a call within seven days after a telephone conversation, per person and per debt, with the (b)(3) exclusions (1006.14(b)(2)); the 8 a.m.-9 p.m. local-time presumption, convenient in every location the collector's information points to (1006.6(b)(1)(i), comment 6(b)(1)(i)-2); no calls to a consumer known to be represented by an attorney (1006.6(b)(2)); and keeping each call recording for three years after the call (1006.100(b)). Oral disputes are recorded but the stop-collection duty in 1006.38(d)(2) applies only to written disputes in the validation period. Regulation X, 12 CFR 1024.39-1024.41 (as amended at 90 FR 20792, 16 May 2025, which removed the expired COVID-19 live-contact provisions): live contact with a delinquent borrower by day 36, and informing the borrower about loss-mitigation options if appropriate (1024.39(a)); accurate information from assigned personnel (1024.40(b)). They apply to a principal residence and not to small servicers (1024.30). Right-party verification is a practice check that supports 1006.6(d), not a rule of its own. Recording consent is state law (e.g. California Penal Code 632). QA triage, not legal advice.

### Draft breach notices and deadlines (`incident-notification-pack`)

For the CISO and general counsel during a cyber incident. From the team's incident log it builds a signed, hash-chained timeline and a deadline clock per report (SEC 8-K Item 1.05, NIS2, DORA, the Cyber Resilience Act, California), computed in code from the entries you tag. It drafts each notice or checks yours sentence by sentence against the log entries it cites: times and counts in code, other facts by a grounding judge. It lists the required elements each notice lacks and the facts two notices state differently. Counsel decides and files.

- What it is: A signed incident timeline, a deadline clock per regime, and every sentence of each notice checked against the log.
- Who it's for: CISOs, general counsel, breach counsel and DFIR firms preparing regulator notices during a cyber incident.
- Page: https://decosa.ai/tools/finance/incident-notification-pack
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/incident-notification-pack):
  - 20 / 20: Planted problems caught, held-out scenarios (test split, n = 20)
  - 35 / 44: Deadlines right, hand-labelled cases, first run (test split, n = 44)
  - 1 / 42: Clean sentences held, held-out scenarios (test split, n = 42)
  - 23.3 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B drafts, reviews and judges; the clocks, times and numbers are plain code. Where: Self-host for real incidents (privileged, possibly material non-public); the hosted demo takes synthetic incidents only.
- Verification: Receipt per model call; every time and number checked in code against the cited entries; hash-chained timeline; signed pack and a signed counsel sign-off (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for Qwen3.8-27B; the clocks, the time and number checks and the record run on CPU
- Output: A Markdown pack with the clocks, the cited notices and every check, a clean copy of each notice without held sentences, a signed JSON report, the sealed timeline record, and counsel's signed sign-off.
- Page: https://decosa.ai/tools/finance/incident-notification-pack · JSON: https://decosa.ai/use-cases/incident-notification-pack.json
- Prompts: https://decosa.ai/prompts/incident-notification-pack-hosted.md · https://decosa.ai/prompts/incident-notification-pack-selfhost.md · https://decosa.ai/prompts/incident-notification-pack-assemble.md
- Tiers:
  - Lite · clocks and timeline, no GPU (Any CPU): POST /incident/clock: the timeline, log issues and every regime's deadline from the tagged entries. No drafting and no sentence checks.
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B drafts notices, judges each sentence against what it cites, and maps required elements; clocks, times, numbers and contradictions are code. This is what the hosted demo runs, on synthetic incidents only.
- Buyer facts:
  - Data retention: Nothing stored: the log lives in memory for the request. The signed pack holds hashes, ids, verdicts, clocks, receipt ids and the facts read from each notice (counts, yes/no fields, matched times and numbers), never the log or notice text itself; the timeline record goes back to you; logs carry counts only.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and only incidents marked synthetic are accepted. Self-hosted on the direct route: nothing leaves the box.
  - What it will not do: Decide whether a regime applies or whether an incident is material, significant, major or severe; file anything; or call a notice compliant. A clean result reads 'every sentence traced to the log (counsel review still required)'.
  - Input formats: Log lines 'time | source | text | #tag' or JSON entries, up to 300 entries and 150,000 characters; every time needs a zone. Up to 6 notices per pack, each up to 60 sentences with [E3, P.name] citations, or left blank to be drafted.
  - Typical run: The ransomware sample: a dozen or so model calls and a fraction of a cent at the gateway list price; a drafted DORA notice adds a few calls. Each run shows its own measured cost.
- Samples:
  - 1. Ransomware saas (`ransomware-saas`): https://decosa.ai/tools/finance/incident-notification-pack?sample=1&autorun=0
  - 2. Cloud leak dora (`cloud-leak-dora`): https://decosa.ai/tools/finance/incident-notification-pack?sample=2&autorun=0
  - 3. Cra exploited vuln (`cra-exploited-vuln`): https://decosa.ai/tools/finance/incident-notification-pack?sample=3&autorun=0
  - 4. Clean pack (`clean-pack`): https://decosa.ai/tools/finance/incident-notification-pack?sample=4&autorun=0
- Regulatory note: Checked 26 Sep 2026 against primary sources (links under Tools). SEC Form 8-K Item 1.05: file within four business days after determining the incident is material (General Instruction B.1; Release 33-11216; compliance from 18 Dec 2023, smaller reporting companies 15 Jun 2024); an undetermined incident may go under Item 8.01 (SEC Corp Fin statement, 21 May 2024); we show 17:30 Eastern because later EDGAR submissions are dated the next business day (17 CFR 232.13). NIS2 Art. 23: early warning within 24 hours of becoming aware, notification within 72 hours, final report one month after the notification (national law applies; amendments proposed 20 Jan 2026, COM(2026) 13, not adopted). DORA Art. 19 and Delegated Regulation (EU) 2025/301 Art. 5: initial notification within 4 hours of classification as major and no later than 24 hours from awareness, intermediate within 72 hours, final within one month; weekend relief to noon of the next working day except for some entity types (Member State bank holidays are not modelled). Cyber Resilience Act Art. 14 (applies from 11 Sep 2026, Art. 71(2)): actively exploited vulnerabilities and severe incidents, early warning 24 hours, notification 72 hours, final report 14 days after a fix (vulnerabilities) or one month (incidents), through ENISA's single reporting platform. California Civ. Code 1798.82 as amended by SB 446 (from 1 Jan 2026): notice within 30 calendar days of discovery, required headings and contents, AG sample within 15 days if more than 500 residents. Unverified: national NIS2 rules, whether CRA format acts were adopted, whether Regulation 1182/71 extends EU day and month periods (shown as an extended date, weekends only). Whether a regime applies, and whether an incident is material, significant, major or severe, is counsel's call; the clocks only run from the entries the team tagged. Not legal advice and never a compliance determination. Model licence: Apache-2.0 (Qwen3.8-27B).

### Tie out MD&A figures (`filing-tieout`)

Ties every figure in a 10-K or 10-Q's MD&A to the filing's tagged statements and notes, in code: the cell it equals, or the change, percentage change or margin it is computed from. A figure that does not tie is flagged with the cell it was compared with: a different value, last year's figure, million for billion, the other direction, a note reference to the wrong note, or a note table that disagrees with a statement. Figures the tables do not show are left for a person. You get a workpaper and a signed record. A preparer's and reviewer's aid, not an audit.

- What it is: Every figure in a 10-K or 10-Q's MD&A tied to the tagged statements and notes, in code, with a signed workpaper.
- Who it's for: SEC reporting teams, controllers and audit staff who tie out 10-K and 10-Q drafts.
- Page: https://decosa.ai/tools/finance/filing-tieout
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/filing-tieout):
  - 11 in 6,039 figures (1.8 per 1,000): False flags on untouched held-out 10-K MD&As (frozen rules) (test split, n = 6039)
  - 79/79: Note table that disagrees with a statement, caught (test split, n = 79)
  - 88/142 (62%): Scale error (million/billion) caught (test split, n = 142)
  - 9.5 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: The tie-out is plain code; Qwen3.8-27B optionally reads claims in words. Where: Self-host for drafts (pre-release numbers are MNPI); the hosted demo ties public EDGAR filings.
- Verification: Every tie names a fact id; receipt per claim read; signed record over hashes, statuses and fact ids (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: Any CPU for the tie-out; 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for Qwen3.8-27B if you want claims in words read
- Output: A Markdown and CSV workpaper with every figure's cell or derivation, and a signed JSON record of hashes, statuses and fact ids.
- Page: https://decosa.ai/tools/finance/filing-tieout · JSON: https://decosa.ai/use-cases/filing-tieout.json
- Prompts: https://decosa.ai/prompts/filing-tieout-hosted.md · https://decosa.ai/prompts/filing-tieout-selfhost.md · https://decosa.ai/prompts/filing-tieout-assemble.md
- Tiers:
  - Lite · figures only, no GPU (Any CPU): Every figure, period, scale, direction and note reference tied in code; no model, no receipts. Runs on any CPU.
  - Standard · one GPU for claims in words (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): The lite tier plus Qwen3.8-27B reading up to 12 claims in words ('gross margin improved') against the named lines' figures.
- Alternates (smaller models for the same job):
  - MD&A tables and Word/PDF drafts (1x RTX PRO 6000 96 GB (estimate)), not built: Tie the tables inside MD&A and read drafts from Word or PDF, not only iXBRL and pasted text. The reader is not chosen yet; it must be licence-clean.
- Buyer facts:
  - Data retention: Nothing stored but a cache of public EDGAR documents you asked for. Drafts and tables live in memory for the request; the signed record holds hashes, statuses and fact ids, never text; logs carry counts only.
  - What leaves the box: Only an EDGAR request (CIK and accession number) when you ask for a public filing. Hosted: claim reads go through our gateway to the GPU serving Qwen3.8-27B. Self-hosted on the direct route with EDGAR off: nothing.
  - What it will not do: Audit, give an opinion or say 'compliant'. A clean result reads 'every figure tied to the tables (review still required)'; untraced figures are left for a person.
  - Input formats: An iXBRL 10-K or 10-Q up to 20 MB, a CIK and accession number, or up to 30 CSV tables with a draft of up to 120,000 characters; an optional trial balance CSV (account, name, balance, line).
  - Typical run: The figures are tied in code, a whole 10-K MD&A in about a second with no model. Each claim in words that is read is one gateway call, up to the run's claims cap, so a 10-K with many claims costs several times the planted sample; the cost per filing shown is measured on whole filed 10-Ks with their claims read.
- Samples:
  - 1. Amd fy2025 planted (`amd-fy2025-planted`): https://decosa.ai/tools/finance/filing-tieout?sample=1&autorun=0
  - 2. Aaoi fy2025 (`aaoi-fy2025`): https://decosa.ai/tools/finance/filing-tieout?sample=2&autorun=0
  - 3. Amd fy2025 (`amd-fy2025`): https://decosa.ai/tools/finance/filing-tieout?sample=3&autorun=0
  - 4. Synthetic tables (`synthetic-tables`): https://decosa.ai/tools/finance/filing-tieout?sample=4&autorun=0
- Regulatory note: Checked 26 Sep 2026. Pre-release financial statements are material non-public information, so this is self-host first: the hosted demo ties public EDGAR filings, the bundled samples and text you declare public, and refuses the rest. MD&A content is set by SEC Regulation S-K Item 303 (eCFR, current text); the tie-out checks that the figures agree with the statements, not that the discussion is complete or its explanations right. Statements and notes are tagged in inline XBRL under Regulation S-T Rule 405; it reads the tags as filed and does not validate them. EDGAR requests follow SEC's fair-access policy (a declared User-Agent, well under 10 requests a second). PCAOB QC 1000 (a firm's system of quality control) is context for audit firms; we understand it to be effective 15 December 2026 after a postponement, but could not confirm the date on the PCAOB docket page (unverified). It is not an audit, gives no opinion and never says 'compliant'. Not legal or accounting advice. Model licence: Apache-2.0 (Qwen3.8-27B).

### Disposition a sanctions alert (`sanctions-disposition-record`)

Clears one sanctions screening alert at a time, for the analyst. Code compares the customer with the OFAC, EU or UK list entry field by field: name across transliterations and name order, date of birth, nationality, ID numbers, IMO number, address and the kind of party. It proposes true match, false positive or needs more information by a named rule. An open model gives an independent reading and writes a rationale, and every sentence of it is checked against the fields it cites. The analyst decides, and the decision is sealed in a signed, hash-chained record that pins the list version, for a 10-year log you keep.

- What it is: Clear a screening alert field by field, with a rule-based proposal, a rationale checked against the fields, and a signed record that pins the list version.
- Who it's for: Teams in finance and insurance and compliance and trust.
- Page: https://decosa.ai/tools/finance/sanctions-disposition-record
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/sanctions-disposition-record):
  - 0 of 1,094: Same-party pairs proposed as false positive (the risky direction) (test split, n = 1094)
  - 635 / 635 (100%): False-positive clearance precision (test split, n = 635)
  - 863 (78.9%): Same-party pairs proposed as true match (test split, n = 1094)
  - 8.4 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B explains; the comparison, the proposal and the checks are plain code. Where: Self-host for real alerts; the hosted demo takes synthetic customers only (list entries are real public data).
- Verification: Receipt per model call, embedded in a signed, hash-chained record with the list snapshot's hash; every rationale sentence checked against its fields (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for Qwen3.8-27B; the comparison, the proposal and the record run on CPU
- Output: A signed, hash-chained disposition record (list version, fields compared, proposal, rationale, analyst, time), a Markdown memo, and a signed audit sample on request.
- Page: https://decosa.ai/tools/finance/sanctions-disposition-record · JSON: https://decosa.ai/use-cases/sanctions-disposition-record.json
- Prompts: https://decosa.ai/prompts/sanctions-disposition-record-hosted.md · https://decosa.ai/prompts/sanctions-disposition-record-selfhost.md · https://decosa.ai/prompts/sanctions-disposition-record-assemble.md
- Tiers:
  - Lite · the comparison and the proposal, no GPU (Any CPU (about 200 MB of RAM for the list snapshot)): POST /sanctions/analyze compares every field and proposes a disposition by rule, and /decide seals the analyst's decision; no model reading, no rationale.
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Adds Qwen3.8-27B's independent typed reading and a rationale whose every sentence is checked against the fields it cites. This is what the hosted demo runs, on synthetic customers only.
- Alternates (smaller models for the same job):
  - Names in every script (1x RTX PRO 6000 96 GB (estimate)), not built: Compare names written in Arabic, Persian, Chinese and Korean script directly, not only their Latin forms. The model is not chosen yet; it must be licence-clean.
- Buyer facts:
  - Data retention: Nothing stored: the alert lives in memory for the request, and the review and the record go back to you. Logs carry the list, rule and decision, never names, dates of birth or numbers. You keep the records (10 years under 31 CFR 501.601).
  - What leaves the box: Hosted: the two model calls go through our gateway to the GPU serving Qwen3.8-27B, and only synthetic customers are accepted. Self-hosted on the direct route: nothing leaves the box. Refreshing the lists downloads the public list files; no customer data is sent.
  - What it will not do: Decide, screen your customer base, or say 'compliant'. It proposes; the analyst decides and the record says who.
  - Lists: OFAC SDN and Consolidated, EU consolidated, UK Sanctions List: 32,452 entries in the 26 Sep 2026 snapshot, each source file's sha256 and publish date recorded. Rebuild the snapshot any day with one command.
  - Typical run: Two model calls, a fraction of a cent at the gateway list price; the analysis and the record are free.
- Samples:
  - 1. Translit true match (`translit-true-match`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=1&autorun=0
  - 2. Common name false positive (`common-name-false-positive`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=2&autorun=0
  - 3. Dob mismatch (`dob-mismatch`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=3&autorun=0
  - 4. Alias true match (`alias-true-match`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=4&autorun=0
  - 5. Name only (`name-only`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=5&autorun=0
  - 6. Company vs vessel (`company-vs-vessel`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=6&autorun=0
  - 7. Vessel imo (`vessel-imo`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=7&autorun=0
  - 8. Cyrillic true match (`cyrillic-true-match`): https://decosa.ai/tools/finance/sanctions-disposition-record?sample=8&autorun=0
- Regulatory note: Checked 26 Sep 2026 against primary sources (links under Tools). OFAC's recordkeeping rule, 31 CFR 501.601, now requires a full and accurate record of each transaction subject to OFAC's regulations, kept at least 10 years after the transaction; blocked-property records must be kept 10 years after unblocking. The change from 5 to 10 years came in an interim final rule (89 FR 74832, 13 Sep 2024, effective 12 Mar 2025), which a final rule adopted without change (90 FR 13286, published and effective 21 Mar 2025). It matches the 10-year statute of limitations for IEEPA and TWEA violations that took effect on 24 Apr 2024 (Pub. L. 118-50, as summarised in those rules). The rule text speaks of transactions and blocked property; it does not name screening alerts. Keeping an alert's disposition with the transaction record for the same 10 years is our conservative reading, not quoted text, so the record's retain_until is 10 years from the later of the decision and the transaction date. The proposal rules follow OFAC FAQ 5 (updated 9 Sep 2026). Step 3 there says to compare all the details on the list entry, that an address alone may not be enough, and that many potential matches are false positives. Step 4 says multiple matching identifiers suggest a valid match, and the organisation decides under its own procedures. Weak aliases follow OFAC FAQs 122-123. For the UK, the OFSI Consolidated List closed on 28 Jan 2026, and the UK Sanctions List (FCDO) is now the only source; the snapshot uses it. OFAC notes that EU data-retention limits can conflict with the 10-year rule (90 FR 13286); where they do, that is a question for your counsel. This is decision support, not legal advice and never a compliance determination: the analyst decides. The hosted demo takes synthetic customers only. Model licence: Apache-2.0 (Qwen3.8-27B). List data: US government works, EU list (Commission Decision 2011/833/EU reuse), UK list (Open Government Licence assumed, not stated in the file).

### Review a Medicare sales call (`medicare-call-record`)

QA and a signed record for recorded Medicare Advantage and Part D sales calls. On the call it answers yes, no or unclear, with a quote and a time: the TPMO disclaimer and whether it came before benefits were discussed, the Scope of Appointment, "free" for a $0 premium, claiming to be Medicare, gifts, pressure, and the steps of an enrollment. Every benefit the agent states is checked against the plan's Summary of Benefits. Signed record with the retention dates of the 1 Jun 2026 rule.

- What it is: A recorded Medicare Advantage or Part D sales call checked against the CMS marketing rules, each answer quoted with its time, every benefit claim checked against the plan's Summary of Benefits, sealed in a signed record.
- Who it's for: TPMO and FMO compliance teams, independent Medicare agents, and plan broker-oversight teams.
- Page: https://decosa.ai/clinics/medicare-call-record
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/medicare-call-record):
  - 40/40: Per-rule accuracy, test B run 1 (never used to change anything) (held out, n = 40)
  - 101/101: Per-rule accuracy, test run 4 (test split, n = 101)
  - 9/9: Planted typed answers found, test run 4 (test split, n = 9)
  - 7.5 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: MOSS-Transcribe-Diarize · Qwen3.8-27B. Where: Self-host for real calls (hosted demo: synthetic role-plays only).
- Verification: Receipt per model call; every yes backed by a quote found in the transcript; each benefit claim judged against the plan facts with the span cited; signed hash-chained record with retention dates (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Qwen3.8-27B and the diarizer; the call-sheet items and the record need no GPU
- Output: A QA checklist of typed answers with quotes and times, the call-sheet items, each benefit claim with its verdict against the plan facts, and a signed record with its retention dates.
- Page: https://decosa.ai/clinics/medicare-call-record · JSON: https://decosa.ai/use-cases/medicare-call-record.json
- Prompts: https://decosa.ai/prompts/medicare-call-record-hosted.md · https://decosa.ai/prompts/medicare-call-record-selfhost.md · https://decosa.ai/prompts/medicare-call-record-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same checks and claim judge on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B extracts, answers every typed check and judges every benefit claim; MOSS-Transcribe-Diarize turns recordings into transcripts; the call-sheet items are code. Every model call receipted.
  - Best · two 96 GB cards (2x RTX PRO 6000 Blackwell 96 GB): A larger model for long calls and many benefit claims. Not measured on this task.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Patient calls stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: On each recorded sales call: the TPMO disclaimer and whether it came before benefits, a Scope of Appointment read on the call, "free" for a $0 premium, claiming to be Medicare, an Advantage plan called a supplement, gifts, pressure, and for an enrollment the notice, a clear yes, where the Summary of Benefits is, the needs topics and the pre-enrollment checklist. From the call sheet: the SOA's date and scope, non-health products, cold calls, the disclaimer numbers. Every benefit claim against the Summary of Benefits.
  - What it does not do: It gives no compliance verdict. It checks benefit claims against the plan facts you give, not the plan's filed benefits: a partial Summary of Benefits produces false flags. It does not check the written SOA form, SMIDs, the Star Ratings document, the plan's outbound enrollment verification, agent licensing, compensation or lead-generation data sharing.
  - Record-keeping dates: The record carries the dates the rule sets since 1 Jun 2026: audio for 3 years, audio or a complete and accurate transcript to year 6 (42 CFR 422.2274(g)(2)(ii)), and 10 years for the enrollment portion of a telephonic enrollment. A machine transcript counts as complete and accurate only after a person checks it; the record takes that attestation.
  - Consent: Every run needs a statement that the call was recorded with the consent the law requires (some states need every party's). The statement goes into the signed record, and a check looks for the recording notice on the call.
  - Data retention: Nothing on the server. Transcripts, call sheets, plan facts and records live in memory for the request, and logs carry counts and timings only. You keep the signed record with the recording until its retention dates.
  - What leaves the box (hosted demo): The transcript and the plan facts go to Qwen3.8-27B through our gateway, which is Decosa-operated. The call sheet is processed in code and not sent to the model. The gateway's receipts hold hashes, not text. Self-hosted, nothing leaves.
  - Model calls per call: Two extractions, 8 or 9 typed checks and one grounding call per benefit claim (up to 12): 13 to 19 calls on the eval calls; the smoke sample used 15 calls and 15,120 tokens, about $0.006.
- Samples:
  - 1. Tb1 showcase (`tb1-showcase`): https://decosa.ai/clinics/medicare-call-record?sample=1&autorun=0
  - 2. T6 cold call pressure (`t6-cold-call-pressure`): https://decosa.ai/clinics/medicare-call-record?sample=2&autorun=0
  - 3. T8 enroll missing elements (`t8-enroll-missing-elements`): https://decosa.ai/clinics/medicare-call-record?sample=3&autorun=0
  - 4. T9 supplement gift giveback (`t9-supplement-gift-giveback`): https://decosa.ai/clinics/medicare-call-record?sample=4&autorun=0
  - 5. T7 outside scope life (`t7-outside-scope-life`): https://decosa.ai/clinics/medicare-call-record?sample=5&autorun=0
  - 6. T1 clean enroll hmo (`t1-clean-enroll-hmo`): https://decosa.ai/clinics/medicare-call-record?sample=6&autorun=0
  - 7. T3 clean pdp enroll soa on call (`t3-clean-pdp-enroll-soa-on-call`): https://decosa.ai/clinics/medicare-call-record?sample=7&autorun=0
  - 8. T4 clean free allowed (`t4-clean-free-allowed`): https://decosa.ai/clinics/medicare-call-record?sample=8&autorun=0
- Regulatory note: Checked 26 Sep 2026 against eCFR (title 42 current to 24 Sep 2026) and the Federal Register. Medicare Advantage and Part D communications and marketing, 42 CFR 422.2260-422.2274 and 423.2260-423.2274, as amended by the CY2027 final rule (91 FR 17384, 6 Apr 2026, effective 1 Jun 2026; CMS's tip sheet 12237-P applies the marketing changes to CY2027 marketing from 1 Oct 2026). Encoded: the TPMO disclaimer, now said before any discussion of benefits (was: within the first minute) and without the SHIP reference (422.2267(e)(41)); the Scope of Appointment agreed and recorded before a personal marketing appointment, no 48-hour wait any more, no products beyond its scope, valid 12 months (422.2264(c)(3), 422.2274(b)(3)); no non-health products such as annuities (422.2263(b)(4); life insurance is treated as non-health here, our reading); no unsolicited calls (422.2264(a)); no "free" for a $0 premium or reduced cost sharing, no claim of Medicare endorsement or misleading use of the Medicare name, no implying a plan is a supplement (422.2262(a)(1)); no cash or non-nominal gifts (422.2263(b)); the needs topics before an enrollment (422.2274(c)(12)), the pre-enrollment checklist reviewed and the Summary of Benefits location given on a telephonic enrollment (422.2267(e)(4), (e)(5)); the telephonic SOA elements from the MCMG (16 Mar 2022). Retention: marketing and sales calls recorded in full and kept 6 years, in audio for years 1-3 and audio or a complete and accurate transcript for years 4-6 (422.2274(g)(2)(ii) as amended; 10 years before 1 Jun 2026); the enrollment portion of a telephonic enrollment is the enrollment form and stays under the 10-year record rule (422.504(d); 91 FR 17463-17464). "Pressure tactics" is not a defined term; it is checked as misleading conduct (422.2262). Recording consent is state law (e.g. California Penal Code 632). Sales calls carry PHI: HIPAA applies to plans and their business associates. QA triage, not legal advice.

### Appeal a denial (`denial-appeal-packet`)

For billing, revenue-cycle and denials staff, and patient advocates. Give it the denial (EOB, remittance or letter with its CARC/RARC codes), the chart excerpts and the payer's coverage criteria. It computes the next appeal level's deadline with the federal rule and the arithmetic (ERISA and ACA, Medicare, Medicare Advantage, Medicaid managed care) and marks each criterion met, not met or undocumented, with a quote from the chart. It drafts a letter only when the chart supports every criterion, and checks every sentence against the chart and the policy. When the chart doesn't support it, it says don't appeal, or what to get first. The packet is signed.

- What it is: A denial, the chart and the payer's criteria in; the appeal deadline, a criteria checklist with quotes, and a letter only when the chart supports it.
- Who it's for: Billing, revenue-cycle and denials staff at providers and DME suppliers, and patient advocates.
- Page: https://decosa.ai/clinics/denial-appeal-packet
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/denial-appeal-packet):
  - 34 of 40: Held-out phrasing cases, recommendation right (held out, n = 40)
  - 36 of 40: Recommendation right, fresh held-out set (test split, n = 40)
  - 26 of 26: Unsupported cases with no appeal and no letter (test split, n = 26)
  - 12.9 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real patients (hosted demo: synthetic cases only).
- Verification: Receipt per model call; every criterion quote found word for word in the chart; every letter sentence grounded; signed hash-chained packet (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the deadline rules and the signed packet need no GPU
- Output: A recommendation, the filing deadline with its rule, a criteria checklist with quotes, a draft letter when the chart supports it, a Markdown packet and a signed record.
- Page: https://decosa.ai/clinics/denial-appeal-packet · JSON: https://decosa.ai/use-cases/denial-appeal-packet.json
- Prompts: https://decosa.ai/prompts/denial-appeal-packet-hosted.md · https://decosa.ai/prompts/denial-appeal-packet-selfhost.md · https://decosa.ai/prompts/denial-appeal-packet-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B reads the denial, maps and checks the criteria and drafts the letter; the deadline, the quote checks and the recommendation are plain code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long charts and dense payer policies.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Patient records stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it gives you: A recommendation (appeal, don't appeal, get documentation first, fix the claim), the next level's filing deadline with its CFR rule and arithmetic, each criterion met, not met or undocumented with quotes, a draft letter when the chart supports it, a Markdown packet and a signed record.
  - What it does not do: It does not file, fax or submit to a portal, predict outcomes, read plan documents or provider contracts, know state rules, or fetch payer policies (paste the criteria). Non-covered indications are not mapped. Scanned charts need OCR first.
  - Data retention: Nothing kept on the server. Denials, charts and packets live in memory for the request; logs carry counts only. You keep the signed packet with the claim.
  - What leaves the box (hosted demo): The denial, chart excerpts and policy go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for synthetic cases only.
  - Model calls per packet: One denial read, one criteria map, one check per criterion; with a letter, one draft and one grounding call per letter sentence: about 1 to 21 calls on the demo cases.
  - Typical run cost: A fraction of a cent at the gateway list price for the CPAP case with a letter (a couple of dozen calls); less when it says don't appeal. Each run shows its own measured cost.
  - Deadlines: Federal minimums for ERISA and ACA plans, Medicare A/B, Medicare Advantage and Medicaid managed care, from the notice date (Medicare presumes receipt 5 days later). Give the receipt date and any shorter provider-contract window; the earlier date wins.
- Samples:
  - 1. Cpap appeal (`cpap-appeal`): https://decosa.ai/clinics/denial-appeal-packet?sample=1&autorun=0
  - 2. Cpap no appeal (`cpap-no-appeal`): https://decosa.ai/clinics/denial-appeal-packet?sample=2&autorun=0
  - 3. Bariatric gather (`bariatric-gather`): https://decosa.ai/clinics/denial-appeal-packet?sample=3&autorun=0
  - 4. Mri external review (`mri-external-review`): https://decosa.ai/clinics/denial-appeal-packet?sample=4&autorun=0
  - 5. Admin missing npi (`admin-missing-npi`): https://decosa.ai/clinics/denial-appeal-packet?sample=5&autorun=0
- Regulatory note: Charts are protected health information under HIPAA: run it on the provider's own hardware; the hosted demo takes synthetic cases only. Deadline rules read on 26 Sep 2026 from eCFR (titles current to 24 Sep 2026): ERISA group health plans, at least 180 days after receipt of the denial to appeal (29 CFR 2560.503-1(h)(3)(i)), with plan decisions due in 72 hours (urgent), 30 days (pre-service) or 60 days (post-service) (2560.503-1(i)(2)); ACA individual coverage, the same rules with one internal level (45 CFR 147.136(b)(3)); federal external review, four months after receipt of the final internal denial, the first day of the next month when that day does not exist, moved off weekends and federal holidays (29 CFR 2590.715-2719(d)(2)(i), 45 CFR 147.136(d)(2)(i)); state processes must allow at least four months. Medicare Parts A and B: redetermination 120 days, reconsideration 180 days, ALJ hearing, Council review and court 60 days each, all from receipt presumed 5 days after the notice (42 CFR 405.942, 405.962, 405.1002, 405.1102, 405.1130/405.1136); amount in controversy $200 for an ALJ hearing and $1,960 for court in 2026, $2,000 for court from 1 Jan 2027 (Federal Register, 4 Dec 2025 and 16 Sep 2026). Medicare Advantage: reconsideration 60 days after presumed receipt (42 CFR 422.582(b)); ALJ 60 days (422.602(b)). Medicaid managed care: plan appeal 60 days from the date on the notice (42 CFR 438.402(c)(2)(ii)); state fair hearing 90 to 120 days as the state sets (438.408(f)(2)). These are federal minimums: plan documents, state programs and provider contracts can differ. Criteria come only from the policy text pasted in; the demo uses CMS National Coverage Determinations (public domain) and an invented payer policy. CARC/RARC labels are our own summaries; the official descriptions belong to X12. Not legal or medical advice, and it never predicts an outcome.

### Check if a video is made for kids (`kids-content-preflight`)

Before a video, short or app listing goes out, checks whether it is directed to children. Each of the FTC's COPPA factors (subject matter, characters, music, the age of the people on screen, language, ads) gets its own typed answer and probability, with the lines that show it. The call is compared with your made-for-kids setting: kids-directed but set not-for-kids by a channel default is the pattern in the FTC's $10M Disney order. When it is kids-directed, it lists what has to change: comments, personalised ads, links, requests for children's details, undisclosed paid promotion, AI labels, and for apps purchases, third-party ads and analytics. A named reviewer makes the designation, sealed in a signed review record. A back-catalogue audit checks many uploads at once. Not a COPPA compliance determination.

- What it is: Before a video, short or app listing goes out: is it directed to children, factor by factor with the lines that show it; does your made-for-kids setting match; what has to change if it does; and a named reviewer's designation in a signed review record.
- Who it's for: Trust-and-safety and legal leads at brand channels, MCNs, and kids' app and game studios.
- Page: https://decosa.ai/tools/media/kids-content-preflight
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/kids-content-preflight):
  - 10 / 12: Frame check: picture-only stills, vision on (held out, n = 12)
  - 39 / 40: Made for kids or not (full check) (test split, n = 40)
  - 37 / 40: Audience, 4-way (primary / mixed / general / mature) (test split, n = 40)
  - 7.2 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B (typed answers per factor; optional frame descriptions) · MOSS-Transcribe-Diarize for video audio. Where: Hosted or self-host; unreleased episodes and app builds on your own machine.
- Verification: Receipt per model call; signed ASR receipt for video audio; signed review record naming the reviewer and their designation (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: Qwen3.8-27B (1× RTX 5090 32 GB or larger); speech to text for videos needs MOSS-Transcribe-Diarize on a GPU; OCR and the record run on CPU
- Output: The audience call with a probability per factor and the lines behind it, the setting check, what to change, and a signed review record with the reviewer's designation.
- Page: https://decosa.ai/tools/media/kids-content-preflight · JSON: https://decosa.ai/use-cases/kids-content-preflight.json
- Prompts: https://decosa.ai/prompts/kids-content-preflight-hosted.md · https://decosa.ai/prompts/kids-content-preflight-selfhost.md · https://decosa.ai/prompts/kids-content-preflight-assemble.md
- Tiers:
  - Lite · audience only (the catalogue audit) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (not measured)): One typed audience question per upload (5 calls on the gateway, 1 with logprobs self-hosted) and the setting check; no factor answers, evidence or content flags. Feature flags still come from what you declare.
  - Standard · full check on text (hosted demo) (1x RTX PRO 6000 96 GB (measured); the diarizer on a second GPU in the hosted setup): Evidence lines per FTC factor, the typed audience call, nine typed factor answers, the setting check and every flag; speech to text and frame OCR for videos.
  - Best · adds the frame check (vision) (1x RTX PRO 6000 96 GB with the model served with its vision tower (measured through the hosted gateway)): Everything in Standard, plus four frames of a video described by Qwen3.8-27B's image input (one more receipted call, about 3,500 more prompt tokens), so the picture counts, not only the words.
- Buyer facts:
  - What it checks: Whether an upload is directed to children under 13 (one typed answer per FTC factor, with the lines that show it), whether your made-for-kids setting matches, and, when it is kids-directed, comments, personalised ads, links, requests for children's details, undisclosed paid promotion, AI labels, and for apps purchases without a parental gate, third-party ads and analytics, chat and the data collected.
  - What it does not do: It is not a COPPA compliance determination or legal advice, and a named person makes the designation. Feature flags come from what you declare; audience analytics are not read unless you put them in the text. Without the vision option, the picture itself is read only by OCR.
  - Data retention: An uploaded video is deleted when its run ends. Reports and records are kept in memory for one hour, for the token or key that made them. Logs carry run ids, statuses and counts, never titles, transcripts, names or links. You keep the signed review record.
  - What leaves the box (hosted demo): The upload's text (and, with the vision option, four frames) goes to Qwen3.8-27B through our gateway, which Decosa operates; video audio is transcribed on Decosa's hosted service. The gateway's receipts hold hashes, not text. Self-hosted, nothing leaves.
  - Children's data: None is needed. The pre-flight reads what a channel publishes about the upload, never viewers' data; the demo samples are synthetic.
  - Cost per upload: A fraction of a cent for the full check, and less in the audience-only audit, at the gateway list price (measured on the test set).
  - Output: The audience call with probabilities, nine factors with their lines, the setting check, what to change, and a signed review record naming the reviewer and their designation (Markdown and JSON exports; verify at /record/verify).
- Samples:
  - 1. Preschool puppet (`preschool-puppet`): https://decosa.ai/tools/media/kids-content-preflight?sample=1&autorun=0
  - 2. Momo video (`momo-video`): https://decosa.ai/tools/media/kids-content-preflight?sample=2&autorun=0
  - 3. Parenting vlog (`parenting-vlog`): https://decosa.ai/tools/media/kids-content-preflight?sample=3&autorun=0
  - 4. Game trailer (`game-trailer`): https://decosa.ai/tools/media/kids-content-preflight?sample=4&autorun=0
  - 5. Adult cartoon (`adult-cartoon`): https://decosa.ai/tools/media/kids-content-preflight?sample=5&autorun=0
  - 6. Kids app listing (`kids-app-listing`): https://decosa.ai/tools/media/kids-content-preflight?sample=6&autorun=0
  - 7. Toy unboxing (`toy-unboxing`): https://decosa.ai/tools/media/kids-content-preflight?sample=7&autorun=0
  - 8. Ai science host (`ai-science-host`): https://decosa.ai/tools/media/kids-content-preflight?sample=8&autorun=0
  - 9. Catalogue audit (`catalogue-audit`): https://decosa.ai/tools/media/kids-content-preflight?sample=9&autorun=0
- Regulatory note: Not legal advice and not a COPPA compliance determination; checked against the primary texts on 26 Sep 2026. COPPA Rule, 16 CFR 312.2 (as amended by 90 FR 16918, published 22 Apr 2025, effective 23 Jun 2025, compliance by 22 Apr 2026): to decide whether a website or online service, or a portion of one, is directed to children under 13, the FTC considers its subject matter, visual content, animated characters or child-oriented activities and incentives, music or other audio, the age of models, child celebrities or celebrities who appeal to children, language, and whether ads promoting or appearing on it are directed to children, plus evidence of audience composition and of the intended audience (the 2025 amendment adds marketing materials, representations, reviews and the age of users of similar services). A mixed-audience service is directed to children under those factors without targeting them as its primary audience, and must not collect personal information before determining age neutrally. 16 CFR 312.5 requires verifiable parental consent before collecting personal information from children. United States v. Disney (FTC complaint and proposed order 2 Sep 2025; court-approved order announced 31 Dec 2025): $10M civil penalty for child-directed videos on channels set 'not made for kids' at channel level, and a programme to review whether videos posted to YouTube should be designated made for kids, unless YouTube implements age assurance for all users or stops letting creators label videos. YouTube's help pages (undated; read 26 Sep 2026) treat content as made for kids when children are the primary audience, or a secondary audience that the content is still directed at, list the features turned off on made-for-kids content (comments, personalised ads, the notification bell, live chat, end screens and others), and say YouTube may override a setting. Apple App Review Guidelines 1.3, 2.3.8 and 5.1.4 and Google Play's Families policies govern kids' apps and their listings. 16 CFR 255.6: endorsements addressed to children may be of special concern. Model licences: Apache-2.0 (Qwen3.8-27B, MOSS-Transcribe-Diarize); Tesseract Apache-2.0.

### Check HCC codes for MEAT evidence (`hcc-evidence-file`)

For Medicare Advantage risk-adjustment and compliance teams. Send a member's submitted diagnosis codes and the notes for the year. For each code that maps to a CMS-HCC V28 payment HCC it says supported, not supported or insufficient, quotes the sentences that show the condition was monitored, evaluated, assessed or treated (MEAT) or says none do, and checks each record the way a RADV reviewer would: face to face, signed with an acceptable credential, in the year, no audio-only calls, radiology reports or late addenda. Deletions come first. It never suggests codes to add. Out comes the net effect, an evidence file and a signed record. A review aid: certified coders decide.

- What it is: Each submitted risk-adjustment code checked against the notes: keep, hold or delete, with the MEAT quotes, and never a suggestion to add one.
- Who it's for: Medicare Advantage risk-adjustment and compliance teams, and provider groups in risk contracts, preparing for RADV audits.
- Page: https://decosa.ai/clinics/hcc-evidence-file
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/hcc-evidence-file):
  - 189 / 200: Verdict accuracy, 3 classes (test split, n = 200)
  - 96 / 100: Supported vs not supported (test split, n = 100)
  - 0 / 161: Unsupported or held codes called supported (false keep) (test split, n = 161)
  - 23.9 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B reads the notes (evidence, typed judgment, grounding); the V28 map, hierarchies and record checks are plain code. Where: Self-host for real records (PHI); the hosted demo takes synthetic members only.
- Verification: Receipt per model call; every quote is a sentence of the note with offsets; signed hash-chained record with no note text (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for Qwen3.8-27B; the V28 map, the record checks and the signed record run on CPU
- Output: Per submitted code a delete, hold or keep with MEAT quotes, the net effect in codes and paying HCCs, a Markdown evidence file (deletions first) and a signed decosa.record.v1.
- Page: https://decosa.ai/clinics/hcc-evidence-file · JSON: https://decosa.ai/use-cases/hcc-evidence-file.json
- Prompts: https://decosa.ai/prompts/hcc-evidence-file-hosted.md · https://decosa.ai/prompts/hcc-evidence-file-selfhost.md · https://decosa.ai/prompts/hcc-evidence-file-assemble.md · https://decosa.ai/prompts/hcc-evidence-file-mac.md
- Tiers:
  - Lite · map and record checks only, no GPU (Any CPU): POST /hcc/codes maps codes to V28 payment HCCs after the hierarchies; POST /hcc/records runs the RADV record checks on every note and addendum. No reading of the notes, so no MEAT verdicts.
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B reads the notes per code (evidence, typed judgment, grounding); the map, record checks, verdict rules and guard are code. This is what the hosted demo runs, on synthetic members only.
  - Best · scanned charts too (adds the document reader) (1x RTX PRO 6000 96 GB for Qwen3.8, and the document reader on about 6 GB of a GPU (measured 27 Sep: 5.7 GB peak for parser and layout, 1.6 to 2.4 s a page, the GPU shared) or on a Mac Studio (measured: 3.3 s a page, 2.9 GB)): Everything in Standard, plus scanned PDFs and images: the document reader turns each page into a note with its date, record type, provider, credential and signature read as fields tied to boxes, and every MEAT quote is cited to a page and box of the scan.
- Buyer facts:
  - Data retention: Nothing stored: the member file lives in memory for the request. The signed record holds hashes, verdicts and receipt ids, never note text; logs carry counts only.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and only synthetic members are accepted. Self-hosted on the direct route: nothing leaves the box. Scanned charts: the document reader runs on Decosa's hosted service (self-hosted: on yours).
  - What it will not do: Suggest codes to add, search notes for new codes, or draft provider queries. It checks only the submitted codes, lists deletions first, and a certified coder decides.
  - Pricing basis: Per member reviewed (or a flat self-host licence). Never a share of recovered, protected or captured revenue.
  - Input formats: JSON: member id and service year (age and sex optional: they only pick CMS's age and sex edits in the code map), up to 20 submitted codes, up to 12 notes of up to 12,000 characters (60,000 in total) with date, record type, provider, credential, specialty, signature and addenda. Or, through the API only (not on this page), a scanned chart (PDF or image, one record per page, up to 10 pages and 12 MB hosted) through POST /hcc/read-chart, read by the document reader.
  - Typical run: The clean sample: a handful of model calls and a fraction of a cent at the gateway list price, under a minute on the shared gateway. Each run shows its own measured cost.
- Samples:
  - 1. Mixed file (`mixed-file`): https://decosa.ai/clinics/hcc-evidence-file?sample=1&autorun=0
  - 2. Clean file (`clean-file`): https://decosa.ai/clinics/hcc-evidence-file?sample=2&autorun=0
  - 3. Edge cases (`edge-cases`): https://decosa.ai/clinics/hcc-evidence-file?sample=3&autorun=0
- Regulatory note: Checked 26 Sep 2026 against CMS primary sources (links under Tools). CMS said on 21 May 2025 that it will audit all eligible MA contracts for each payment year (about 550, up from about 60) and grow its medical coders from 40 to about 2,000. The PY 2024 RADV Audit Methods and Instructions (28 Aug 2026) define a valid record as a legibly signed record from a face-to-face visit with a provider in the data collection period, accept audio-and-video telehealth as face to face (HPMS memo, 4 May 2022), and list the invalid-record codes this tool uses (INV2 no signature, INV4 no date, INV5 invalid source such as lab only, home health, hospice, superbill or non-face-to-face, INV7 no credential, INV14 outside the period). The Contract-Level RADV Medical Record Reviewer Guidance (10 Jan 2020) excludes diagnostic radiologists and telephone contacts, treats amendments as timely at up to 90 days and only from the treating provider, says a third party may not amend a record or query the provider for additional diagnoses, and lists acceptable specialties (Appendix B, PY 2015; CMS points to its CSSC site for the current list, which we did not check). The ICD-10-CM Official Guidelines FY 2026 (IV.H and IV.J) bar coding uncertain diagnoses in the outpatient setting and conditions that no longer exist. MEAT (monitor, evaluate, assess, treat) is the coding industry's documentation test, not a CMS term. The Kaiser Permanente affiliates' $556M settlement (14 Jan 2026) concerned diagnoses added through addenda; we could not load the DOJ page on 26 Sep 2026, so that detail is from our research notes (unverified here). Clinical notes are PHI under HIPAA: self-host first; the hosted demo takes synthetic members only. The tool never suggests adding codes, and any price will be per member reviewed, never a share of recovered or protected revenue (prices aren't published yet). This is not legal or coding advice and not an audit determination: certified coders decide. Model licence: Apache-2.0 (Qwen3.8-27B). The V28 table is a US government work (public domain).

### Draft VEX for scanner findings (`vex-triage`)

For product security teams and software vendors. Give it a Grype or Trivy report for a container image, its SBOM, and an evidence bundle from a small collector that runs next to the image. For each CVE, code checks the package, the version against the fix and the advisory ranges, the distro changelog, what actually loads the library, the settings the advisory names, and the platform. An open model then picks a VEX status with quoted evidence. Not affected needs positive evidence; everything else stays affected, fixed or under investigation. You get signed OpenVEX and CycloneDX VEX drafts and a signed evidence record for a security engineer to sign off.

- What it is: A Grype or Trivy report in; a VEX status per CVE with the evidence from the image, signed OpenVEX and CycloneDX VEX out.
- Who it's for: Product security (PSIRT) and AppSec teams at software vendors shipping container images, and platform teams that own base images.
- Page: https://decosa.ai/tools/developer/vex-triage
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/vex-triage):
  - 154 / 159: not_affected precision against Canonical's VEX (test split, n = 159)
  - 5 / 378: False not_affected on vendor-affected findings (test split, n = 378)
  - 154 / 474: not_affected recall (test split, n = 474)
  - 14.3 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B. Where: Hosted or self-host (air-gapped with a local advisory store).
- Verification: Receipt per model call; every quote found in the advisory or check it cites; not_affected only on a strong code check; DSSE-signed VEX documents; signed hash-chained evidence record (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the collector, the checks, rules mode and the signing run on CPU
- Output: Draft VEX statements with evidence, signed OpenVEX and CycloneDX VEX documents (DSSE), a Markdown report and a signed evidence record.
- Page: https://decosa.ai/tools/developer/vex-triage · JSON: https://decosa.ai/use-cases/vex-triage.json
- Prompts: https://decosa.ai/prompts/vex-triage-hosted.md · https://decosa.ai/prompts/vex-triage-selfhost.md · https://decosa.ai/prompts/vex-triage-assemble.md
- Tiers:
  - Lite · rules only, no GPU (Any CPU): The same checks and gates with no model call (mode "rules"): the same statuses on the held-out set, but no written impact statement beyond the check's own text, and nothing routed to under investigation.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): The checks plus one Qwen3.8-27B judgment per finding the rules do not settle: a written impact statement with quotes, and unclear cases sent to a person. Every model call receipted.
- Buyer facts:
  - What it gives you: A VEX status per finding with its evidence (the checks and the advisory lines it rests on), signed OpenVEX and CycloneDX VEX drafts in DSSE envelopes, a Markdown report and a signed evidence record listing every advisory's hash.
  - What it does not do: It does not scan (it triages a Grype or Trivy report), trace calls inside programs, see configs mounted at run time or dlopen by path, or decide attacker control or mitigations. It never says an image is secure; a security engineer signs off.
  - Data retention: Nothing kept but advisory text: reports, SBOMs and evidence bundles live in memory for the request; logs carry counts only. The image never reaches the API: the collector runs on your machine.
  - What leaves the box: Hosted: the scanner report, SBOM and evidence bundle (search hits, loader chains, config excerpts with secret-looking lines redacted) go to decosa-api, and the prompts to Qwen3.8-27B through our gateway, whose receipts hold hashes. Both routes: vulnerability ids to OSV.dev and NVD, which you can switch off.
  - Round trip: Tested: grype --vex with the signed OpenVEX from the nginx demo hides exactly the four findings it proves not affected and keeps the other 546.
  - Typical run cost: One model call per finding the rules do not settle: about a cent for the nginx sample at the gateway list price. Rules mode costs no model tokens.
  - Findings per run: 25 by default, 60 hosted (known-exploited first, then severity and EPSS); self-host raises it with DECOSA_VEX_TRIAGE_MAX.
- Samples:
  - 1. Nginx 1.20.0 (`nginx-1.20.0`): https://decosa.ai/tools/developer/vex-triage?sample=1&autorun=0
  - 2. Nginx 1.20.0 image filter (`nginx-1.20.0-image-filter`): https://decosa.ai/tools/developer/vex-triage?sample=2&autorun=0
  - 3. Python 3.10.0 slim (`python-3.10.0-slim`): https://decosa.ai/tools/developer/vex-triage?sample=3&autorun=0
- Regulatory note: Checked 26 Sep 2026. Statuses and justifications follow CISA's Minimum Requirements for Vulnerability Exploitability eXchange (VEX) (21 Apr 2023, https://www.cisa.gov/resources-tools/resources/minimum-requirements-vulnerability-exploitability-exchange-vex) and OpenVEX v0.2.0 (https://github.com/openvex/spec); CycloneDX 1.6 VEX uses its own justification vocabulary (mapping in GET /vex/info). EU Cyber Resilience Act, Regulation (EU) 2024/2847: vulnerability and incident reporting (Art. 14) applies from 11 Sep 2026 and the main obligations, including documenting components with a software bill of materials (Annex I Part II), from 11 Dec 2027 (Art. 71(2)); we could not re-read the Official Journal text on 26 Sep 2026, so treat these dates as from our 26 Sep incident-pack check, not independently verified here. VEX is a way to state exploitability; this tool drafts statements and never says an image is secure, and it makes no CRA or other compliance determination. Model licence: Apache-2.0 (Qwen3.8-27B). Advisory sources: OSV.dev (per-record licence; GitHub advisories CC BY 4.0) and NVD (US government work).

### Triage a device complaint for MDR (`device-mdr-triage`)

For medical device regulatory and quality teams. Give it a complaint and the product's labeling. It answers the 21 CFR 803.50(a) questions one by one, each with the words from the complaint that decide it: death or serious injury, could the device have caused or contributed, a malfunction, likely to cause serious harm if it recurred. It suggests reportable, not reportable or needs investigation, and starts the 30-day or 5-work-day clock from the day any employee heard, which the complaint itself may name. It checks the complaint record against 820.35(a), drafts the 3500A event description with every sentence checked, and signs a triage record and the named person's decision. A trend view groups complaints by failure mode. A triage aid: a person decides.

- What it is: A device complaint in; the 21 CFR 803.50 questions answered with quotes, the MDR clock from the right day, and a signed record of who decided.
- Who it's for: Regulatory affairs and quality teams at medical device makers, and complaint-handling service providers.
- Page: https://decosa.ai/tools/life-sciences/device-mdr-triage
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/device-mdr-triage):
  - 6 of 96 (6.3%): Reportable complaints called not reportable (test split, n = 96)
  - 0 of 58: Not-reportable complaints called reportable (test split, n = 58)
  - 53 of 58: Not-reportable complaints cleared (test split, n = 58)
  - 11.9 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B (typed answers, the narrative and its grounding check); the decision rule, the clocks and the file check are plain code. Where: Self-host for real complaints (they can hold PHI); the hosted demo takes public MAUDE narratives and synthetic complaints only.
- Verification: Receipt per model call; every answer that settles anything has a quote found word for word in the complaint; narrative sentences grounded; signed triage and decision records (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) for Qwen3.8-27B; the clocks, the file check and the signed records run on CPU
- Output: A suggestion with the reasoning chain and quotes, the clock with its rule, the complaint-record check, a grounded 3500A draft, a Markdown memo, and signed triage and decision records.
- Page: https://decosa.ai/tools/life-sciences/device-mdr-triage · JSON: https://decosa.ai/use-cases/device-mdr-triage.json
- Prompts: https://decosa.ai/prompts/device-mdr-triage-hosted.md · https://decosa.ai/prompts/device-mdr-triage-selfhost.md · https://decosa.ai/prompts/device-mdr-triage-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B answers the questions and drafts the narrative; the decision rule, the quote checks, the clocks and the file check are plain code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, messy complaint files.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Patient complaints stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it gives you: The 803.50(a) questions answered with quotes, a suggestion (reportable, not reportable, needs investigation) with the reasoning, the clock with its CFR rule and arithmetic, the 820.35(a) complaint-record check and missing 3500A items, a grounded draft event description, a Markdown memo, a signed triage record and a signed decision record naming the decider.
  - What it does not do: It does not decide or file: no eMDR submission, no event problem codes, no remedial-action or recall decision (21 CFR 806), no EU MDR vigilance. It never says a firm or file is compliant. It reads the text you send; scanned forms need OCR first.
  - Data retention: Nothing kept on the server. Complaints and records live in memory for the request; logs carry counts and decisions only. You keep the signed records in the MDR event file.
  - What leaves the box (hosted demo): The complaint and labeling go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for public MAUDE narratives and synthetic complaints only.
  - Model calls per triage: One facts read and 2 to 4 questions; with the narrative, one draft and one grounding call per sentence: about 3 to 11 calls on the demo complaints.
  - Typical run cost: A fraction of a cent at the gateway list price for the catheter complaint with the narrative; less for a cosmetic complaint that does not look reportable. Each run shows its own measured cost.
  - Clocks: 30 calendar days, 5 work days (FDA's written request, or remedial action from the day a supervisor knew), 10 work days for user facilities, and 30 days for supplements, from the day any employee had the information. Deadlines are not moved off weekends or holidays.
- Samples:
  - 1. Rep told earlier (`rep-told-earlier`): https://decosa.ai/tools/life-sciences/device-mdr-triage?sample=1&autorun=0
  - 2. Maude pod infection (`maude-pod-infection`): https://decosa.ai/tools/life-sciences/device-mdr-triage?sample=2&autorun=0
  - 3. Maude meter no start (`maude-meter-no-start`): https://decosa.ai/tools/life-sciences/device-mdr-triage?sample=3&autorun=0
  - 4. Maude pacemaker header (`maude-pacemaker-header`): https://decosa.ai/tools/life-sciences/device-mdr-triage?sample=4&autorun=0
  - 5. Cpap lid cosmetic (`cpap-lid-cosmetic`): https://decosa.ai/tools/life-sciences/device-mdr-triage?sample=5&autorun=0
  - 6. Meter reads high (`meter-reads-high`): https://decosa.ai/tools/life-sciences/device-mdr-triage?sample=6&autorun=0
- Regulatory note: Rules read on 26 Sep 2026 from eCFR (title 21 current to 24 Sep 2026). Manufacturers report deaths, serious injuries and malfunctions likely to cause or contribute to one if they recurred no later than 30 calendar days after they become aware (21 CFR 803.50(a)); awareness is when any employee has the information (803.3(b)(2)); 5-work-day reports when a reportable event needs remedial action to prevent an unreasonable risk of substantial harm, or FDA asks in writing (803.53); work days exclude federal holidays (803.3(y)); supplemental reports within 30 calendar days of new information (803.56); importers 30 calendar days and user facilities 10 work days (803.10). A qualified person (physician, nurse, risk manager or biomedical engineer) may conclude an event is not reportable, and the MDR event file keeps that information and the deliberations (803.20(c)(2), 803.18). The Quality Management System Regulation took effect on 2 Feb 2026: part 820 incorporates ISO 13485:2016 by reference, 820.35(a) lists what a complaint record must hold, and FDA now inspects under compliance programme 7382.850 instead of QSIT (FDA QMSR page, content current 2 Feb 2026). The malfunction factors, the two-year presumption and the start of the remedial-action clock are from FDA's guidance Medical Device Reporting for Manufacturers (8 Nov 2016), which is not binding. Complaints can hold patient health information: run it on your own hardware; the hosted demo takes public MAUDE narratives and synthetic complaints only. A triage aid, not a reportability determination or legal advice; it never says a firm or file is compliant and files nothing with FDA.

### Map evidence to NIST 800-171 (`cmmc-evidence-map`)

For small defence contractors and the consultants who prepare them for CMMC Level 2. Send the SSP's statements and the evidence behind them: policies, config exports, logs, records, scans, screenshots and verified test runs. For each NIST SP 800-171 Rev 2 requirement and each of its NIST SP 800-171A objectives it says evidence present, partial or missing, quotes the lines, and flags artefacts that cannot count: drafts, stale exports, systems outside the scope. Gaps come first. Out come a gap map, a draft POA&M marking what may never go on one, and a signed evidence index an assessor can check. An evidence-gap map, never a certification or a score.

- What it is: Each NIST SP 800-171 objective checked against the evidence you send: present, partial or missing, with quotes. A gap map, never a score.
- Who it's for: Small defence contractors preparing a CMMC Level 2 assessment, and the RPOs and MSPs who prepare them.
- Page: https://decosa.ai/tools/finance/cmmc-evidence-map
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/cmmc-evidence-map):
  - 237 / 286: Objective status, 3 classes (test split, n = 286)
  - 0 / 121: Partial or missing objectives called present (false present) (test split, n = 121)
  - 0 / 59: Requirements called fully evidenced that are not (test split, n = 59)
  - 149.5 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B reads the evidence (lines per objective, typed judgment, grounding) and transcribes screenshots; the catalog, point values, POA&M rules and artefact checks are plain code. Where: Self-host for real SSPs (CUI stays on your box); the hosted demo takes synthetic sets only and refuses marked CUI.
- Verification: Receipt per model call; every quote is a line of an artefact you sent; test-run certificates verified; signed hash-chained evidence index with no artefact text (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1× RTX PRO 6000 (96 GB) or 1× RTX 5090 (32 GB) for Qwen3.8-27B; the catalog, artefact checks, POA&M rules and signed record run on CPU
- Output: Per NIST SP 800-171A objective evidence present, partial or missing with quotes, a draft POA&M with 32 CFR 170.21 eligibility, a Markdown gap map and a signed decosa.record.v1 evidence index.
- Page: https://decosa.ai/tools/finance/cmmc-evidence-map · JSON: https://decosa.ai/use-cases/cmmc-evidence-map.json
- Prompts: https://decosa.ai/prompts/cmmc-evidence-map-hosted.md · https://decosa.ai/prompts/cmmc-evidence-map-selfhost.md · https://decosa.ai/prompts/cmmc-evidence-map-assemble.md · https://decosa.ai/prompts/cmmc-evidence-map-mac.md
- Tiers:
  - Lite · catalog and artefact checks only, no GPU (Any CPU): GET /cmmc/catalog gives every requirement's objectives, points and POA&M rule; POST /cmmc/check flags draft, stale, undated and out-of-scope artefacts and verifies test-run certificates. No reading of the evidence, so no statuses.
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B reads the evidence per requirement and objective (lines, typed judgment, grounding) and transcribes screenshots; the catalog, artefact checks, status rules, POA&M rules and signed index are code. This is what the hosted demo runs, on synthetic sets only.
- Alternates (smaller models for the same job):
  - Evidence binders as PDFs (document reader) (1x RTX PRO 6000 96 GB (estimate)), not built: Read SSPs and evidence binders as PDF or Word with page references in quotes. Waits on the document reader block (page 48); not built.
- Buyer facts:
  - Data retention: Nothing stored: the SSP and artefacts live in memory for the request. The signed index holds hashes, statuses and receipt ids, never artefact text; logs carry counts only.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and only synthetic sets are accepted (marked CUI is refused before any call). Self-hosted on the direct route: nothing leaves the box, and in real-data mode the service refuses to run if model calls would.
  - What it will not do: Say MET, certify, score or affirm. It maps evidence to objectives, lists the gaps first, and leaves MET to the assessor and the affirmation to the senior official.
  - POA&M rules: Every draft row carries its 32 CFR 170.21 eligibility: 1-point requirements only (3.13.11 conditionally), never 3.1.20, 3.1.22, 3.10.3, 3.10.4, 3.10.5 or 3.12.4.
  - Input formats: JSON: the system in scope (name, assessment date, scope list), SSP statements (or the SSP pasted as text), and up to 30 artefacts of up to 20,000 characters (120,000 in total), plus screenshots and Decosa test-run certificates. Up to 12 requirements per hosted run, 110 self-hosted.
  - Typical run: The clean sample: a few dozen model calls and a fraction of a cent at the gateway list price, from under a minute to a few minutes on the shared gateway; self-hosted, seconds. Each run shows its own measured cost.
- Samples:
  - 1. Harbor precision (`harbor-precision`): https://decosa.ai/tools/finance/cmmc-evidence-map?sample=1&autorun=0
  - 2. Bluefin clean (`bluefin-clean`): https://decosa.ai/tools/finance/cmmc-evidence-map?sample=2&autorun=0
  - 3. Tern valley pasted (`tern-valley-pasted`): https://decosa.ai/tools/finance/cmmc-evidence-map?sample=3&autorun=0
- Regulatory note: Checked 26 Sep 2026 against primary sources (links under Tools). 32 CFR Part 170 (the CMMC Program rule, 89 FR 83092, effective 16 Dec 2024) sets Level 2 as the 110 requirements of NIST SP 800-171 Rev 2, assessed per NIST SP 800-171A (June 2018), both incorporated by reference in 170.2; NIST's Rev 3 (May 2024) is not what CMMC assesses. 170.24 defines MET, NOT MET and N/A, requires evidence in final form (no drafts or unapproved policies), and sets the point values (5, 3, 1; 3.5.3 and 3.13.11 variable). 170.21 limits a POA&M: a score of at least 0.8 of 110, only 1-point requirements (3.13.11 at 3 points when encryption is employed but not FIPS-validated), never 3.1.20, 3.1.22, 3.10.3, 3.10.4, 3.10.5 or 3.12.4, closed out within 180 days. The DFARS rule (DFARS Case 2019-D041, 90 FR 43560) put CMMC into DoD contracts from 10 Nov 2025 (Phase 1). A pause of the move to Phase 2 (Level 2 C3PAO certification in new contracts from 10 Nov 2026) for a 60-day review from 13 Jul 2026 (DoD memo 26-P-1023) is reported by secondary sources; we did not find the memo itself, so that is unverified here, and we do not know the review's outcome. NIST's 800-171A CSV labels 3.10.1 as Personnel Security; the catalog takes the family from the section number. An SSP and its evidence are often CUI: self-host first; the hosted demo takes synthetic sets only and refuses marked CUI. This is not legal advice, an assessment or a certification: an assessor decides MET, and the senior official affirms in SPRS. Model licence: Apache-2.0 (Qwen3.8-27B). NIST publications and 32 CFR are US government works.

### Check a lay summary's numbers (`eu-trial-lay-summary`)

For sponsors' clinical disclosure teams, medical writers and CROs. Give it a trial with results on ClinicalTrials.gov and it drafts the lay summary the EU Clinical Trials Regulation requires (Article 37(4), content in Annex V), or checks the one you wrote. Every number is traced in code to a results cell, or to arithmetic on the cells the sentence cites, and a wrong one is flagged with the right figure. Comparisons are checked against the table, a result that could be chance has to say so, serious side effects and deaths must be given for every group, and promotional or softening words are flagged. Each claim goes through a grounding judge; Annex V's ten elements and the reading grade are reported. Out come a review file and a signed record of every check. It never says a summary is compliant: the sponsor's medical writer approves.

- What it is: Draft or check the EU lay summary of a trial's results, with every number traced to the results table.
- Who it's for: Clinical disclosure and transparency teams at sponsors, medical writers and CROs who write or review lay summaries for the EU database.
- Page: https://decosa.ai/tools/life-sciences/eu-trial-lay-summary
- Last reviewed: 2026-09-27
- Key numbers (details and caveats: https://decosa.ai/metrics/eu-trial-lay-summary):
  - 212 / 217 (97.7%): Planted errors caught after fixes, fresh trials (held out, n = 217)
  - 555 / 577 (96.2%): Planted errors caught, drafts with citations (code checks) (test split, n = 577)
  - 9 / 812 (1.1%): Correct sentences flagged (false flags), code checks (test split, n = 812)
  - 8.2 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B drafts the sections and judges each claim; number tracing, comparisons, hedges, side-effect checks, coverage and readability are plain code. Where: Hosted or self-host; self-host for results that are not public yet.
- Verification: Receipt per model call; every number traced to a results cell id; signed hash-chained record of the sources, the draft and every check (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: Qwen3.8-27B (1× RTX 5090 32 GB or larger); the number tracing, checks and record run on CPU
- Output: A lay summary in Annex V's order with every number traced to a results cell, the sentences to fix or check, Annex V coverage, the reading grade, a review file and a signed decosa.record.v1.
- Page: https://decosa.ai/tools/life-sciences/eu-trial-lay-summary · JSON: https://decosa.ai/use-cases/eu-trial-lay-summary.json
- Prompts: https://decosa.ai/prompts/eu-trial-lay-summary-hosted.md · https://decosa.ai/prompts/eu-trial-lay-summary-selfhost.md · https://decosa.ai/prompts/eu-trial-lay-summary-assemble.md · https://decosa.ai/prompts/eu-trial-lay-summary-mac.md
- Tiers:
  - Lite · check a draft in code, no GPU (Any CPU): POST /laysummary/check with grounding off: every number traced, comparisons, hedges, side effects, word lists, Annex V coverage and readability. No drafting and no claim-by-claim grounding.
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B drafts the six narrative sections and judges each sentence; every check and the record are code. This is what the hosted demo runs. Member-State versions: Hy-MT2-7B translates each sentence, and every number is checked again in each language.
- Buyer facts:
  - Data retention: Nothing stored: the study record, the draft and the result live in memory for the request, and logs carry counts only. A registry record fetched by NCT number is cached for a day. The signed record holds hashes, statuses and receipt ids, and the text of each sentence.
  - What leaves the box: Hosted: every model call goes through our gateway to the GPU serving Qwen3.8-27B, and an NCT number you type is sent to ClinicalTrials.gov. Self-hosted with DECOSA_LAYSUMMARY_FETCH=0 on the direct route: nothing leaves the box.
  - What it will not say: That a summary is compliant or meets Annex V. It says what it checked and marks what the registry does not hold (EU trial number, contact details, participants per country, follow-up plans) for the sponsor to add.
  - Input formats: A ClinicalTrials.gov NCT number or API v2 study record (CTIS and EudraCT tables have to be pasted in that shape), an optional protocol synopsis (20,000 characters), and for checking, a Markdown draft of up to 16,000 characters in English, German, French, Italian, Dutch or Spanish. Member-State versions: up to six of de, fr, es, it, nl, pl, pt, cs per request.
  - Typical run: A full draft: dozens of model calls, a few cents at the gateway list price, a minute or more on the shared gateway (faster self-hosted on the direct route). A short check with grounding: a few calls, a fraction of a cent. Member-State versions with the optional meaning check: two more calls per sentence per language, which adds minutes for many sentences in several languages on the busy gateway.
- Samples:
  - 1. Icodec weekly insulin (`icodec-weekly-insulin`): https://decosa.ai/tools/life-sciences/eu-trial-lay-summary?sample=1&autorun=0
  - 2. Ruxolitinib covid (`ruxolitinib-covid`): https://decosa.ai/tools/life-sciences/eu-trial-lay-summary?sample=2&autorun=0
  - 3. Delgocitinib teens (`delgocitinib-teens`): https://decosa.ai/tools/life-sciences/eu-trial-lay-summary?sample=3&autorun=0
  - 4. Icodec check (`icodec-check`): https://decosa.ai/tools/life-sciences/eu-trial-lay-summary?sample=4&autorun=0
- Regulatory note: Checked 26 Sep 2026 against the primary sources (links under Tools). Regulation (EU) No 536/2014, Article 37(4): within one year of the end of a trial in all Member States concerned, whatever the outcome, the sponsor submits a summary of the results to the EU database, accompanied by a summary understandable to laypersons whose content is set out in Annex V. Annex V lists ten elements (identification; sponsor name and contact details; where, when, objectives and reasons; the population, including numbers in the Member State concerned, the Union and third countries, age and gender breakdown and eligibility; the investigational medicinal products; adverse reactions and their frequency; overall results; comments on the outcome; whether follow-up trials are foreseen; where to find more). The 6-month deadline for paediatric trials is not in Article 37 itself: the Good Lay Summary Practice (adopted by the Clinical Trials Expert Group, 9 Jul 2021) gives 12 months, 6 for paediatric studies and up to 30 for non-therapeutic phase 1, citing the EU portal specifications (EMA/42176/2014). The expert group's recommendations (v2, 22 Feb 2018) ask for no promotional content, plain language and numeracy, explain that a non-significant difference should be explained to the reader, and call a 6th-grade Flesch-Kincaid level ideal. A cross-sectional study of the 7,547 phase II-IV trials registered in CTIS by November 2025 found that of the 234 legally required to report, 116 (49.6%) fully reported results on time (Bruckner et al., medRxiv preprint, 5 Apr 2026, not peer reviewed). Annex V asks for adverse reactions; the posted tables list adverse events whatever their cause, and the draft says so. This is a drafting and checking aid, not legal or regulatory advice, and it never states that a summary meets Annex V. Model licence: Apache-2.0 (Qwen3.8-27B). ClinicalTrials.gov records are cited by NCT number; results are facts reported by sponsors.

### Investigate a Reg E dispute (`reg-e-dispute-file`)

For bank, credit union and fintech disputes teams, and the compliance officers who answer for their files. Give it the consumer's notice of an electronic fund transfer error (a chat, call transcript, form or letter), the account's transactions as CSV, and the investigation notes, events and results letter. It computes every Regulation E clock in code with the rule cited: notice within 60 days of the statement, 10 or 20 business days, provisional credit and its notice, 45 or 90 days, the report within 3 business days and the correction within 1. It quotes the notice's required elements, lists the file's gaps, and checks that a no-error letter explains the findings for this case and notes the right to request the documents relied on. It never decides a claim: a named investigator does, and the signed record keeps their decision as entered.

- What it is: A Reg E dispute investigation file for disputes teams: every error-resolution clock computed from the dates with its CFR rule, the notice's required elements quoted, the file's gaps listed, and the no-error letter checked. The investigator decides every claim.
- Who it's for: Disputes operations teams and compliance officers at banks, credit unions, fintechs and remittance providers.
- Page: https://decosa.ai/tools/finance/reg-e-dispute-file
- Last reviewed: 2026-09-26
- Key result: On a held-out set of 40 synthetic notices it flagged 43 of 45 missing elements with one false flag, and caught every generic letter in 20 templated tests; all synthetic, written by the same author.
- Key numbers (details and caveats: https://decosa.ai/metrics/reg-e-dispute-file):
  - 0.956 (43/45): Missing notice elements flagged (recall) (test split, n = 45)
  - 0.977 (43/44): Missing-element flags that were right (precision) (test split, n = 44)
  - 37 / 40: Notices with all five elements right (test split, n = 40)
  - 37.8 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real accounts (hosted demo: synthetic disputes only).
- Verification: Receipt per model call; every clock computed in code with its CFR rule; every quote found word for word; signed hash-chained investigation record (receipt coverage: full). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the clocks, the checklist and the signed record need no GPU
- Output: The file status, every clock with its rule and a timeline, the notice's elements with quotes, the checklist of gaps, the letter check, a Markdown file and a signed record with the investigator's determination as entered.
- Page: https://decosa.ai/tools/finance/reg-e-dispute-file · JSON: https://decosa.ai/use-cases/reg-e-dispute-file.json
- Prompts: https://decosa.ai/prompts/reg-e-dispute-file-hosted.md · https://decosa.ai/prompts/reg-e-dispute-file-selfhost.md · https://decosa.ai/prompts/reg-e-dispute-file-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. The clocks are code and identical; the reading accuracy on this task is unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B reads the notice, the notes and the letter; the clocks, the quote checks, the checklist and the file status are plain code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long call transcripts and dense case notes.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Dispute files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it gives you: Every Reg E clock with its rule, start, due date and arithmetic; a timeline; the notice's required elements with quotes; a checklist of the file's gaps; a check of the results letter; coverage questions with the rule text; a Markdown file and a signed record.
  - What it does not do: It never decides a claim, denies one or drafts a denial letter: a named investigator does, and the file keeps their determination as entered. It does not cover Regulation Z credit-card disputes, card-network chargeback rules or state law, and it does not read your core system: send it the CSV.
  - Data retention: Nothing kept on the server. Notices, account data and files live in memory for the request; logs carry counts only. You keep the signed record with the dispute, at least two years (12 CFR 1005.13(b)).
  - What leaves the box (hosted demo): The notice, notes, letter and the disputed transfers go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for synthetic disputes only.
  - Model calls per file: One for the notice, one for the notes, one for the letter and one grounding call per letter sentence: 1 to 14 calls on the demo files. The clocks are code.
  - Typical run cost: About a cent or less at the gateway list price for the card-not-present file; a little more for the P2P scam file with a longer letter. Each run shows its own measured cost.
  - Business days: Weekdays minus the Federal Reserve Banks' holidays (a Saturday holiday is not moved). Reg E counts your own business days, so list any other day you are closed.
- Samples:
  - 1. Cnp generic denial (`cnp-generic-denial`): https://decosa.ai/tools/finance/reg-e-dispute-file?sample=1&autorun=0
  - 2. Atm short cash new account (`atm-short-cash-new-account`): https://decosa.ai/tools/finance/reg-e-dispute-file?sample=2&autorun=0
  - 3. P2p scam consumer sent (`p2p-scam-consumer-sent`): https://decosa.ai/tools/finance/reg-e-dispute-file?sample=3&autorun=0
  - 4. P2p takeover open (`p2p-takeover-open`): https://decosa.ai/tools/finance/reg-e-dispute-file?sample=4&autorun=0
  - 5. Remittance not received (`remittance-not-received`): https://decosa.ai/tools/finance/reg-e-dispute-file?sample=5&autorun=0
- Regulatory note: Consumer notices and account data are nonpublic personal information (Gramm-Leach-Bliley): run it on the institution's own hardware; the hosted demo takes synthetic disputes only. Rules read on 26 Sep 2026 from eCFR (the version of 1 Sep 2026; 12 CFR 1005.11 last amended 13 Feb 2018): notice within 60 days after the statement first showing the error is sent (1005.11(b)(1)(i)); determine within 10 business days (1005.11(c)(1)) or, with provisional credit within 10 business days and notice of it within 2, within 45 days (1005.11(c)(2)); 20 business days for a transfer within 30 days after the first deposit (1005.11(c)(3)(i)) and 90 days for a POS debit card transaction (including mail, phone and online orders, not ATMs; comment 11(c)(3)-1), a transfer not initiated in a state, or a new account (1005.11(c)(3)(ii)); report within 3 business days and correct within 1 (1005.11(c)(1), (c)(2)(iii)-(iv)); for no error or a different error, a written explanation and the right to request the documents relied on (1005.11(d)(1)); prepaid accounts without statements, 120 days from posting (1005.18(e)(2)(ii)); remittances, 180 days' notice and 90 days to decide, with an explanation that addresses the sender's complaint (1005.33(b), (c), (d)); records kept 2 years (1005.13(b)). A transfer the consumer was tricked into sending is initiated by the consumer, which the definition of an unauthorized transfer (1005.2(m)) does not cover; transfers made with access information obtained from the consumer by fraud are unauthorized (comment 2(m)-3; CFPB EFT FAQs, updated 13 Dec 2021 and online on 26 Sep 2026). The CFPB's December 2024 suit over Zelle fraud was dismissed with prejudice on 5 Mar 2025. The tool shows these as questions with the rule text; the investigator decides. Business days default to the Federal Reserve Banks' holidays (Board schedule K.8); Reg E counts the institution's own (1005.2(d)). Not covered: Regulation Z credit-card disputes, card-network rules, state law. Not legal advice.

### Stage a listing photo with disclosure (`virtual-staging`)

Stages an empty-room listing photo with open image models, adding movable furniture only, then checks that the walls, windows, doors, views, floors and visible damage were left alone, pixel by pixel and with a side-by-side model comparison. Requests to remove power lines, hide cracks or make a room look bigger are refused. It also checks images staged anywhere else. A photo that passes gets a burned-in 'Virtually staged' label linking to the unaltered original on a public page, a C2PA credential naming the original as its parent, and a signed pair record: the disclosure California's AB 723 has asked for since 1 Jan 2026. Not legal advice.

- What it is: Stage an empty listing photo with furniture only, check that walls, windows, doors, views, floors and damage were left alone, and ship the AB 723 disclosure: a burned-in label, the original on a public page, a C2PA credential and a signed pair record.
- Who it's for: Listing agents, brokerages, listing photographers and MLS vendors publishing staged photos of California homes.
- Page: https://decosa.ai/tools/operations/virtual-staging
- Last reviewed: 2026-09-26
- Key numbers (details and caveats: https://decosa.ai/metrics/virtual-staging):
  - 38 / 40: Planted edits caught (check) (test split, n = 40)
  - 0 / 10: False alarms on clean stagings (test split, n = 10)
  - 38 / 40: Planted edits caught, pixel check alone (test split, n = 40)
  - 15.2 s: Median end-to-end run, hosted (QA sweep 2026-09-26)

- Status: live. Models: Wan2.2-VACE-Fun-A14B (furniture inpainting) · Qwen3.8-27B (structure map, furniture boxes, side-by-side check). Where: Hosted or self-host.
- Verification: Receipt per model call; C2PA credential with the original as parent; signed pair record binding both images, the check and the label (receipt coverage: partial). Hosted QA 2026-09-26: pass. Self-host QA 2026-09-26: pass.
- Hardware: Wan2.2-VACE-Fun-A14B (fp8) and Qwen3.8-27B: one 96 GB GPU (measured on two shared ones); the check, label, credential and record run on CPU
- Output: A staged photo with a burned-in 'Virtually staged' label, the original on a public page, a C2PA credential naming the original, and a signed pair record.
- Page: https://decosa.ai/tools/operations/virtual-staging · JSON: https://decosa.ai/use-cases/virtual-staging.json
- Prompts: https://decosa.ai/prompts/virtual-staging-hosted.md · https://decosa.ai/prompts/virtual-staging-selfhost.md · https://decosa.ai/prompts/virtual-staging-assemble.md
- Tiers:
  - Lite · check only (any staging tool) (1x RTX 5090 32 GB for the model (not measured) or the hosted gateway): No render: check an image staged elsewhere against its original, then label, credential and pair it when it passes. Two image calls per check.
  - Standard · stage and check (hosted demo) (Qwen3.8-27B and VACE A14B on two 96 GB cards (measured, shared)): Furniture drawn on the open floor by Wan2.2-VACE-Fun-A14B, only the furniture kept, the same check; a render the check rejects is drawn again (up to 3 seeds) and never labelled.
- Alternates (smaller models for the same job):
  - instruction image editing (not measured), not served: Qwen-Image-Edit-2511 (Apache-2.0) edits by instruction and keeps the rest of the photo by design; not installed here.
- Buyer facts:
  - What it checks: Whether a staged photo changes the property: windows painted over or hidden, doors, the view through a window, wall colour, floor patterns painted out, fixtures, and visible damage hidden or covered, plus framing. Refuses instructions to remove, hide or improve anything.
  - What it does not do: It is not legal advice or a compliance determination, and it can miss things (2 of 40 planted edits in the held-out test, a redrawn floor patch). A person should look at the pair before it is published. It does not check MLS or portal rules.
  - Data retention: Uploads live in a temporary folder for the run and are deleted when it ends. A disclosed pair (the original, the staged image and the record) is kept for 30 days on a public page, because the label links to it: download the pair to host it on your own listing site. Runs are kept in memory for one hour for the token or key that made them. Logs carry ids, verdicts and counts, never images or instructions.
  - What leaves the box (hosted demo): The photos go to Qwen3.8-27B through our gateway and to the studio GPU, both operated by Decosa. The disclosed pair is public by design. Self-hosted, nothing leaves the machine unless you publish the pair.
  - Cost per image: A fraction of a cent to check an image staged elsewhere (two image calls) and a little more plus a short GPU render to stage one, at the gateway list price. Staging services typically charge tens of dollars per image.
  - Output: A JPEG with 'Virtually staged' and the original's link burned in, a C2PA credential naming the original as its parent, a public page with the unaltered original, and a signed pair record (verify at /record/verify).
- Samples:
  - 1. House living (`house-living`): https://decosa.ai/tools/operations/virtual-staging?sample=1&autorun=0
  - 2. Berlin living (`berlin-living`): https://decosa.ai/tools/operations/virtual-staging?sample=2&autorun=0
  - 3. Old room (`old-room`): https://decosa.ai/tools/operations/virtual-staging?sample=3&autorun=0
  - 4. Room marble (`room-marble`): https://decosa.ai/tools/operations/virtual-staging?sample=4&autorun=0
  - 5. Check view replaced (`check-view-replaced`): https://decosa.ai/tools/operations/virtual-staging?sample=5&autorun=0
  - 6. Check clean (`check-clean`): https://decosa.ai/tools/operations/virtual-staging?sample=6&autorun=0
  - 7. Refuse power lines (`refuse-power-lines`): https://decosa.ai/tools/operations/virtual-staging?sample=7&autorun=0
  - 8. Refuse bigger (`refuse-bigger`): https://decosa.ai/tools/operations/virtual-staging?sample=8&autorun=0
- Regulatory note: Not legal advice or a compliance determination; checked against the primary text on 26 Sep 2026. California Business and Professions Code section 10140.8, added by AB 723 (Pellerin; Stats. 2025, ch. 497; approved and filed 10 Oct 2025; as a regular statute with no date of its own it took effect 1 Jan 2026, Cal. Const. art. IV, sec. 8(c)): a real estate broker or salesperson, or a person acting for them, who includes a digitally altered image in an advertisement or promotional material for the sale of real property must include a reasonably conspicuous statement, on or adjacent to the image, that it has been altered, and a link, URL or QR code to a publicly accessible site that includes and clearly identifies the original, unaltered image; on a website they control, the unaltered image must be in the posting or linked. 'Digitally altered' covers adding, removing or changing elements with photo-editing software or AI, including furniture, fixtures, flooring, walls, paint colour, landscape and elements visible from the property such as utility poles, views through windows and neighbouring properties; lighting, white balance, cropping, exposure and similar adjustments that do not change the representation of the property are excluded. A willful violation of the Real Estate Law is a crime. MLSs and portals may have their own rules (not checked here). Model licences: Apache-2.0 (Qwen3.8-27B, Wan2.2-VACE-Fun-A14B, Wan2.2-Lightning).

### Write a due-diligence red-flag memo (`ma-dd-redflags`)

For deal counsel and corporate development teams. Give it a data room: contracts, board minutes, the disclosure schedule, revenue by customer, financial summaries. For ten red-flag categories (change of control, anti-assignment, exclusivity and non-compete, MFN pricing, uncapped liability, debt covenants, IP assignment gaps, litigation missing from the disclosure schedule, customer concentration, key person), a reranked search pulls the passages that could hold one and an open model says which do, quoting the exact words. Code checks every quote word for word against the passage and byte span it cites, and a grounding check reads each finding against that passage. You get a red-flag memo with what to ask the seller, and a signed record. A triage aid for counsel, not legal advice.

- What it is: A data room in, a red-flag memo for counsel out: every flag quotes the exact words and the byte span it came from.
- Who it's for: Deal counsel, corporate development teams and private-equity deal teams doing buy-side diligence.
- Page: https://decosa.ai/legal/ma-dd-redflags
- Last reviewed: 2026-09-27
- Key numbers (details and caveats: https://decosa.ai/metrics/ma-dd-redflags):
  - 9 / 12: CUAD anti-assignment clauses found (held out, n = 12)
  - 3 / 7: CUAD exclusivity and non-compete clauses found (held out, n = 7)
  - 20 / 20: Planted red flags caught (test split, n = 20)
  - 92.9 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B (one call per category, grounding of each finding) · Qwen3-Embedding + Qwen3-Reranker (the evidence retrieval block). Where: Self-host for a real data room; the hosted demo takes invented rooms only.
- Verification: Receipt per model call; signed search receipts with the index hash; every quote found word for word at the cited byte span; findings grounded; signed hash-chained record (receipt coverage: full). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B plus about 10 GB for the embedder and reranker; the quote checks, memo and record run on CPU
- Output: A red-flag memo for counsel (flag, severity, document, verbatim quote with byte span, why it matters, what to ask) and a signed decosa.record.v1 with the retrieval index hash.
- Page: https://decosa.ai/legal/ma-dd-redflags · JSON: https://decosa.ai/use-cases/ma-dd-redflags.json
- Prompts: https://decosa.ai/prompts/ma-dd-redflags-hosted.md · https://decosa.ai/prompts/ma-dd-redflags-selfhost.md · https://decosa.ai/prompts/ma-dd-redflags-assemble.md
- Tiers:
  - Lite · BM25 retrieval, no retrieval service (1x GPU for Qwen3.8-27B): The same review with DECOSA_MADD_RETRIEVAL=bm25: keyword search picks the passages, no embedder or reranker, so only the language model needs a GPU. Same quote gate, grounding and record.
  - Standard · retrieval service + the model (hosted demo) (1x RTX PRO 6000 96 GB (measured on shared cards)): The evidence retrieval block (Qwen3-Embedding-0.6B + Qwen3-Reranker-4B) picks the passages per category, Qwen3.8-27B names the flags, code checks the quotes and the grounding block checks the findings. This is what the hosted demo runs, on invented rooms only.
- Alternates (smaller models for the same job):
  - Scanned data rooms (document reader) (1x RTX PRO 6000 96 GB (estimate)), not built: Read scanned PDFs into text with page and box cites first, then review. The document reader block exists; the two are not wired together here yet.
- Buyer facts:
  - Data retention: Nothing stored: the documents live in memory for the request. The signed record holds hashes (documents, index, quotes) and receipt ids, never document text; logs carry counts only.
  - What leaves the box: Self-hosted on the direct route: nothing. The model, the retrieval service and the checks run on the same machine. Hosted: model calls go through the Decosa API, and only invented rooms are accepted.
  - What it checks: Ten categories: change of control, anti-assignment, exclusivity and non-compete, MFN pricing, uncapped liability, debt covenants, IP assignment gaps, litigation missing from the disclosure schedule, customer concentration, key person.
  - What it is not: Legal advice or a materiality call. A triage aid: every flag quotes the data room so counsel can confirm it in seconds.
- Samples:
  - 1. Kestrel (`kestrel`): https://decosa.ai/legal/ma-dd-redflags?sample=1&autorun=0
  - 2. Linnet (`linnet`): https://decosa.ai/legal/ma-dd-redflags?sample=2&autorun=0
- Regulatory note: Not a regulated activity in itself, but the output feeds legal advice, so a lawyer must review it. Written 27 Sep 2026. ABA Formal Opinion 512, Generative Artificial Intelligence Tools (29 Jul 2024; https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf) reads the duties of competence (Model Rule 1.1), confidentiality (1.6) and supervision (5.1, 5.3) as requiring lawyers to understand a tool's limits, check its output and protect client information put into it (unverified: the ABA site blocked our fetch on 27 Sep 2026, so the date and summary are from the published opinion as we recall it; read it before relying on this). This tool supports those duties by quoting its sources word for word and running on the firm's own box; it does not replace the lawyer's review. Data rooms are usually under NDA: check the NDA before sending any document to a hosted service, which is why the hosted demo takes invented rooms only.

### Draft a tariff classification memo (`tariff-classification`)

For customs brokers and importers' trade-compliance teams. Describe a product. It searches public CBP rulings on similar products and the HTS text of the candidate headings, and an open model proposes a heading, subheading and statistical number with the General Rules of Interpretation it applied. Code checks that every code is printed in the current HTS release and that every quote from a ruling is found word for word in the text it retrieved. When the description leaves out a fact the code turns on, or the evidence is thin, it says so and sends it to a broker. You get a memo and a signed record naming the ruling set and the index it searched. A research aid: a licensed broker decides.

- What it is: A research aid for HTS classification: describe a product and it finds CBP rulings on similar products, proposes a heading, subheading and statistical number with the GRI applied, and quotes the rulings word for word. A licensed customs broker decides.
- Who it's for: Licensed customs brokers and importers' trade-compliance teams who write classification memos for new products.
- Page: https://decosa.ai/tools/finance/tariff-classification
- Last reviewed: 2026-09-27
- Key result: On 200 held-out CBP rulings, the memo's top-1 subheading matched CBP's 145 / 200 times, against 57 / 200 for the same model without the rulings; queries were CBP's own descriptions.
- Key numbers (details and caveats: https://decosa.ai/metrics/tariff-classification):
  - 145 / 200: Memo top-1 subheading (6-digit) (test split, n = 200)
  - 171 / 200: Memo top-1 heading (4-digit) (test split, n = 200)
  - 172 / 200: Memo top-3 subheading (test split, n = 200)
  - 32.1 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B (the proposal); Qwen3 embedder and reranker (the search over rulings). Where: Hosted or self-host.
- Verification: Receipt per model call; signed search receipts with the ruling-set index hash; every ruling quote verified word for word with its byte offsets; codes checked against the HTS release; signed hash-chained memo record (receipt coverage: full). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the embedder and reranker fit in about 10 GB beside it or on a second card; the checks and signing run on CPU
- Output: A classification memo in Markdown for a broker (proposed code or 'needs a broker', GRI reasoning, verified quotes, alternatives) and a signed memo record.
- Page: https://decosa.ai/tools/finance/tariff-classification · JSON: https://decosa.ai/use-cases/tariff-classification.json
- Prompts: https://decosa.ai/prompts/tariff-classification-hosted.md · https://decosa.ai/prompts/tariff-classification-selfhost.md · https://decosa.ai/prompts/tariff-classification-assemble.md
- Tiers:
  - Lite · smaller reranker (1x 80-96 GB card): Qwen3-Reranker-0.6B instead of the 4B: about 7 GB less GPU memory. Not measured on this tool.
  - Standard · the hosted demo (1x RTX PRO 6000 Blackwell 96 GB (or the LLM and retrieval on two cards)): Hybrid search with the 4B reranker, one Qwen3.8-27B call, the checks and the signed record.
- Buyer facts:
  - What it gives you: A memo: the rulings found (with the matching passages), a proposed heading, subheading and statistical number with the GRI applied and verified quotes, up to two alternatives, the facts a broker should confirm, and a signed record naming the ruling set and the index searched.
  - Coverage: It covers CBP New York rulings in 35 headings of 11 chapters. Chapters: 39, 42, 61, 62, 63, 64, 73, 84, 85, 94, 95. Headings: 3923, 3924, 3926, 4202, 6104, 6109, 6110, 6114, 6204, 6211, 6307, 6402, 6403, 6404, 7323, 7326, 8414, 8419, 8471, 8479, 8481, 8504, 8516, 8517, 8518, 8528, 8543, 8544, 9401, 9403, 9405, 9503, 9504, 9505, 9506. A product outside them comes back as "out of coverage", naming the heading it would need; a licensed customs broker can classify it. Inside them, a description that lacks a deciding fact comes back as "needs a broker".
  - What it does not do: It does not decide, file or rule: a licensed customs broker or the importer of record decides, and only CBP issues binding rulings. It searches 1,992 New York rulings in 35 headings, not all of CROSS, and not HQ rulings, court cases or the Explanatory Notes. It does not apply Chapter 99 duties or decide origin or value.
  - Data retention: Nothing stored: the description lives in memory for the request. Logs carry counts and timings only.
  - Evidence it keeps: The record commits to the description's hash, the ruling-set hash, the HTS release, the retrieval index hash (a Merkle root over every chunk), the search and model receipts, the proposal and every check.
- Samples:
  - 1. Insulated lunch bag (`insulated-lunch-bag`): https://decosa.ai/tools/finance/tariff-classification?sample=1&autorun=0
  - 2. Knit pullover (`knit-pullover`): https://decosa.ai/tools/finance/tariff-classification?sample=2&autorun=0
  - 3. Wall heater (`wall-heater`): https://decosa.ai/tools/finance/tariff-classification?sample=3&autorun=0
  - 4. Storage bin (`storage-bin`): https://decosa.ai/tools/finance/tariff-classification?sample=4&autorun=0
  - 5. Vague bag (`vague-bag`): https://decosa.ai/tools/finance/tariff-classification?sample=5&autorun=0
- Regulatory note: Checked 27 Sep 2026. Importers must use reasonable care to classify (19 U.S.C. 1484, https://www.law.cornell.edu/uscode/text/19/1484); binding rulings come only from CBP under 19 CFR Part 177 (https://www.ecfr.gov/current/title-19/chapter-I/part-177), and a ruling binds only the transaction it describes. Customs business, including classification for others, is for licensed customs brokers (19 U.S.C. 1641, https://www.law.cornell.edu/uscode/text/19/1641). The General Rules of Interpretation are quoted from the HTS (USITC, 2026 Revision 19, https://hts.usitc.gov). The memo is not a ruling, not advice, and not a filing; Chapter 99 duties (Section 301, 232, IEEPA) are not applied.

### Check CSR numbers against tables (`csr-number-verifier`)

For medical writing and regulatory operations teams at sponsors and CROs. Give it a clinical study report: the narrative of sections 10 to 12 with its in-text tables and TLFs, as text and rows or as a PDF (born-digital or scanned, read by the document reader). Every number in the narrative is traced to the table cell it reports: an open model points each number at the cell for the right arm, row and timepoint, and code compares the value at the printed decimals and recomputes percentages, differences, totals and relative reductions. It flags numbers that don't match their cell, numbers with no source, and derived numbers that don't compute, with the likely slip (the other arm's value, a stale value from an earlier data cut, transposed digits, rounding, a wrong denominator). In-text tables are checked against their source TLF. You get a QC report and a signed record. A medical-writing QC aid: the medical writer and the QC reviewer decide.

- What it is: Every number in a clinical study report traced to the table cell it reports, checked in code, with the slips flagged.
- Who it's for: Medical writing and regulatory operations teams at sponsors and CROs, and medical-writing QC contractors.
- Page: https://decosa.ai/tools/life-sciences/csr-number-verifier
- Last reviewed: 2026-09-27
- Key numbers (details and caveats: https://decosa.ai/metrics/csr-number-verifier):
  - 170 / 178: Planted number errors caught, ClinicalTrials.gov tables (held out, n = 178)
  - 92 / 96: Planted number errors caught (test split, n = 96)
  - 93 / 96: Planted number errors caught, second writer (test split, n = 96)
  - 6.5 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B (points each number at its cell) · Qwen3-Embedding + Qwen3-Reranker (finding the right table in a long report) · PaddleOCR-VL + Docling layout (reading PDFs and scans). Where: Self-host for a real report; the hosted demo takes invented or public reports only.
- Verification: Receipt per model call; signed search receipts with the index hash; every comparison and recomputation done in code; signed hash-chained record of each number's cell and verdict (receipt coverage: full). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B, plus about 10 GB for the embedder and reranker and about 6 GB for the document reader; the comparisons, report and record run on CPU
- Output: A number-QC report in Markdown (each flag with its section, cell, expected value and likely slip; in-text table mismatches) and a signed decosa.record.v1 of every number's cell and verdict.
- Page: https://decosa.ai/tools/life-sciences/csr-number-verifier · JSON: https://decosa.ai/use-cases/csr-number-verifier.json
- Prompts: https://decosa.ai/prompts/csr-number-verifier-hosted.md · https://decosa.ai/prompts/csr-number-verifier-selfhost.md · https://decosa.ai/prompts/csr-number-verifier-assemble.md
- Tiers:
  - Lite · text and rows, keyword search (1x GPU for Qwen3.8-27B): Send the narrative as text and the tables as rows (exported from the TLF outputs), and run with DECOSA_CSR_RETRIEVAL=bm25: no retrieval service and no document reader, so only the language model needs a GPU. Same pointer, checks and record.
  - Standard · retrieval + document reader + the model (hosted demo) (1x RTX PRO 6000 96 GB (measured on shared cards)): The retrieval block picks the tables per paragraph, the document reader reads PDFs and scans, Qwen3.8-27B points each number at its cell and code checks everything. This is what the hosted demo runs, on invented or public reports only.
- Alternates (smaller models for the same job):
  - Number-consistency checker tier (CPU) (CPU, in process), not built: For numbers the pointer leaves untraced: a small encoder trained on planted number errors reads the sentence against the table; it can mark a number supported but never overrules a code mismatch. The hook is built; the checker is on an unmerged branch and not measured with this tool.
- Buyer facts:
  - Data retention: Nothing stored: the report lives in memory for the request. The signed record holds each number's status, cell and expected value, table hashes and receipt ids, never the report text; logs carry counts and timings only.
  - What leaves the box: Self-hosted on the direct route: nothing. The model, the retrieval service, the document reader and the checks run on the same machine. Hosted: model calls go through the Decosa API, and only invented or public reports are accepted.
  - What it checks: Every number in the narrative (counts, Ns, percentages, means, differences, CIs, hazard ratios, p-values) against the cell it reports, derived numbers recomputed, in-text tables against their source TLF, and the likely slip for each mismatch.
  - What it is not: A validated (GxP) system or a regulatory submission tool. It does not check the TLFs themselves, figures, listings or interpretation. The medical writer and the QC reviewer decide.
- Samples:
  - 1. A fictional zenavotide phase 3 report with planted errors (sections 10 to 12, 11 tables) (`zenavotide-301`): https://decosa.ai/tools/life-sciences/csr-number-verifier?sample=1&autorun=0
  - 2. The same fictional report with no errors (`zenavotide-301-clean`): https://decosa.ai/tools/life-sciences/csr-number-verifier?sample=2&autorun=0
  - 3. The report with planted errors, as a PDF made from the text (`zenavotide-301-pdf`): https://decosa.ai/tools/life-sciences/csr-number-verifier?sample=3&autorun=0
  - 4. The report with planted errors, as a 12-page scan (`zenavotide-301-scan`): https://decosa.ai/tools/life-sciences/csr-number-verifier?sample=4&autorun=0
- Regulatory note: No regulation requires this check; it supports the manual number QC that sponsors and CROs already run on clinical study reports. Written 27 Sep 2026. ICH E3 (Structure and Content of Clinical Study Reports) defines the CSR's sections, including the efficacy (11) and safety (12) evaluations and the tables that support them (not re-read for this note). The FDA's draft guidance 'Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products' (January 2025; https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological) is still a draft and is about AI used to produce information supporting regulatory decisions; a QC aid that only flags numbers for a person to check is not that, but a sponsor's own quality system decides how to qualify any tool it uses. The EMA reflection paper on the use of AI in the medicinal product lifecycle (adopted 30 Sep 2024; https://www.ema.europa.eu/en/use-artificial-intelligence-ai-medicinal-product-lifecycle-scientific-guideline) asks sponsors to take a risk-based approach and keep a human accountable. This tool is not a validated (GxP) system and not a regulatory submission tool: the medical writer and the QC reviewer decide what to change, and the TLFs remain the biostatistician's responsibility. Dates and scope from the Decosa wiki's sourced notes (page 33, S35 and S36, checked 26 Sep 2026); read the documents before relying on this.

### Build an EU GPSR listing (`gpsr-listing-pack`)

Turns a supplier's product sheet and the product label (text or a photo) into the listing information Article 19 of the EU General Product Safety Regulation (EU) 2023/988 asks every online offer to show: manufacturer, EU responsible person when the maker is outside the Union, product identification, and the warnings in the buyer's language. Every value is looked up in the documents, each Article 19 element is marked found or missing, a label that disagrees with the sheet is flagged, and the warnings come back in up to five languages with every number, unit, negation and name checked. A preparation aid with a signed record, never a statement that an offer meets the Regulation.

- What it is: The GPSR listing pack prepares the GPSR listing information Article 19 of Regulation (EU) 2023/988 asks an online offer to show, from a supplier's sheet and the label, and checks it: every value found in your documents, each element found or missing, the label compared with the sheet, the warnings translated with their numbers checked.
- Who it's for: Marketplace sellers shipping to the EU and responsible-person services that check their clients' listings and labels.
- Page: https://decosa.ai/tools/finance/gpsr-listing-pack
- Last reviewed: 2026-09-30
- Key result: On 24 held-out synthetic products it found 15 of 19 missing elements with no false alarm (18 of 18 after one fix) and all 5 label-vs-sheet mismatches; translations caught 98.9% of planted number errors.
- Key numbers (details and caveats: https://decosa.ai/metrics/gpsr-listing-pack):
  - 15 / 19: Missing Article 19 elements found, frozen (test split, n = 19)
  - 18 / 18: Missing elements found after that fix (test split, n = 18)
  - 5 / 5: Label-vs-sheet mismatches found (test split, n = 5)
  - 12.0 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: live. Models: Qwen3.8-27B (label reading, field extraction; translation into 15 EU languages) · Hy-MT2-7B (translation into 8, language-pack block). Where: Hosted or self-host.
- Verification: Gateway receipts for Qwen3.8 calls; model-call attestation per translation call; signed record of the fields, elements, flags and translations (receipt coverage: partial). Hosted QA 2026-09-30: pass.
- Hardware: 1x RTX PRO 6000 (96 GB): Qwen3.8-27B NVFP4 plus Hy-MT2-7B (about 18 GB); the checks run on CPU
- Output: The listing fields per language (CSV/JSON), the Article 19 checklist with articles, the label-vs-sheet flags, the checked translations and a signed record.
- Page: https://decosa.ai/tools/finance/gpsr-listing-pack · JSON: https://decosa.ai/use-cases/gpsr-listing-pack.json
- Prompts: https://decosa.ai/prompts/gpsr-listing-pack-hosted.md · https://decosa.ai/prompts/gpsr-listing-pack-selfhost.md · https://decosa.ai/prompts/gpsr-listing-pack-assemble.md
- Tiers:
  - Standard · Qwen3.8-27B plus Hy-MT2-7B (hosted demo) (2x RTX PRO 6000 96 GB shared (measured); 1x 96 GB should fit (not measured)): What the hosted demo runs: Qwen3.8-27B extracts, Hy-MT2-7B translates, code checks and signs.
  - Best · the 30B translation model (not measured) (1x RTX PRO 6000 96 GB (estimate: 20 GB Qwen3.8 NVFP4 weights plus 31 GB MT)): Hy-MT2-30B-A3B-FP8 for translation; not run by us.
- Buyer facts:
  - Data retention: Nothing stored: the sheet, the label and the result live in memory for the request, and logs carry counts only. The signed record holds hashes, statuses, flags, the extracted fields and the translations.
  - What leaves the box: Hosted: the Qwen3.8 calls go through our gateway to the GPU serving the model; translation runs on Decosa's hosted service, and the meaning check sends each English line with its translation and back-translation to Qwen3.8 through the gateway. Self-hosted on the direct route: nothing leaves the box.
  - What it will not say: That an offer or a product complies with Regulation (EU) 2023/988. It lists the Article 19 elements found or missing and what disagrees; Decosa is not a responsible person.
  - Languages: Any EU official language as a target (23 besides English, up to five per pack). Thirteen of the fifteen added on 27 Sep met the language pack's FLORES bar; Irish and Maltese are returned marked "draft, needs a reviewer".
  - Typical run: One Qwen3.8-27B call to read the sheet (two with a label photo), a few translation calls per language, and for every translated line a back-translation and one Qwen3.8 judgment (the meaning check, on by default). The meaning check is most of the time and model cost of a run; the measured figures are shown on this page. Send "meaning": false to skip it: numbers, units, negations and names are still checked in code.
- Samples:
  - 1. Toy with rp (`toy-with-rp`): https://decosa.ai/tools/finance/gpsr-listing-pack?sample=1&autorun=0
  - 2. Charger no rp (`charger-no-rp`): https://decosa.ai/tools/finance/gpsr-listing-pack?sample=2&autorun=0
  - 3. Kettle label mismatch (`kettle-label-mismatch`): https://decosa.ai/tools/finance/gpsr-listing-pack?sample=3&autorun=0
  - 4. Candle label photo (`candle-label-photo`): https://decosa.ai/tools/finance/gpsr-listing-pack?sample=4&autorun=0
- Regulatory note: Checked 27 Sep 2026 against the text of Regulation (EU) 2023/988 (GPSR), OJ L 135, 23.5.2023, p. 1 (CELEX 32023R0988, read from the EU Publications Office). It applies from 13 December 2024 (Article 52). Article 19: an offer made online or at a distance must clearly and visibly indicate (a) the manufacturer's name, registered trade name or trade mark and postal and electronic address; (b) where the manufacturer is not established in the Union, the name, postal and electronic address of the responsible person under Article 16(1) or Article 4(1) of Regulation (EU) 2019/1020; (c) information allowing the product to be identified, including a picture, its type and any other product identifier; and (d) any warning or safety information to be affixed to the product or packaging or included in an accompanying document, in a language consumers can easily understand, as determined by the Member State where the product is made available. Article 9(5)-(7) sets the manufacturer's labelling (type, batch or serial number; name and addresses; instructions and safety information in the consumer's language), and Article 22(9) asks online marketplaces to let traders provide the Article 19 information. The pack checks for these elements and compares the label with the sheet; it does not judge whether a product is safe, it does not check harmonisation legislation (toys, electrical equipment, cosmetics and others add their own rules), which language a Member State requires is for that Member State, and Decosa is not a responsible person. Not legal advice. Model licences: Apache-2.0 (Qwen3.8-27B, Hy-MT2-7B).

### Build a medical chronology (`medical-chronology`)

For personal-injury and med-mal attorneys, life and disability insurers and APS reviewers. Give it a packet of medical records: typed notes, scanned and faxed forms, handwriting, lab tables. The document reader reads every page with boxes, the pages go into a retrieval index as chunks that keep their page and box, and an open model lists the encounters, diagnoses, procedures, imaging, labs and medications on each page. Code keeps only events whose words are on the page and whose date is written there. Copies from different providers are merged, conflicting dates for the same procedure are flagged, and so are gaps in treatment and conditions documented before the injury. Every line cites the file, page and box. A paralegal and reviewer aid, not medical or legal advice.

- What it is: A medical chronology tool for record reviewers: every line of the chronology cites the file, page and box it came from, and copies, conflicting dates, gaps in treatment and pre-existing conditions are flagged.
- Who it's for: Personal-injury and med-mal attorneys and paralegals, nurse reviewers, and life and disability insurers reviewing attending physician statements.
- Page: https://decosa.ai/legal/medical-chronology
- Last reviewed: 2026-09-27
- Key result: On six held-out synthetic packets (96 pages) it found 189 of 198 planted events, each with the right date and a cite on the right page and box, and 2 of 224 lines were not planted events; real records are not measured yet.
- Key numbers (details and caveats: https://decosa.ai/metrics/medical-chronology):
  - 189 / 198: Planted events found (test split, n = 198)
  - 189 / 189: Date right, of events found (test split, n = 189)
  - 189 / 189: Cite on the right page and box, of events found (test split, n = 189)
  - 25.3 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B (events per page with the page image, typed judgments) · document reader (Docling layout + PaddleOCR-VL) · Qwen3-Embedding + Qwen3-Reranker (retrieval over the page boxes). Where: Self-host for real records (PHI); the hosted demo takes synthetic records only.
- Verification: Receipt per model call; a signed document-reader receipt per file; retrieval index and layout hashes with signed search receipts; every quote found on its page; signed hash-chained record (receipt coverage: full). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B, plus about 6 GB for the document reader and 10 GB for the embedder and reranker; the checks, merging and record run on CPU
- Output: A dated chronology in Markdown where every line cites file, page and box, with copies merged, conflicting dates, gaps and pre-existing conditions flagged, and a signed decosa.record.v1.
- Page: https://decosa.ai/legal/medical-chronology · JSON: https://decosa.ai/use-cases/medical-chronology.json
- Prompts: https://decosa.ai/prompts/medical-chronology-hosted.md · https://decosa.ai/prompts/medical-chronology-selfhost.md · https://decosa.ai/prompts/medical-chronology-assemble.md
- Tiers:
  - Lite · document reader and the model, no retrieval service (1x GPU with about 26 GB free for Qwen3.8-27B and the reader): DECOSA_CHRONO_RETRIEVAL=bm25: the same reading, extraction, checks, merging and record; cites are located in a keyword index and the conflict search uses BM25 without the reranker, so about 10 GB less GPU memory.
  - Standard · reader, retrieval service and the model (hosted demo) (1x RTX PRO 6000 96 GB (measured on shared cards)): The document reader, scans into retrieval with Qwen3-Embedding and Qwen3-Reranker, and Qwen3.8-27B per page with the page image. This is what the hosted demo runs, on synthetic records only.
- Buyer facts:
  - Data retention: Nothing stored: files and page images live in memory for the request (the sample packet's reader output is cached in memory by file hash). The signed record holds hashes and receipt ids, never record text or the patient's name; logs carry counts only.
  - What leaves the box: Self-hosted on the direct route: nothing. The reader, the retrieval service, the model and the checks run on the same machine. Hosted: page text and page images go through the Decosa API, and only synthetic records are accepted.
  - What every line carries: Date, type, label, the quote, and a cite to file, page and box (plus the chunk and byte span in the index). Click-through to the box on the sample packet.
  - What it is not: Medical or legal advice. It does not judge causation, standard of care or damages; a paralegal, nurse reviewer or attorney checks each line at its box.
- Samples:
  - 1. Whitlock (`whitlock`): https://decosa.ai/legal/medical-chronology?sample=1&autorun=0
- Regulatory note: Written 27 Sep 2026. Medical records held by HIPAA covered entities (providers, health plans) and their business associates are protected health information; 45 CFR 160.103 names legal and consulting services among the services that make a vendor a business associate (https://www.law.cornell.edu/cfr/text/45/160.103, read 27 Sep 2026), and 45 CFR Part 164 Subpart E sets the privacy rules (https://www.hhs.gov/hipaa/for-professionals/privacy/index.html). A plaintiff's firm that gets records with the patient's authorization is usually not a business associate, and life and disability income insurance are excepted benefits outside HIPAA's 'health plan' (PHS Act 2791(c)(1), cited by 160.103; the list itself was not re-read today, unverified here), but state medical-privacy laws, protective orders and professional duties still apply to the same records. That is why real records are self-host only and the hosted demo refuses anything not marked synthetic. The tool lists what the records say with their page and box; it does not give medical opinions (causation, standard of care, prognosis) or legal advice, and a paralegal, nurse reviewer or attorney checks each line. Model licences: Apache-2.0 (Qwen3.8-27B, PaddleOCR-VL-1.6, Qwen3-Embedding, Qwen3-Reranker), MIT and Apache-2.0 (Docling and its layout weights).

### Screen a claim for Federal IDR (`nsa-idr-packet`)

For out-of-network emergency, anesthesia, radiology and pathology groups, their billing companies, and the health plans on the other side. Give it the EOB or remittance, the open negotiation notice and the supporting documents. It reads the claim facts with quotes, then screens the claim for Federal IDR: coverage, qualified service, state law, notice and consent, the 30-business-day open negotiation start, the 4-business-day initiation window, cooling-off and batching, each check with its dated rule and the business-day arithmetic. When nothing fails, it states the offer as a dollar amount and a percentage of the QPA and drafts a brief from facts it can quote; billed charges, usual and customary charges and Medicare rates are left out. A screen-only mode makes one model call. The packet is signed. A preparation aid: a person files, and a certified IDR entity decides.

- What it is: A No Surprises Act IDR eligibility screen that reads the EOB and notices, shows each check with its dated rule and the business-day arithmetic, and drafts the offer brief only from facts it can quote.
- Who it's for: Out-of-network emergency, anesthesia, radiology and pathology groups, their billing companies and IDR vendors, and health plans answering disputes.
- Page: https://decosa.ai/clinics/nsa-idr-packet
- Last reviewed: 2026-09-27
- Key result: On a held-out set of 61 synthetic disputes, read from the documents alone, the screen got 61 of 61 verdicts right; the documents are templated, so real files will be harder.
- Key numbers (details and caveats: https://decosa.ai/metrics/nsa-idr-packet):
  - 61 of 61: Eligibility verdict right from documents, held out (test split, n = 61)
  - 32 of 32: Planted ineligible disputes caught on the planted check (test split, n = 32)
  - 29 of 29: Eligible disputes passed, including traps (test split, n = 29)
  - 52.1 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real claims (hosted demo: synthetic disputes only).
- Verification: Receipt per model call; every fact's quote found word for word with its date or amount on the line; every brief sentence grounded; signed hash-chained packet (receipt coverage: full). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the clocks, the screen from structured facts and the signed packet need no GPU
- Output: An eligibility screen with each check's rule and arithmetic, the facts with quotes, the offer as a percentage of the QPA, a checked brief when nothing fails, a Markdown packet and a signed record.
- Page: https://decosa.ai/clinics/nsa-idr-packet · JSON: https://decosa.ai/use-cases/nsa-idr-packet.json
- Prompts: https://decosa.ai/prompts/nsa-idr-packet-hosted.md · https://decosa.ai/prompts/nsa-idr-packet-selfhost.md · https://decosa.ai/prompts/nsa-idr-packet-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B reads the facts, gathers the factor evidence and drafts the brief; the screen, the clocks, the offer arithmetic and the prohibited-factor rule are plain code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long files and dense correspondence.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a fact or brief sentence stands when they agree; disagreements go to the reviewer. Claim files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it gives you: An eligibility screen (likely eligible, likely ineligible or needs review) with each check, its dated rule and the business-day arithmetic; the claim facts with quotes; the factor evidence; the offer as a dollar amount and a percentage of the QPA; a checked draft brief when nothing fails; a Markdown packet and a signed record.
  - What it does not do: It does not file in the IDR portal, predict or promise an outcome, recompute or audit a QPA, read state statutes (it uses CMS's chart and says when to check the state's letter), or decide the similar-condition batching test, which is the certified IDR entity's call. Scanned EOBs need OCR first.
  - Data retention: Nothing kept on the server. Documents and packets live in memory for the request; logs carry counts only. You keep the signed packet with the claim.
  - What leaves the box (hosted demo): The documents go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for synthetic disputes only.
  - Model calls per run: Screen mode: one call. Full packet: one fact read, one factor-evidence call, one brief and one grounding call per brief sentence, about 15 to 23 calls on the samples. The clocks and a screen from structured facts need no model.
  - Typical run cost: A fraction of a cent at the gateway list price for a screen from documents; a few cents for the emergency anesthesia packet with its brief. Each run shows its own measured cost.
  - Rules and dates: 45 CFR 149.510 read on 27 Sep 2026 with the staged applicability of the 2026 final rule: $15 fee from 11 Jun 2026, new batching rules for negotiations from 1 Nov 2026, new portal steps only after each is announced. Plans and providers use the same screen.
- Samples:
  - 1. Er anesthesia pa (`er-anesthesia-pa`): https://decosa.ai/clinics/nsa-idr-packet?sample=1&autorun=0
  - 2. Radiology batch tx (`radiology-batch-tx`): https://decosa.ai/clinics/nsa-idr-packet?sample=2&autorun=0
  - 3. Late initiation az (`late-initiation-az`): https://decosa.ai/clinics/nsa-idr-packet?sample=3&autorun=0
  - 4. Ny state law (`ny-state-law`): https://decosa.ai/clinics/nsa-idr-packet?sample=4&autorun=0
  - 5. Plan consent defense (`plan-consent-defense`): https://decosa.ai/clinics/nsa-idr-packet?sample=5&autorun=0
  - 6. Plan offer er co (`plan-offer-er-co`): https://decosa.ai/clinics/nsa-idr-packet?sample=6&autorun=0
- Regulatory note: EOBs and clinical records are protected health information under HIPAA: run it on your own hardware; the hosted demo takes synthetic disputes only. Rules read on 27 Sep 2026 from primary sources. 45 CFR 149.510 (eCFR, current to 24 Sep 2026): open negotiation within 30 business days of receiving the initial payment or denial ((b)(1)); initiation in the 4 business days starting on business day 31 of negotiation ((b)(2)(i)); no IDR after a valid notice-and-consent waiver ((b)(2)(ii)), which never covers ancillary services (149.420(b)); a 90-calendar-day cooling-off period after a determination ((c)(4)(vi), text in force before 3 Aug 2026). The Federal IDR Operations final rules (91 FR 33900, 4 Jun 2026, effective 3 Aug 2026; correction 91 FR 55462, 28 Aug 2026) apply in stages: the $15 administrative fee to disputes initiated on or after 11 Jun 2026; the new batching rules (at most 50 line items, three similar-condition tests, a 30-business-day batched cooling-off) to negotiations starting on or after 1 Nov 2026 (CMS notice, 3 Aug 2026); the new open negotiation, initiation and registry steps only 90 days after the Departments announce each IDR Gateway function (expected from Spring 2027), so until then the earlier text applies (149.510(h)). State law: CMS's applicability chart (state information current as of 11 Jan 2023) and each state's enforcement letter. Courts: TMA II (5th Cir., 2 Aug 2024) affirmed striking the rule on how arbiters weigh the QPA; TMA III (5th Cir. en banc, 11 Aug 2026) held that ghost rates may not be counted in the QPA and incentive payments may not be left out; CMS said guidance would follow (13 Aug 2026). TMA I and TMA IV are described from CMS guidance (opinions not read). It does not recompute QPAs, read state statutes or file anything. Not legal advice, and it never predicts an outcome.

### Quote a job from a walkthrough video (`walkthrough-to-quote`)

For painters, flooring, remodel, restoration, landscaping and moving crews. The customer or estimator walks the job with a phone video and talks. It drafts a line-item scope: rooms and areas, quantities with how each was estimated, materials mentioned, and damage, hazards and access problems seen. Every line cites a time range and a keyframe and is marked seen, said or assumed. Quantities from the video are ranges; it never invents a dimension and says what to measure first. Prices come only from your own price sheet, the sums are done in code and signed. A draft for the estimator: you set the final price.

- What it is: Walkthrough to quote: walk the job with your phone and talk, and the draft quote comes back as line items, each with the moment of the video it comes from, marked seen, said or assumed, with quantities that say where they came from and prices only from your own sheet.
- Who it's for: Owners and estimators at painting, flooring, remodel, restoration, landscaping and moving businesses who quote from a site visit.
- Page: https://decosa.ai/tools/operations/walkthrough-to-quote
- Last reviewed: 2026-09-27
- Key result: On 4 held-out synthetic walkthroughs it found all 33 planted items and every amount re-computed exactly, but only 2 of 5 areas estimated from the video were within its range, so it asks you to measure first.
- Key numbers (details and caveats: https://decosa.ai/metrics/walkthrough-to-quote):
  - 33 / 33: Planted items found (test split, n = 33)
  - 36 / 37: Lines and conditions matching a planted item (test split, n = 37)
  - 26 / 33: Basis right (seen / said / seen and said) (test split, n = 33)
  - 54.0 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B (watches the video, drafts the scope) · MOSS-Transcribe-Diarize (the narration). Where: Hosted or self-host.
- Verification: Receipt per model call, with the video's sha256 and sampling in the request hash; signed ASR receipt; signed record of every line's quantity, unit price and amount (receipt coverage: partial). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: Qwen3.8-27B with video input and MOSS-Transcribe-Diarize: one 96 GB GPU (measured on two shared ones); ffmpeg, pricing and the record on CPU
- Output: A draft quote in Markdown, CSV and JSON: scope lines with time ranges, keyframes, bases, quantities with their method, unit prices and amounts, conditions, where to measure first, and a signed record.
- Page: https://decosa.ai/tools/operations/walkthrough-to-quote · JSON: https://decosa.ai/use-cases/walkthrough-to-quote.json
- Prompts: https://decosa.ai/prompts/walkthrough-to-quote-hosted.md · https://decosa.ai/prompts/walkthrough-to-quote-selfhost.md · https://decosa.ai/prompts/walkthrough-to-quote-assemble.md
- Tiers:
  - Lite · scope and quote from the video only (1x RTX PRO 6000 96 GB (measured)): The areas, conditions and scope lines the video shows, each with a time range and a keyframe, priced from your sheet. No speech to text: send a transcript if you have one, otherwise nothing is marked said and spoken measurements are not used.
  - Standard · narration, scope and quote (hosted demo) (1x RTX PRO 6000 96 GB (measured); the diarizer on a second GPU in the hosted setup): Speech to text for the narration, the video survey, requests and spoken measurements tied to transcript lines, scope lines marked seen, said, seen and said or assumed, a second look at said-only lines, and the priced draft with measure-first and a signed record.
- Buyer facts:
  - What it does: Turns one narrated walkthrough video and your price sheet into scope lines (each with a time range, a keyframe and a basis: seen, said, seen and said, or assumed), conditions (damage, hazards, access), materials mentioned, quantities with where they came from, prices from your sheet, a subtotal range and the lines to measure first. You enter measurements, re-price, and sign; exports in Markdown, CSV and JSON.
  - What it does not do: It does not price from anything but your sheet (no market prices), add tax, markup or minimum charges, or send a quote to a customer. It does not see behind walls or under floors, check codes or permits, or measure: video quantities are estimates, and on held-out clips areas were often far off. Videos over 4 minutes are watched in 2-minute parts at one frame a second, so brief views can be missed.
  - Data retention: The video and its clip are deleted when the run ends. The report, with its keyframes, is kept in memory for one hour for the token or key that made it. Logs carry run ids and counts, never titles, narration or line text. You keep the exports and the signed record, which holds hashes and numbers only.
  - What leaves the box (hosted demo): The clip (no audio), the narration text and your price sheet's codes and descriptions go to Qwen3.8-27B through our gateway, which Decosa operates; the audio is transcribed on Decosa's hosted service. The gateway's receipts hold hashes, not content. Self-hosted, nothing leaves.
  - Accuracy: On 4 held-out synthetic walkthroughs: 33 of 33 planted items found, 36 of 37 lines and conditions matched a planted item, the basis right on 26 of 33 (2 of 5 said-only items were wrongly marked seen), counts exact on 11 of 14, but only 2 of 5 video areas and lengths had the truth inside the range (median error 68.7%). Every priced amount re-computed exactly.
  - Cost per walkthrough: A few cents for a short walkthrough at gateway list price (a few model calls, measured on the demo sample), plus speech to text on Decosa's hosted service.
  - Output: A draft quote with lines, time ranges, keyframes, bases, quantities and their method, unit prices and amounts; a decosa.record.v1 record of the arithmetic, re-signed when you re-price and when you sign (verify at /record/verify), and POST /walkthrough/check-arithmetic to re-run the sums.
- Samples:
  - 1. Painter bedroom hall (`painter-bedroom-hall`): https://decosa.ai/tools/operations/walkthrough-to-quote?sample=1&autorun=0
  - 2. Flooring living dining (`flooring-living-dining`): https://decosa.ai/tools/operations/walkthrough-to-quote?sample=2&autorun=0
- Regulatory note: Not legal advice; a draft for the estimator, not a contract or a sendable quote. Checked 27 Sep 2026: if the quote leads to a sale agreed at the customer's home, the FTC's Cooling-Off Rule may apply: 16 CFR 429.0(a) defines a door-to-door sale as one where the buyer's agreement is made 'at a place other than the place of business of the seller', at $25 or more at the buyer's residence, and the buyer then has three business days to cancel (https://www.ecfr.gov/current/title-16/chapter-I/subchapter-D/part-429, read via the Cornell LII copy at https://www.law.cornell.edu/cfr/text/16/429.0). Many states also set rules for home-improvement contracts; we did not check them. Recording a conversation: in some US states everyone must consent (California Penal Code 632(a) makes it an offence to record a confidential communication 'without the consent of all parties', https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=632.&lawCode=PEN); film with the customer's knowledge. Video of a home can be personal data; keep people out of frame. The demo walkthroughs are synthetic.

### Turn an expert video into an SOP (`expert-to-sop`)

Turns a screen or bench recording of an expert doing a procedure into a draft SOP. Every step cites a time range and a keyframe from the recording and is checked against what the expert said. Steps that were said but never shown are marked needs confirmation, steps shown but never said are marked, and a named reviewer signs each revision into a record anyone can verify. A draft for a person to approve, not a certified procedure.

- What it is: SOP from video: an expert records the screen or the bench while doing the procedure and talking, and the draft comes back as numbered steps, each with the time it happens, a keyframe from that moment and a check against what was said.
- Who it's for: CI, training and quality leads who keep work instructions current, and IT and ops teams who own runbooks.
- Page: https://decosa.ai/tools/operations/expert-to-sop
- Last reviewed: 2026-09-27
- Key result: On 4 held-out synthetic screen recordings, 21 of 25 planted steps were found, all in order, and 4 of 5 steps said but never shown were flagged; on one recording the cited times were wrong and a warning says so.
- Key numbers (details and caveats: https://decosa.ai/metrics/expert-to-sop):
  - 21 / 25: Planted on-screen steps found (test split, n = 25)
  - 100%: Found steps in the right order (test split, n = 21)
  - 19 / 21: Cited start within 3 s (test split, n = 21)
  - 20.2 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: live. Models: Qwen3.8-27B (watches the recording, checks each step) · MOSS-Transcribe-Diarize (the narration). Where: Hosted or self-host.
- Verification: Receipt per model call, with the video's sha256 and sampling in the request hash; signed ASR receipt; signed revision record with step, keyframe and receipt hashes (receipt coverage: partial). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: Qwen3.8-27B with video input and MOSS-Transcribe-Diarize: one 96 GB GPU (measured on two shared ones); ffmpeg, keyframes and the record on CPU
- Output: A draft SOP in Markdown and JSON: numbered steps with time ranges, keyframes and statuses, and a signed revision record naming the reviewer.
- Page: https://decosa.ai/tools/operations/expert-to-sop · JSON: https://decosa.ai/use-cases/expert-to-sop.json
- Prompts: https://decosa.ai/prompts/expert-to-sop-hosted.md · https://decosa.ai/prompts/expert-to-sop-selfhost.md · https://decosa.ai/prompts/expert-to-sop-assemble.md
- Tiers:
  - Lite · steps and keyframes, no narration check (1x RTX PRO 6000 96 GB (measured)): The steps the video shows, each with a time range and a keyframe, and the sign-off record. No speech to text, so nothing is checked against what was said and said-only steps are not found.
  - Standard · steps checked against the narration (hosted demo) (1x RTX PRO 6000 96 GB (measured); the diarizer on a second GPU in the hosted setup): Speech to text for the narration, the steps from the video, each step judged against what was said near its time, and instructions said but not shown checked back against the video and marked needs confirmation.
- Buyer facts:
  - What it does: Turns one narrated recording of a procedure into numbered steps, each with a time range, a keyframe and a status: seen and said, seen but not narrated, narration differs, or said but not shown (needs confirmation). A named reviewer decides every flagged step and signs the revision; exports in Markdown and JSON.
  - What it does not do: It does not judge whether a procedure is safe, correct or compliant, and a draft is not an approved procedure. It does not hear sound effects or read audio cues other than speech, merge several recordings, or diff revisions. Recordings over 4 minutes are sampled more sparsely (one frame every 2.5 s at 10 minutes), so short actions can be missed.
  - Data retention: The recording and its clip are deleted when the run ends. The report, with its keyframes, is kept in memory for one hour for the token or key that made it. Logs carry run ids and counts, never titles, narration or step text. You keep the exports and the signed record, which holds hashes only.
  - What leaves the box (hosted demo): The clip (no audio) and the narration text go to Qwen3.8-27B through our gateway, which Decosa operates; the audio is transcribed on Decosa's hosted service. The gateway's receipts hold hashes, not content. Self-hosted, nothing leaves.
  - Accuracy: On 4 held-out synthetic screen recordings: 21 of 25 planted on-screen steps found, all in order; 4 of 5 said-but-not-shown steps flagged; no seen step wrongly flagged. On 1 of the 4 the model cited times in 2-second slots instead of reading them; the draft then shows a timing warning.
  - Cost per recording: A few cents or less for a short recording at gateway list price (measured on the demo sample). Each run shows its own measured cost.
  - Output: A draft SOP with steps, time ranges, keyframes and statuses; after sign-off a decosa.record.v1 revision record with the recording hash, each step's hashes, the receipts and the reviewer (verify at /record/verify).
- Samples:
  - 1. Labeldesk add printer (`labeldesk-add-printer`): https://decosa.ai/tools/operations/expert-to-sop?sample=1&autorun=0
  - 2. Hmi crimp calibration (`hmi-crimp-calibration`): https://decosa.ai/tools/operations/expert-to-sop?sample=2&autorun=0
- Regulatory note: Not legal or safety advice; a draft for a qualified person to review, not a validated or approved procedure. Checked against the primary text on 27 Sep 2026: where OSHA's control of hazardous energy standard applies, 29 CFR 1910.147(c)(4)(i) says procedures 'shall be developed, documented and utilized', (c)(4)(ii) that they 'shall clearly and specifically outline the scope, purpose, authorization, rules, and techniques', and (c)(6)(i) requires a periodic inspection of each procedure at least annually (https://www.osha.gov/laws-regs/regulations/standardnumber/1910/1910.147). A drafted SOP does not meet those duties by itself: the employer writes, authorizes and inspects the procedure. Recording people at work can be personal data (for example under the GDPR where workers are identifiable); record with the knowledge of the people in the recording and keep faces out where you can. The demo recordings are synthetic.

### Check generated product images (`honest-product-imagery`)

For e-commerce sellers and brands that make product and lifestyle images with AI. Send photos of the real product and the AI image: it checks the net quantity, colour, pack text and claims, logo, warnings, and features or extra units the product does not have, with deterministic measurements (colour difference in Lab, the document reader for the pack text, logo and unit matching) plus a side-by-side open-model comparison. Any person in the image must be a consent-ledger identity whose consent covers this advertising use. An image that passes gets the AI metadata and label each marketplace or law you pick asks for, a C2PA credential and a signed record. It checks images made anywhere; it does not generate them. Not legal advice.

- What it is: Honest AI product images start with a check: this compares an AI-made product or lifestyle shot with photos of the real product, stops it when the size, colour, pack text, logo or features are wrong or a person in it has no consent for the ad, and adds the metadata and label each marketplace asks for.
- Who it's for: E-commerce sellers, DTC brands and the agencies that make AI product and lifestyle images for them.
- Page: https://decosa.ai/tools/media/honest-product-imagery
- Last reviewed: 2026-09-27
- Key result: On the four test products (thresholds frozen before the run) it flagged 48 of 48 planted misrepresentations and 3 of 24 faithful images; it still misses small warning text the image model garbled.
- Key numbers (details and caveats: https://decosa.ai/metrics/honest-product-imagery):
  - 48 / 48: Planted misrepresentations flagged (test split, n = 48)
  - 3 / 24: Faithful images flagged (test split, n = 24)
  - 5 / 5: Generator misrepresentations flagged (not planted) (test split, n = 5)
  - 12.9 s: Median end-to-end run, hosted (QA sweep 2026-09-27)

- Status: preview. Models: Qwen3.8-27B (finds the product, logo and people; compares the two images) · document reader block: Docling layout + PaddleOCR-VL-1.6 (pack text). Where: Hosted or self-host.
- Verification: Receipt per model call and per page-parse call; signed consent decisions; C2PA credential with the fidelity result; signed hash-chained record (receipt coverage: partial). Hosted QA 2026-09-27: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B with its vision tower; the document reader's layout model and parser run beside it; the checks, credential and record on CPU
- Output: A verdict with each misrepresentation found, the consent decisions, the disclosure per target with sources, the approved JPEG with IPTC metadata and a C2PA credential, and a signed record.
- Page: https://decosa.ai/tools/media/honest-product-imagery · JSON: https://decosa.ai/use-cases/honest-product-imagery.json
- Prompts: https://decosa.ai/prompts/honest-product-imagery-hosted.md · https://decosa.ai/prompts/honest-product-imagery-selfhost.md · https://decosa.ai/prompts/honest-product-imagery-assemble.md
- Tiers:
  - Lite · pixels and the model, no document reader (1x RTX 5090 32 GB for the model (not measured) or the hosted gateway): Alignment, colour, logo and unit checks in code plus the model's comparison; the quantity, claim and warning text is judged by the model only, so those findings stay warnings.
  - Standard · pixels, pack text and the model (hosted demo) (Qwen3.8-27B on one 96 GB card; the document reader beside it (about 6 GB)): Adds the document reader: printed quantities, claims and warnings compared in code, so a 750 ml label on a 500 ml product fails without the model's say-so.
- Alternates (smaller models for the same job):
  - Lifestyle scenes for demos (not part of the service) (shared studio GPU, one job at a time; about 30-45 s per 1152x768 still (measured)), self-host only: Wan2.2-VACE-Fun-A14B drew the demo and eval scenes around synthetic packs. The service checks images from any generator; it does not generate them.
- Buyer facts:
  - What it checks: Net quantity and count, product colour, pack text and claims, the logo, warnings, and parts, accessories or extra units the product does not have, against your own product photos and the facts you state. Any person or human likeness in the image needs a consent-ledger identity whose consent covers this advertising use.
  - What it does not do: It does not generate images, give legal advice or say a listing or ad is lawful. It cannot know about a product change you did not tell it about. It misses small warning or claim text that the image model garbled: in the eval it passed all 17 such images (a frontier model caught 15). It does not identify people: it only checks the identities you name.
  - Data retention: Images are held in memory and a temporary folder for the run and deleted when it ends; an approved image is kept with the run for one hour for the token or key that made it, so you can download it. Consent decisions on the demo go to a private in-memory ledger. Logs carry ids, verdicts and counts, never images or product facts.
  - What leaves the box (hosted demo): The images go to Qwen3.8-27B through our gateway and to the document reader service, both operated by Decosa. Self-hosted, nothing leaves the machine.
  - Cost per image: A fraction of a cent of model calls for a full check (a few image calls) and less for a consent refusal, at the gateway list price (measured).
  - Output: A verdict with each misrepresentation found and how, the consent decisions, the disclosure per marketplace or law with its source, the approved JPEG with IPTC metadata (and Amazon's synthetic-performer keyword when it applies), a burned-in label where one is required, a C2PA credential, and a signed record (verify at /record/verify).
- Samples:
  - 1. Faithful shelf (`faithful-shelf`): https://decosa.ai/tools/media/honest-product-imagery?sample=1&autorun=0
  - 2. Bystanders (`bystanders`): https://decosa.ai/tools/media/honest-product-imagery?sample=2&autorun=0
  - 3. Pump swap (`pump-swap`): https://decosa.ai/tools/media/honest-product-imagery?sample=3&autorun=0
  - 4. Size 750ml (`size-750ml`): https://decosa.ai/tools/media/honest-product-imagery?sample=4&autorun=0
  - 5. Colour shift (`colour-shift`): https://decosa.ai/tools/media/honest-product-imagery?sample=5&autorun=0
  - 6. Warning dropped (`warning-dropped`): https://decosa.ai/tools/media/honest-product-imagery?sample=6&autorun=0
  - 7. Extra units (`extra-units`): https://decosa.ai/tools/media/honest-product-imagery?sample=7&autorun=0
  - 8. Synthetic model (`synthetic-model`): https://decosa.ai/tools/media/honest-product-imagery?sample=8&autorun=0
  - 9. Wrong campaign (`wrong-campaign`): https://decosa.ai/tools/media/honest-product-imagery?sample=9&autorun=0
  - 10. No identity (`no-identity`): https://decosa.ai/tools/media/honest-product-imagery?sample=10&autorun=0
  - 11. Withdrawn consent (`withdrawn-consent`): https://decosa.ai/tools/media/honest-product-imagery?sample=11&autorun=0
  - 12. Consented but wrong cap (`consented-but-wrong-cap`): https://decosa.ai/tools/media/honest-product-imagery?sample=12&autorun=0
- Regulatory note: Not legal advice or a determination that a listing or ad is lawful; sources read on the primary pages on 27 Sep 2026 unless marked. Misrepresentation: FTC Act s.5 and the FTC Policy Statement on Deception (14 Oct 1983, https://www.ftc.gov/legal-library/browse/ftc-policy-statement-deception); EU Unfair Commercial Practices Directive 2005/29/EC Art. 6(1)(b) (main characteristics such as composition, accessories, quantity; https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32005L0029). AI marking and deep fakes: EU AI Act (Regulation (EU) 2024/1689) Art. 50(2) and 50(4), applying from 2 Aug 2026 (Art. 113); generators on the market before that date have until 2 Dec 2026 for 50(2) (Art. 111(4), added by Regulation (EU) 2026/1744, OJ 24 Jul 2026). Synthetic performers: New York GBL s.396-b (https://www.nysenate.gov/legislation/laws/GBS/396-B) requires an advertiser with actual knowledge to disclose a synthetic performer conspicuously; in force from 9 Jun 2026 per the disclosure pre-flight's reading of chapter 617 of 2025 (exact date not re-verified). Model consent: New York Fashion Workers Act, Labor Law Art. 36, s.1037(7) (clear and conspicuous prior written consent to create or use a model's digital replica, stating scope, purpose, rate of pay and duration; https://www.nysenate.gov/legislation/laws/LAB/1037), effective 19 Jun 2025 per the NY Department of Labor FAQ. Marketplaces: Amazon's product image guide (G1881) asks images to represent the product accurately and to tag photorealistic AI-generated people (keyword contains-synthetic-performer); Google Merchant Center asks for AI metadata such as IPTC DigitalSourceType TrainedAlgorithmicMedia; Etsy's Creativity Standards (10 Jun 2025) require disclosure for items made with AI, not for AI photos of real items; Meta labels images carrying C2PA or IPTC AI indicators, and whether it requires AI disclosure in ads outside political and social-issue ads was not verified. Model licences: Apache-2.0 (Qwen3.8-27B, PaddleOCR-VL-1.6, the Heron layout weights), MIT (Docling).

### Draft an adverse-event case (`pv-intake`)

For drug safety teams at marketing authorisation holders and the CROs that run case intake for them. Send an adverse event report as it arrived: an email, a web form, a call to the medical information line, or a scanned MedWatch or CIOMS form, in any of nine languages. It drafts an E2B(R3)-shaped case with every field quoted to its source line (and the reporter's own words for a translated report), checks the four minimum criteria, marks the seriousness criteria per event, checks expectedness against the reference label you send, and works out day 0, which the report itself may move earlier, and the US 15-day and EU 15- and 90-day dates. Coding suggestions are free text, not MedDRA. Anything missing comes back as follow-up questions in the reporter's language. The safety physician decides and signs.

- What it is: Pharmacovigilance case intake that drafts an ICSR from an adverse event report as it arrived: an email, a web form, a call recording or a scanned MedWatch or CIOMS form, in nine languages. Every field is quoted to its source line and the safety physician decides.
- Who it's for: Drug safety teams at marketing authorisation holders, CROs running case intake, and medical information centres.
- Page: https://decosa.ai/tools/life-sciences/pv-intake
- Last reviewed: 2026-09-28
- Key result: On 80 synthetic reports run once (held out): seriousness 60 of 60, expectedness 56 of 60, day 0 exact 55 of 59; one labeller's synthetic cases, not real files.
- Key numbers (details and caveats: https://decosa.ai/metrics/pv-intake):
  - 17 / 20: Missing minimum criteria caught (held out, n = 20)
  - 1 / 300: Criteria wrongly called missing (held out, n = 300)
  - 60 / 60: Seriousness right (held out, n = 60)
  - 7.1 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Qwen3.8-27B (fields with quotes, seriousness, expectedness, the narrative and its grounding check) · MOSS-Transcribe-Diarize and Qwen3-ASR-1.7B (calls) · Hy-MT2-7B (translation) · document reader (scans); the criteria and the clocks are plain code. Where: Self-host for real reports (they hold health data); the hosted demo takes synthetic reports only.
- Verification: Receipt per model call; every field kept has a quote found word for word in the report; narrative sentences grounded; signed case and decision records (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B, Hy-MT2-7B, the call recogniser and the document reader; the criteria, clocks and records run on CPU
- Output: A draft case: the four minimum criteria, fields quoted to their lines, seriousness and expectedness per event, day 0 and the 15- and 90-day dates, follow-up questions, an E2B(R3)-shaped JSON draft, a Markdown memo, and signed case and decision records.
- Page: https://decosa.ai/tools/life-sciences/pv-intake · JSON: https://decosa.ai/use-cases/pv-intake.json
- Prompts: https://decosa.ai/prompts/pv-intake-hosted.md · https://decosa.ai/prompts/pv-intake-selfhost.md · https://decosa.ai/prompts/pv-intake-assemble.md
- Tiers:
  - Lite · one 48-80 GB card (1x L40S or A100 80 GB (not measured)): The same pipeline with Gemma 4 26B A4B instead of Qwen. Faster; accuracy on this task not measured.
  - Standard · the hosted demo (1x RTX PRO 6000 Blackwell 96 GB (estimate for everything on one card; the demo uses two)): Qwen3.8-27B reads and judges; the quote checks, criteria, clocks and records are code. Every model call receipted.
  - Best · a larger judge (2x RTX PRO 6000 96 GB (192 GB weights)): DeepSeek-V4-Flash for seriousness and expectedness, where the frontier comparison shows the gap. Not measured here.
  - Wanted: the best setup · a large judge with room to spare (2-4x RTX PRO 6000 96 GB or a large-memory server): DeepSeek-V4-Flash as the judge for every event, with the recognisers and translation beside it. Needs more than one 96 GB card.
- Buyer facts:
  - What it gives you: A draft case: the four minimum criteria each met by a quote (or what is missing), seriousness criteria and expectedness per event with quotes from the report and the label, day 0 and the US 15-day and EU 15- and 90-day dates with the arithmetic, follow-up questions in the reporter's language, an E2B(R3)-shaped JSON draft, free-text coding suggestions, an optional checked narrative, a Markdown memo, and signed case and decision records.
  - What it does not do: It is not a safety database and not a validated system. It does not submit E2B XML to FAERS or EudraVigilance, code in MedDRA (no MedDRA terms are shipped), assess causality, search for duplicates, screen literature, or handle devices, vaccines (VAERS) or clinical-trial SUSARs.
  - Data retention: Nothing kept on the server. Reports, audio, scans and records live in memory for the request (the bundled samples' transcripts are cached by hash); logs carry counts and the case summary only. You keep the signed records with the case.
  - What leaves the box (hosted demo): Report text goes to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text; audio, scans and translation stay on Decosa's hosted service. Self-hosted, nothing leaves. The hosted demo is for synthetic reports only.
  - Model calls per report: One fields read, one seriousness call and one expectedness call per event; plus speech recognition for a call, the document reader for a scan, line-by-line translation for a report not in English, and a draft plus one grounding call per sentence when the narrative is on.
  - Typical run cost: A fraction of a cent in model time for a text report without the narrative (a few calls on the demo email, gateway list price). A call adds speech recognition; a translated report adds a translation call per line; the narrative adds a draft and one check per sentence.
  - Clocks: Day 0 is the first day anyone at the company or acting for it (sales reps, call centres, partners) had the four criteria; the report itself can move it earlier. US 15 calendar days for serious and unexpected; EU 15 days for serious, 90 for non-serious. EU deadlines are not moved off weekends; for US ones the date FDA's 2001 draft guidance would accept is shown beside it.
- Samples:
  - 1. Email angioedema (`email-angioedema`): https://decosa.ai/tools/life-sciences/pv-intake?sample=1&autorun=0
  - 2. Call febrile neutropenia (`call-febrile-neutropenia`): https://decosa.ai/tools/life-sciences/pv-intake?sample=2&autorun=0
  - 3. Scan cioms sepsis (`scan-cioms-sepsis`): https://decosa.ai/tools/life-sciences/pv-intake?sample=3&autorun=0
  - 4. Email fr rash (`email-fr-rash`): https://decosa.ai/tools/life-sciences/pv-intake?sample=4&autorun=0
  - 5. Email de no patient (`email-de-no-patient`): https://decosa.ai/tools/life-sciences/pv-intake?sample=5&autorun=0
  - 6. Email es no product (`email-es-no-product`): https://decosa.ai/tools/life-sciences/pv-intake?sample=6&autorun=0
  - 7. Web form nausea (`web-form-nausea`): https://decosa.ai/tools/life-sciences/pv-intake?sample=7&autorun=0
  - 8. Call er fall (`call-er-fall`): https://decosa.ai/tools/life-sciences/pv-intake?sample=8&autorun=0
- Regulatory note: Rules read on 27 Sep 2026. US: 21 CFR 314.80 (drugs) and 600.80 (biologics), eCFR as of 24 Sep 2026: each adverse experience that is both serious and unexpected is reported no later than 15 calendar days from initial receipt; the rest go in periodic reports (quarterly for 3 years from approval, then annually); unexpected includes events of greater severity or specificity than the labeling. FDA's March 2001 draft guidance (still a draft, not binding) names the four data elements, makes the day the company knows them day 0, and lets a 15th day on a weekend or federal holiday move to the next working day. EU: Directive 2001/83/EC Article 107(3): serious suspected adverse reactions within 15 days and non-serious ones within 90 days of the day the MAH gained knowledge; EMA GVP Module VI Rev 2 (2017): the four minimum criteria, day 0 when any MAH personnel (including medical representatives and contractors) has them, calendar days with no move for weekends or holidays, and the reporter's qualification and country for a valid ICSR; ICH E2D(R1) (Step 4, 15 Sep 2025; in effect in the EU from 18 Mar 2026) says the same for the criteria and day 0. E2B(R3) element ids are checked against the EU ICSR implementation guide only. A triage and drafting aid, not a validated safety database and not regulatory advice: the safety physician decides; it submits nothing to FAERS or EudraVigilance and ships no MedDRA terminology. Reports hold health data: run it on your own hardware; the hosted demo takes synthetic reports only.

### Compare safety sections across labels (`label-consistency-check`)

For regulatory affairs and labelling teams. Give it the label documents for one medicine: the company core data sheet (CCDS), the US Prescribing Information, the EU SmPC and package leaflet in several languages, and the carton text or a scan of the carton. It aligns the sections (indications, dosing, contraindications, warnings, adverse reactions, storage, strengths) and flags every difference with a quote from each document: a dose, interval or strength that disagrees, a warning or contraindication that is missing, a different storage temperature, a translation that dropped a negation or changed a number, an SmPC heading that is not the QRD template's. Each flag is triaged as a likely error, likely deliberate (a regional convention such as US Fahrenheit storage wording, or a line in your deviation log) or unclear. A labelling QC aid with a signed record; regulatory affairs decides.

- What it is: A CCDS, a US PI, an EU SmPC and leaflet in several languages and the carton in; every difference in the safety sections out, quoted from each document and triaged as likely error, likely deliberate or unclear.
- Who it's for: Regulatory affairs and labelling teams at pharma and biotech companies, and labelling service providers.
- Page: https://decosa.ai/tools/life-sciences/label-consistency-check
- Last reviewed: 2026-09-28
- Key numbers (details and caveats: https://decosa.ai/metrics/label-consistency-check):
  - 31 / 31: Planted drifts caught (test split, n = 31)
  - 30 / 31: Quotes tight to the drift (test split, n = 31)
  - 6 / 6: Logged twins triaged deliberate (test split, n = 6)
  - 2.3 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Qwen3.8-27B (section comparison with quotes, triage, meaning judgment) · Hy-MT2-7B (back-translation, language-pack block) · Docling layout + PaddleOCR-VL (scanned cartons, document reader). Where: Self-host for unapproved labelling; the hosted demo takes invented or public labels only.
- Verification: Receipt per model call; every quote found in its document at a character offset; numbers with units compared in code; signed hash-chained record of the document hashes, flags and triage (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB): Qwen3.8-27B NVFP4 plus Hy-MT2-7B (about 18 GB) and the document reader (about 6 GB); sections, number checks and the record run on CPU
- Output: The flags as CSV (each with its section, quotes from both documents, severity and triage) and a signed decosa.record.v1 of the document hashes, flags and receipt ids.
- Page: https://decosa.ai/tools/life-sciences/label-consistency-check · JSON: https://decosa.ai/use-cases/label-consistency-check.json
- Prompts: https://decosa.ai/prompts/label-consistency-check-hosted.md · https://decosa.ai/prompts/label-consistency-check-selfhost.md · https://decosa.ai/prompts/label-consistency-check-assemble.md
- Tiers:
  - Lite · text documents, meaning check off (1x GPU for Qwen3.8-27B): Text documents only, sent with "meaning": false: the model compares sections and triages; translations get the code checks only (numbers, units, negations, QRD headings), no back-translation. Only the language model needs a GPU.
  - Standard · Qwen3.8-27B, Hy-MT2-7B and the document reader (hosted demo) (1x RTX PRO 6000 96 GB (measured on shared cards)): What the hosted demo runs: Qwen3.8-27B compares and triages, Hy-MT2-7B back-translates for the meaning check, the document reader reads scanned cartons, code checks numbers, headings and quotes and signs.
  - Best · the 30B translation model (not measured) (1x RTX PRO 6000 96 GB (estimate: 20 GB Qwen3.8 NVFP4 weights plus 31 GB MT)): Hy-MT2-30B-A3B-FP8 for back-translation; not run by us.
- Buyer facts:
  - Data retention: Nothing stored: the documents live in memory for the request. The signed record holds document hashes, sections, each flag's place, kind and triage and the receipt ids, never the label text; logs carry counts and timings only.
  - What leaves the box: Self-hosted on the direct route: nothing. The model, the translation model, the document reader and the checks run on the same machine. Hosted: model calls go through the Decosa API, and only invented or public labels are accepted.
  - What it compares: Indications, dosing, contraindications, warnings, adverse reactions, storage and strengths: US PI and English SmPC against the CCDS, leaflet and carton against the SmPC, each language against its English text, SmPC headings against the QRD template (English, German, French).
  - What it is not: A regulatory decision, an approval of a deviation or an artwork proofreader. Interactions, pregnancy sections, pharmacology and layout are not compared. Regulatory affairs decides each flag.
- Samples:
  - 1. Norvexa: full set with five drifts (`norvexa-drift`): https://decosa.ai/tools/life-sciences/label-consistency-check?sample=1&autorun=0
  - 2. Norvexa: consistent set with regional differences (`norvexa-clean`): https://decosa.ai/tools/life-sciences/label-consistency-check?sample=2&autorun=0
  - 3. Norvexa: CCDS against the US PI (quick) (`norvexa-quick`): https://decosa.ai/tools/life-sciences/label-consistency-check?sample=3&autorun=0
  - 4. Norvexa: scanned carton against the SmPC (`norvexa-scan`): https://decosa.ai/tools/life-sciences/label-consistency-check?sample=4&autorun=0
- Regulatory note: Checked 27 Sep 2026. US: 21 CFR 201.57(c) sets the Full Prescribing Information sections (1 Indications and usage, 2 Dosage and administration, 3 Dosage forms and strengths, 4 Contraindications, 5 Warnings and precautions, 6 Adverse reactions ... 16 How supplied/storage and handling; eCFR, current), and 21 CFR 314.70(c)(6)(iii) lets a holder add or strengthen a contraindication, warning, precaution or adverse reaction, or a dosage instruction for safe use, on a changes-being-effected (CBE-0) basis. EU: Directive 2001/83/EC Article 11 sets the SmPC headings in order (4.1 therapeutic indications ... 4.3 contra-indications, 4.4 special warnings ..., 4.8 undesirable effects, 6.4 special precautions for storage) and Article 63(1) requires labelling and leaflet particulars in the official language(s) of the Member State; read on legislation.gov.uk (the EU text as it stood on 31 Dec 2020), so later amendments and the EU pharmaceutical reform are unverified here. The EMA QRD product-information template v10.4 (02/2024; a draft v11 was consulted on until 31 Aug 2025) gives the headings and standard statements in each language; the check uses its English, German and French versions. ICH E2C(R2) (Step 4, 17 Dec 2012) describes the Company Core Data Sheet and its core safety information (CCSI), which is why the CCDS is the reference. The check lists and triages differences; it does not decide whether a label is approvable, whether a deviation is justified or which regulatory route a change needs. Not legal or regulatory advice. Model licences: Apache-2.0 (Qwen3.8-27B, Hy-MT2-7B, PaddleOCR-VL-1.6, Docling Heron layout); MIT (Docling).

### Find accessibility fixes for a shop (`storefront-accessibility-pass`)

For online shops selling into the EU or the US, and the agencies that build them. A private headless browser opens the shop's pages, runs axe-core's automated WCAG 2.2 A/AA rules, tries add-to-cart and the checkout with the keyboard alone (reachable, answers Enter, focus visible, no trap), and asks an open model the questions a rule engine cannot answer: is this alt text right for this photo, is "More" a clear link name here, does "Invalid input" say how to fix the field, is the offer only inside a banner image. You get a fix list, purchase blockers first, each with its WCAG success criterion and a code fix, and a signed record of what was checked. An audit aid, not a certification; not legal advice.

- What it is: A storefront accessibility audit for up to five pages of a shop: axe-core's WCAG 2.2 rules, a keyboard-only pass to add-to-cart and checkout, and an open model that checks alt text against each image, link names and error messages. You get a ranked fix list and a signed record.
- Who it's for: Small online shops selling into the EU or the US, and the agencies and theme developers who build them.
- Page: https://decosa.ai/tools/operations/storefront-accessibility-pass
- Last reviewed: 2026-09-28
- Key result: On the held-out test pages of a made-up shop it caught 122 of 127 planted issues and flagged none of 207 clean items; one synthetic shop template, no real themes tested.
- Key numbers (details and caveats: https://decosa.ai/metrics/storefront-accessibility-pass):
  - 122 / 127: Planted issues caught (test split, n = 127)
  - 0 / 207: Clean items flagged (test split, n = 207)
  - 0 / 7: Clean pages with any finding (test split, n = 7)
  - 1.8 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Qwen3.8-27B with vision (alt text against the image, text in images, link and button names, error messages) · axe-core 4.13 (MPL-2.0, automated WCAG rules) in headless Chromium. Where: Hosted or self-host.
- Verification: Receipt per model call; signed hash-chained record with the axe-core version and hash, the prompts' hashes, the pages and every finding (receipt coverage: partial). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-27: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B with its vision tower; headless Chromium and axe-core on CPU
- Output: A fix list ranked by what blocks a purchase, each finding with its WCAG 2.2 success criterion, the element and a code fix, and a signed record of what was checked.
- Page: https://decosa.ai/tools/operations/storefront-accessibility-pass · JSON: https://decosa.ai/use-cases/storefront-accessibility-pass.json
- Prompts: https://decosa.ai/prompts/storefront-accessibility-pass-hosted.md · https://decosa.ai/prompts/storefront-accessibility-pass-selfhost.md · https://decosa.ai/prompts/storefront-accessibility-pass-assemble.md
- Tiers:
  - Lite · rules, keyboard and forms, no model (CPU) (CPU only (a headless Chromium per audit)): axe-core, the keyboard pass and the form pass only ("judge": false): no GPU, no model calls. Misses wrong or keyword-stuffed alt text, vague names, unclear error messages and text only in images.
  - Standard · rules, keyboard, forms and the model (hosted demo) (Qwen3.8-27B on one 96 GB card, or the hosted gateway; the rest on CPU): Adds Qwen3.8-27B with vision for the judgements a rule engine cannot make: alt text against the image, names in context, error messages, text only in images.
- Buyer facts:
  - What it checks: axe-core's automated WCAG 2.2 A/AA rules; add-to-cart and the checkout with the keyboard alone (reachable, answers Enter or Space, visible focus, no trap); error messages after an empty submit; and, by the model, whether each alt text fits its image, whether link and button names are clear in context, whether errors say how to fix the field, and offer text that exists only inside an image. Up to five pages of one shop per audit, desktop or 390 px wide.
  - What it does not do: It does not certify conformance or say a site meets WCAG, the ADA or the European Accessibility Act, and it is not legal advice. It does not test screen-reader output, reading order, zoom and reflow, captions or cognitive load, every widget, or pages behind a login on the hosted service. It never pays, places an order or submits a form to your server. The model judgements are suggestions for a person to review.
  - Data retention: Nothing is stored: pages, screenshots and images live in memory and a private browser profile for the audit and are dropped when it ends. The result and its signed record are returned to you, not kept. Logs carry counts and ids, never URLs, alt text or page text.
  - What leaves the box (hosted demo): Page images, alt text, link and button names and error messages go to Qwen3.8-27B through our gateway, operated by Decosa. The headless browser only fetches the pages you list, on a domain your key has verified. Self-hosted, nothing leaves the machine except the browser's requests to your own site.
  - Output: A fix list ranked P1 (blocks a purchase) to P4, each finding with its WCAG 2.2 success criteria, the page, the element's HTML and a code fix, plus advisories, what was and was not checked, and a signed record (verify at /record/verify) with the axe-core version and hash, the prompts' hashes and a receipt per model call.
  - Cost per page: A fraction of a cent of model calls per page (a few calls) at the gateway list price, measured over the test pages. Without the model it is CPU only.
- Samples:
  - 1. Shop with issues (`shop-with-issues`): https://decosa.ai/tools/operations/storefront-accessibility-pass?sample=1&autorun=0
  - 2. Clean shop (`clean-shop`): https://decosa.ai/tools/operations/storefront-accessibility-pass?sample=2&autorun=0
  - 3. Product page (`product-page`): https://decosa.ai/tools/operations/storefront-accessibility-pass?sample=3&autorun=0
- Regulatory note: An audit aid, not a certification or a statement that a site conforms to WCAG, the ADA or the European Accessibility Act; not legal advice. Sources read on the primary pages on 27 Sep 2026. EU: the European Accessibility Act, Directive (EU) 2019/882 (https://eur-lex.europa.eu/eli/dir/2019/882/oj/eng), covers e-commerce services to consumers (Art. 2(2)(f), defined in Art. 3(30)) and applies from 28 Jun 2025 (Art. 31(2)); microenterprises providing services (fewer than 10 staff and annual turnover or balance sheet of no more than EUR 2 million, Art. 3(23)) are exempt (Art. 4(5)); harmonised standards give a presumption of conformity (Art. 15), and the one used for web content, EN 301 549, points to WCAG 2.1 AA (the EN 301 549 version was not re-checked); penalties are set by each Member State (Art. 30). US: DOJ's web guidance of 18 Mar 2022 (https://www.ada.gov/resources/web-guidance/) says ADA Title III covers the goods and services businesses open to the public offer online, with no Title III regulation setting a technical standard; WCAG is named as helpful guidance, and lawsuits and settlements usually cite WCAG 2.1 AA. The Title II web rule (28 CFR part 35 subpart H, published 24 Apr 2024, https://www.ada.gov/resources/2024-03-08-web-rule/) sets WCAG 2.1 AA for state and local governments only, not private shops; an interim final rule published 20 Apr 2026 moved its compliance dates to 26 Apr 2027 and 26 Apr 2028. Overlays: the FTC's final order against accessiBe (22 Apr 2025, USD 1 million, https://www.ftc.gov/news-events/news/press-releases/2025/04/ftc-approves-final-order-requiring-accessibe-pay-1-million) bars claims that an automated tool makes a site WCAG-compliant without evidence, which is why this pass reports what it checked and never says compliant. Licences: Qwen3.8-27B Apache-2.0; axe-core MPL-2.0 (vendored unmodified, source and licence shipped with it); Playwright Apache-2.0; Chromium BSD-3-Clause.

### Fill a form from your own papers (`fill-and-stop`)

For anyone re-typing the same facts into a long form: a claim, a benefits renewal, a W-9, a school form. Give it the form and your papers. A PDF form (fillable, flat or scanned) comes back filled, with a source sheet showing the line each answer came from and why each blank is blank. For a form on a website it never touches the site (none of the insurer, benefits, school or payer portals we checked allows automated tools): you get a copy list in the site's own order, each answer with its source and a copy button. It never signs, dates your signature, ticks a statement for you or fills ID, bank or card numbers, and you can review in your language.

- What it is: Fill a PDF form from your own papers, or get a copy list for a form on a website: every answer shows the line it came from, what your papers do not say is left for you, and you sign and send it yourself.
- Who it's for: People re-typing the same facts from papers they already have, often in a second language.
- Page: https://decosa.ai/tools/operations/fill-and-stop
- Last reviewed: 2026-09-28
- Key result: On 50 held-out synthetic PDF forms it wrote 1 wrong answer out of 491 and left all 150 signature and certification boxes alone; forms written by the same author as the checker.
- Key numbers (details and caveats: https://decosa.ai/metrics/fill-and-stop):
  - 1 / 491 (0.20%): PDF forms: wrong answers / answers written (test split, n = 491)
  - 490 / 497 (98.6%): PDF forms: answers the papers give, filled (test split, n = 497)
  - 150 / 150: PDF forms: signature, signing-date and certification boxes left alone (test split, n = 150)
  - 13.7 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: preview. Models: Qwen3.8-27B (points each box at the line of your papers that answers it, checks form text for injection) · document reader for scanned forms and photos · language pack for the review. Where: Hosted (the PDF sample or your own PDF, copy lists, our demo web form) or self-host.
- Verification: Receipt per model call; a signed record of every answer (as a hash) with its source line and every blank with its reason; the filled file's hash (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the sandboxed Chromium, form reader and network hold run on CPU
- Output: The filled PDF (or a copy list for a website form), a source sheet with every answer's line and every blank's reason, and a signed record of the run.
- Page: https://decosa.ai/tools/operations/fill-and-stop · JSON: https://decosa.ai/use-cases/fill-and-stop.json
- Prompts: https://decosa.ai/prompts/fill-and-stop-hosted.md · https://decosa.ai/prompts/fill-and-stop-selfhost.md · https://decosa.ai/prompts/fill-and-stop-assemble.md
- Tiers:
  - Lite · text and PDF sources, English review (Qwen3.8-27B on one 96 GB card, or the hosted gateway; the rest on CPU): Qwen3.8-27B, the sandboxed browser and the checks: fills from text and born-digital PDFs and reviews in English. No scans or photos of papers, no translated review.
  - Standard · adds scans, photos and a review in 24 languages (hosted demo) (One 96 GB card: Qwen3.8-27B (57 GB), document reader (6 GB), Hy-MT2-7B (18 GB)): Adds the document reader for scanned PDFs and photos of papers, and the language pack for a review in any of the 24 EU languages; the values typed into the form are never translated.
  - Wanted: the best setup · a second model on choices and widgets (More than one 96 GB card (estimate)): A larger vision-language model (or Holo3 grounding) as a second opinion on the dropdown and radio choices marked 'check it' and on custom widgets. Not built or measured.
- Buyer facts:
  - What it does: Fills a PDF form (fillable, flat or scanned) from your papers and hands it back with a source sheet; for a form on a website, gives you a copy list in the site's order to paste from, and never touches the site. Every answer is checked to be in the line of your papers it cites. Boxes your papers do not answer are left for you, with the reason.
  - What it never does: Sign, write the date you sign, or tick a certification or consent; fill passwords, card, bank or routing numbers, Social Security, tax ID (EIN), passport, licence or Medicare numbers; follow instructions written on the form; automate a website whose terms forbid tools (copy list instead); choose an answer your papers do not support (inferred choices are marked 'check it'). No immigration, court or customs forms.
  - Measured: PDF forms (50 held-out synthetic forms: 36 fillable, 8 flat, 6 scanned; claim, benefits, W-9 style, school): 1 wrong answer in 491 written (0.20%), 150 of 150 signature, date-of-signing and certification boxes left alone, every blank with the right reason (409 of 409), 0 sensitive numbers leaked; blanks found on flat PDFs 144 of 144 and on scans 114 of 114 (113 labelled right). Web forms (the demo engine): 6 wrong in 1,166 typed, 0 forms sent. Same author wrote the forms and the checker.
  - Where it runs: Hosted: the PDF sample or a PDF you upload, copy lists for any website, and the live engine on our demo web form, all with made-up or public papers. Real papers: run it yourself (self-host), or ask for confidential access.
  - Data retention: Hosted runs keep the filled PDF, the source sheet and the signed record for 24 hours, readable only with your key or demo token, then delete them; 'Delete this run now' deletes them at once. Logs carry counts and ids, never answers or document text.
  - What leaves the box (hosted demo): Box labels and your papers' lines (sensitive numbers withheld) go to Qwen3.8-27B on Decosa's servers; scanned pages go to our document reader for OCR; labels and cited lines go to the language pack if you pick another language. No website is contacted by the PDF or copy modes. Self-hosted, nothing leaves your machine unless you point it at the hosted model.
  - Cost per form: PDF forms: a fraction of a cent of model use per form on average on our test set (about one call; seconds per form, scans slowest); the demo claim on production also costs a fraction of a cent and takes seconds, at the gateway list price.
- Samples:
  - 1. Harbor claim pdf (`harbor-claim-pdf`): https://decosa.ai/tools/operations/fill-and-stop?sample=1&autorun=0
  - 2. Harbor claim pdf es (`harbor-claim-pdf-es`): https://decosa.ai/tools/operations/fill-and-stop?sample=2&autorun=0
  - 3. School meals copy (`school-meals-copy`): https://decosa.ai/tools/operations/fill-and-stop?sample=3&autorun=0
  - 4. Harbor claim (`harbor-claim`): https://decosa.ai/tools/operations/fill-and-stop?sample=4&autorun=0
  - 5. Harbor claim short (`harbor-claim-short`): https://decosa.ai/tools/operations/fill-and-stop?sample=5&autorun=0
- Regulatory note: A filling aid, not advice: it never chooses what to answer, and the person reviews, signs and sends. Website terms: on 28 Sep 2026 we read the terms of 25 consumer, government, vendor and school portals and 13 payer portals; none allows automated filling, and Login.gov and ID.me require people to use them by hand. So websites get the copy list (the person pastes), and the hosted engine only automates our demo forms and forms whose owner consents; a per-site table (default: copy) enforces it. Filling a PDF is offline and involves no website; do not use it on licensed blank forms you were not given (for example ACORD sets). Out of scope on purpose: immigration forms, court filings and customs entries. Health or account data belongs on your own box (self-host); the hosted demo takes synthetic or public papers only. Licences: Qwen3.8-27B Apache-2.0; Hy-MT2-7B Apache-2.0; PaddleOCR-VL-1.6 Apache-2.0 and Docling MIT (the document reader); pypdf BSD-3-Clause, pdfplumber MIT, pypdfium2 Apache-2.0/BSD-3-Clause; DejaVu Sans (free licence) for writing on flat PDFs; Playwright Apache-2.0; Chromium BSD-3-Clause. Written 28 Sep 2026; not legal advice.

### Turn a client call into notes (`privileged-call-notes`)

For solo and small-firm lawyers who take client calls all day. After the call you get an intake memo where every line cites the transcript and is checked against it, a conflicts-check list of every name heard, a time entry in 0.1-hour units with a billing narrative, follow-up tasks and an email draft in the client's language. No notetaker bot joins the call: open models run on the server or your own box, the audio is deleted once the transcript is made, and the session ends in a signed record.

- What it is: Legal call notes without a notetaker bot: record or upload a client call and get an intake memo where every line cites the transcript, the names for your conflicts check, a 0.1-hour time entry and follow-up tasks.
- Who it's for: Solo and small-firm lawyers, and their assistants, who turn client calls into intake memos and time entries.
- Page: https://decosa.ai/legal/privileged-call-notes
- Last reviewed: 2026-09-28
- Key numbers (details and caveats: https://decosa.ai/metrics/privileged-call-notes):
  - 89.3%: Memo fact recall (blind grader) (test split, n = 20)
  - 0 / 12 of 796: Memo lines invented / with a wrong detail (test split, n = 796)
  - 89.6% / 94.0%: Conflicts-name recall, strict / phonetic (test split, n = 20)
  - 148.0 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Voxtral Mini 4B Realtime (live captions) · MOSS-Transcribe-Diarize 0.9B (who said what) · Qwen3.8-27B (memo, names, claim check, time entry). Where: Self-host or the Confidential tier for real client calls; the hosted demo is for synthetic calls.
- Verification: Receipt per model call and per transcript line; signed, hash-chained session record with the consent statement and the audio deletion (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for the language model; Voxtral and the diarizer need about 27 GB more (a second card on our server)
- Output: A cited intake memo with its claim check, a conflicts-check list, a 0.1-hour time entry, follow-up tasks, a client email draft and a signed session record.
- Page: https://decosa.ai/legal/privileged-call-notes · JSON: https://decosa.ai/use-cases/privileged-call-notes.json
- Prompts: https://decosa.ai/prompts/privileged-call-notes-hosted.md · https://decosa.ai/prompts/privileged-call-notes-selfhost.md · https://decosa.ai/prompts/privileged-call-notes-assemble.md
- Tiers:
  - Lite · one 48 GB card, uploaded calls (1x L40S or RTX 6000 Ada 48 GB (not measured)): Diarized transcript and the full memo, conflicts, time entry and tasks on the FP8 checkpoint; no live captions.
  - Standard · the hosted demo (1x RTX PRO 6000 Blackwell 96 GB (hosted demo uses two cards on our server)): Live captions and checklist, then speakers, cited memo, claim check, conflicts, time entry, tasks and email, all in a signed record.
  - Best · DeepSeek-V4-Flash writes and checks (2x RTX PRO 6000 96 GB for the writer plus the standard card): Standard plus a larger memo writer and checker on two more 96 GB cards.
- Buyer facts:
  - Who it's for: Solo and small-firm lawyers who take client calls all day, and the assistants who write the intake recaps and time entries.
  - Data retention: The call audio is held in memory only and dropped as soon as the transcript is made; its hash and the time go into the signed record. Nothing is written to disk. The memo, transcript and signed record are returned to you; the server keeps receipts (hashes), not the text.
  - No notetaker bot: Nothing joins your call. Record in the browser or upload the recording afterwards; speech recognition, speaker labels and the memo run on open models: on Decosa's hosted service, on your own box, or in an attested enclave.
  - Recording consent: You say where you and the client are before recording. When any state involved requires every party's consent (or is mixed or unknown), recording stays off until you confirm everyone agreed.
  - Typical run cost: A few cents or less per call at list price (a few dozen receipted model calls for a short call), plus speech recognition on Decosa's hosted service. Measured on synthetic calls.
  - What leaves the box (self-host): Nothing, with DECOSA_LLM_ROUTE=direct. Anyone can verify a record offline or with POST /record/verify.
- Samples:
  - 1. Custody modification call, family law (synthetic) (`cn-family-custody-ortiz`): https://decosa.ai/legal/privileged-call-notes?sample=1&autorun=0
  - 2. DUI consult in Spanish, criminal defense (synthetic) (`cn-criminal-dui-reyes`): https://decosa.ai/legal/privileged-call-notes?sample=2&autorun=0
  - 3. Slip-and-fall intake, personal injury (synthetic) (`cn-pi-slip-fall-thibodeaux`): https://decosa.ai/legal/privileged-call-notes?sample=3&autorun=0
- Regulatory note: Not legal advice (checked 28 Sep 2026). Recording consent: federal law and most states allow a party to record a call (one-party consent, 18 U.S.C. 2511(2)(d): https://www.law.cornell.edu/uscode/text/18/2511); California and eight other states require every party's consent (Cal. Penal Code 632: https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=PEN&sectionNum=632), and five states are mixed. The tool asks where each side is, applies the strictest rule, and will not record until the consent step is done; it picks a prompt from a dated table and does not decide whether a recording is lawful. Confidentiality: ABA Model Rule 1.6(c) asks lawyers to make reasonable efforts to prevent unauthorized disclosure of client information, and ABA Formal Opinion 512 (29 Jul 2024) applies the duties of competence and confidentiality to generative AI tools. Sending client calls to a third-party notetaker raises privilege-waiver questions; this tool runs on open models on your own box or in an attested enclave, but you decide whether that meets your duties. Not a certified transcript: check names, dates and amounts against the call.

### Check the other side's brief (`check-their-brief`)

For the litigator answering an opposition: upload the other side's filing and get findings for your reply. Every cited case, statute and rule is looked up in public sources (a case none of them has is "not found in the sources searched", with the search trail, never "fake"), quotations are matched word for word against the opinion, holdings are checked against the opinion's text, and the file is scanned for text the judge cannot see (white, tiny or off-page text, invisible characters, metadata, comments) and for instructions aimed at AI tools. Each finding cites the page of their filing and the source, in a Word memo and a signed record.

- What it is: Check opposing brief citations and hidden text before you reply: every citation looked up in public sources with a search trail, quotations matched word for word, and the PDF scanned for text the judge cannot see but an AI tool reads.
- Who it's for: Litigators and paralegals answering an opposition or reply, especially against AI-drafted or pro se filings.
- Page: https://decosa.ai/legal/check-their-brief
- Last reviewed: 2026-09-30
- Key result: On 390 held-out brief excerpts it caught 24 of 31 non-existent citations and 31 of 45 swapped-word misquotes as findings; misstated holdings are not reliably caught.
- Key numbers (details and caveats: https://decosa.ai/metrics/check-their-brief):
  - 24 / 31 (77%): Non-existent citations caught (strict) (held out, n = 31)
  - 34 / 68 (50%): Case name and cite from two different cases (strict) (held out, n = 68)
  - 31 / 45 (69%): Swapped-word misquotations (strict) (held out, n = 45)
  - 8.3 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: live. Models: Qwen3.8-27B (holdings against the opinion; which hidden texts speak to AI tools) · document reader for scanned filings. Where: Hosted or self-host; the confidential tier for anything not yet public.
- Verification: Receipt per model call; search trail per citation; signed record (receipt coverage: full). Hosted QA 2026-09-30: pass. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) or 1x RTX 5090 (32 GB) for the judge; the hidden-text scan, lookups and quotation match run on CPU
- Output: Findings for your reply, each tied to a page of their filing and a public source, as a Word memo, plus a signed record of what was checked, when and where.
- Page: https://decosa.ai/legal/check-their-brief · JSON: https://decosa.ai/use-cases/check-their-brief.json
- Prompts: https://decosa.ai/prompts/check-their-brief-hosted.md · https://decosa.ai/prompts/check-their-brief-selfhost.md · https://decosa.ai/prompts/check-their-brief-assemble.md
- Tiers:
  - Lite · CPU only, no GPU (Any Linux or macOS machine with Python 3.11+): The hidden-text scan with the injection patterns, existence, impossible reporters and quotations. No holdings check, and hidden texts that hit no pattern are listed without a model's reading.
  - Standard · one GPU for the judge (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate, without the document reader)): Adds the holdings check and the model's reading of hidden texts with Qwen3.8-27B, one receipted call each, and scanned filings through the document reader. This is what the hosted API runs.
  - Wanted · a GLM-5.3-Flash holdings judge on your own hardware (Your own hardware: 2x 96 GB cards (NVFP4, about 170 GB, estimate) or a Mac with 192 GB or more (MLX 4-bit, estimate).): A stronger open judge from another family for the holdings check, where the 27B is weakest here. Your notes can be privileged, so it runs on your hardware, never on community providers. Not served yet.
- Alternates (smaller models for the same job):
  - Our citation-support model as a second opinion on holdings (CPU (8 threads, about 0.9 s per holding) or any small GPU), not served: Turns a holding the judge calls unsupported into a finding only when our own fine-tuned model agrees: on held-out pairs 80 of 129 misstated holdings at 7 false flags in 372, against 26 at 22 for the judge alone. Prototype weights, off by default (DECOSA_THEIRBRIEF_SUPPORT_URL).
- Buyer facts:
  - Data retention: Hosted: the filing lives in memory; an uploaded file's text and a finished run are kept for one hour (for the export links), then dropped. Answers from the public sources (the citation queried, public opinion text; never the filing) are cached on disk for 7 days. Logs carry counts only.
  - What leaves the box: Hosted: each model call goes through the gateway to the GPU serving Qwen3.8-27B, with a signed receipt (hashes and token counts, no text). Self-hosted on the direct route: the filing never leaves the box; lookups send only citations and case names to CAP, CourtListener, LII, uscode.house.gov and eCFR; offline mode sends nothing.
  - Input formats: Their filing as a PDF (hidden text is checked), a Word file or pasted text; up to 20 MB and 120,000 characters. Scanned PDFs are read by the document reader (hosted, up to 40 pages).
  - Typical run: One model call reads the hidden texts and one checks each holding. The first check of a filing is slower, while the public sources are looked up; repeat lookups are cached. The page header shows the time and cost measured on production, and each run shows its own.
  - Wording: Nothing is called fake. A case the public sources lack is 'not found in the sources searched', with the URLs searched; a real case at the wrong cite is 'their cite is wrong'.
  - Confidential tier: For papers that are not public (served but unfiled, your notes): the same check can run inside an attested enclave, where neither Decosa nor the cloud host can read the text; enterprise pricing, ask us. A filed brief is public, so the hosted demo is fine for it.
- Samples:
  - 1. Demo opposition (`demo-opposition`): https://decosa.ai/legal/check-their-brief?sample=1&autorun=0
  - 2. Mata ecf21 (`mata-ecf21`): https://decosa.ai/legal/check-their-brief?sample=2&autorun=0
  - 3. Osg case v montana (`osg-case-v-montana`): https://decosa.ai/legal/check-their-brief?sample=3&autorun=0
- Regulatory note: Findings for a lawyer to verify, not legal advice (as of 28 Sep 2026). Before telling a court that a cited case does not exist, pull the reporter: public sources miss some unpublished, very recent and Westlaw- or Lexis-only decisions. Fed. R. Civ. P. 11(b)(2) makes the signer of a filing certify that its legal contentions are warranted by existing law, and courts sanction invented citations (Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023)); a sanctions motion under Rule 11(c)(2) must be served 21 days before it is filed. If the scan finds text under a black box (a failed redaction), ABA Model Rule 4.4(b) requires a lawyer who receives inadvertently sent information to notify the sender; the memo never shows those words. A filed brief is a public record; your notes about it, and papers served but not filed, belong on your own box or the confidential tier.

### Check their discovery responses (`discovery-deficiency`)

For litigation associates and paralegals on either side of written discovery. Paste or upload the other side's interrogatory answers or responses to requests for production or admission. It flags objections with no specifics, incorporated general objections, responses that don't say whether anything is being withheld, promises to produce with no date, evasive or incomplete answers, admission responses that neither admit nor deny, privilege claims with no log, a missing verification or signature, and late service, with the deadline worked out. Each flag quotes the response and cites the rule text (federal courts, California, Texas). Then it turns the flags into a first-draft deficiency letter with the rule for each point: a starting point to rewrite, not a letter ready to send. Run it on your own draft responses before you serve them: it suggests fixes and never removes an objection.

- What it is: A discovery deficiency letter starts with finding what is wrong: this tool reads each response, flags boilerplate objections, unstated withholding, evasive answers, improper admission responses and late service with the rule text, and turns them into a first-draft letter you rewrite.
- Who it's for: Litigation associates, paralegals and solo litigators who review the other side's written discovery responses or check their own before serving.
- Page: https://decosa.ai/legal/discovery-deficiency
- Last reviewed: 2026-09-28
- Key result: On unseen real cases it caught 74% of the responses a motion to compel called deficient, and its splitter found 75% of responses; 6.6% of clean synthetic responses were wrongly flagged.
- Key numbers (details and caveats: https://decosa.ai/metrics/discovery-deficiency):
  - 13 / 197 (6.6%): False flags on clean responses, blind synthetic test (held out, n = 197)
  - 0.897: Deficient-or-not precision, blind synthetic test (held out, n = 325)
  - 0.883: Deficient-or-not recall, blind synthetic test (held out, n = 325)
  - 4.8 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Qwen3.8-27B (one structured read of each response) · code (splitting, rule cites, deadlines, the letter) · document reader for scanned PDFs. Where: Hosted for invented or public responses; self-host for client material.
- Verification: Receipt per model call; every quote found in the response at a character offset; every rule cite from a rules pack of verbatim rule text (read 28 Sep 2026); signed hash-chained record (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; splitting, rule cites, deadlines, the letter and the record run on CPU
- Output: Flags per request (quote, what is missing, rule text), the response and motion deadlines, a meet-and-confer letter as Word (or a pre-service fix list), and a signed decosa.record.v1 of the flags and receipt ids.
- Page: https://decosa.ai/legal/discovery-deficiency · JSON: https://decosa.ai/use-cases/discovery-deficiency.json
- Prompts: https://decosa.ai/prompts/discovery-deficiency-hosted.md · https://decosa.ai/prompts/discovery-deficiency-selfhost.md · https://decosa.ai/prompts/discovery-deficiency-assemble.md
- Tiers:
  - Lite · one 32 GB card (1x RTX 5090 32 GB (estimate)): Text, Word and text-layer PDFs only (no scans): the model on one consumer card, everything else on CPU.
  - Standard · Qwen3.8-27B and the document reader (hosted demo) (1x RTX PRO 6000 96 GB (measured on shared cards)): What the hosted demo runs: Qwen3.8-27B reads each response, code assigns flags, cites rules, computes deadlines and writes the letter; the document reader handles scans.
- Buyer facts:
  - Data retention: Nothing stored: the responses live in memory for the request. The signed record holds the document hash, each flag's category, place and rule ids and the receipt ids, never the response text; logs carry counts and timings only.
  - What leaves the box: Self-hosted on the direct route: nothing. The model and the checks run on the same machine. Hosted: model calls go through the Decosa API, and only invented or public responses are accepted.
  - What it checks: Per response: objections without specifics, incorporated general objections, no statement of whether anything is withheld, no production date, evasive or incomplete answers (including "see documents produced"), admission responses that neither admit nor deny, privilege claims with no log, the pre-2015 "reasonably calculated" test (federal only). Per set: general objections, verification, signature, late service.
  - What it is not: Legal advice, a motion or a filing. It cites rule text, not case law, local rules or standing orders, and does not judge whether the requests themselves were proper. An attorney decides and signs.
  - Time and cost per response set: Model cost at list price: a fraction of a cent per response (one Qwen3.8 call per response; the letter and deadlines are code). The sample runs in seconds; real sets took about a minute on a shared gateway. A blind associate estimated hours to review and letter a set by hand; the letter draft is a starting point, not ready to send (see the eval).
  - Confidential tier: Not yet run in the attested enclave. Every model call here is the same kind of single Qwen3.8 chat call that other tools ran sealed on 28 Sep 2026; the proven setup is your own self-hosted API with its model calls sealed to the enclave. Scanned PDFs also need the document reader, which runs in the enclave's capability build. Until a sealed run is recorded for this tool, use self-hosting for client material.
- Samples:
  - 1. Their RFP responses (federal, N.D. Ill.) (`harbor-rfp`): https://decosa.ai/legal/discovery-deficiency?sample=1&autorun=0
  - 2. Their special interrogatory answers (California) (`marlowe-rogs`): https://decosa.ai/legal/discovery-deficiency?sample=2&autorun=0
  - 3. Our draft RFA responses (pre-service, W.D. Tex.) (`delgado-rfa-draft`): https://decosa.ai/legal/discovery-deficiency?sample=3&autorun=0
- Regulatory note: Checked 28 Sep 2026. A review aid for lawyers and litigation staff, not legal advice and not for people without a lawyer: an attorney decides which deficiencies to raise and signs the letter (Fed. R. Civ. P. 26(g)(1) makes the signer certify discovery papers). Rules cited: Fed. R. Civ. P. 26(b)(1), 26(b)(5)(A), 26(g), 33(b), 33(d), 34(b)(2)(A)-(C), 36(a)(3)-(6), 37(a)(1), (a)(4), 6(a), 6(d) and the 2015 Advisory Committee Notes to Rules 26 and 34 (https://www.law.cornell.edu/rules/frcp); Cal. Code Civ. Proc. 2016.040, 2017.010, 2030.210-2030.300, 2031.210-2031.310, 2033.210-2033.280, 1010.6, 1013 (https://leginfo.legislature.ca.gov); Tex. R. Civ. P. 191-198 and 21a (https://www.txcourts.gov/rules-forms/rules-standards/). Every quote in the rules pack was matched word for word against those pages on 28 Sep 2026; the Texas rules came from the official PDF, which carries no single effective-date stamp. It does not apply case law (courts differ on whether general or boilerplate objections are waived), local rules or a judge's standing order; state deadlines are counted on the US federal holiday list, which can differ from a state court's. Confidentiality: discovery responses carry client facts and are often under a protective order; ABA Formal Opinion 512 (29 Jul 2024) asks lawyers to understand how a tool uses what they put in and to protect client information, so run real matters self-hosted, where nothing leaves the machine. The hosted demo takes invented or public responses only. Model licence: Apache-2.0 (Qwen3.8-27B).

### Answer a payer audit (`payer-audit`)

For the owner or manager of a small practice, or its billing service, when a payer or its review contractor asks for records or audits paid claims. Give it the auditor's letter and the charts. It works out the respond-by date from the letter in code, reads the claim list, matches each claim to its note, and checks every documentation requirement of the payer's published policy: found, with the chart, page and line, or missing. Weak claims are listed first, as plainly as the supported ones. It drafts a cover letter whose sentences are each checked against the sources, builds an indexed, bookmarked response packet, and signs a record of what was checked. It never suggests changing a record, and it never touches a payer portal. The same check runs as a self-audit before a payer asks.

- What it is: An auditor's records request and the charts in; the respond-by date, each claim checked against the payer's policy with page:line cites, weak claims first, and a packet.
- Who it's for: Owners and managers of small practices that take insurance, and the billing services that work for them.
- Page: https://decosa.ai/clinics/payer-audit
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/payer-audit):
  - 9 / 9: Weak claims flagged, new blind BCBSM letter (29 Sep) (held out, n = 9)
  - 7 / 9: Weak claims flagged, new blind Optum letter (29 Sep) (held out, n = 9)
  - 23 / 26: Weak claims flagged, held-out B rerun (29 Sep) (held out, n = 26)
  - 59.9 s: Median end-to-end run, hosted (QA sweep 2026-09-28)

- Status: live. Models: Qwen3.8-27B. Where: Self-host for real records (hosted demo: synthetic cases only; confidential access can't take patient data yet).
- Verification: Receipt per model call; every found element's words located in the note by code; times, dates and signatures read in code; every cover-letter sentence grounded; signed hash-chained record (receipt coverage: full). Hosted QA 2026-09-28: pass. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the deadline, the matching, the packet and the signed record need no GPU
- Output: The respond-by date, a per-claim table of the policy's requirements with cites, weak claims first, a draft cover letter, a bookmarked response packet, a worksheet, a CSV and a signed record.
- Page: https://decosa.ai/clinics/payer-audit · JSON: https://decosa.ai/use-cases/payer-audit.json
- Prompts: https://decosa.ai/prompts/payer-audit-hosted.md · https://decosa.ai/prompts/payer-audit-selfhost.md · https://decosa.ai/prompts/payer-audit-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Accuracy on this task not measured.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B reads the letter and each note; the respond-by date, the matching, times, signatures, dates, the claim status and the packet are code. Every model call receipted.
- Buyer facts:
  - What it gives you: The respond-by date with its arithmetic; each claim's note and each policy requirement found (chart, page, line and the words) or missing; weak claims first; a draft cover letter; a bookmarked PDF packet with an index; the values an upload form asks for with their sources; a worksheet, a CSV and a signed record.
  - Objective measures: BCBS Michigan asks for objective tools to monitor progress: a scored instrument's result (PHQ-9, GAD-7, PCL-5) counts, the client's own rating (SUDS, a 0-10 score) does not (our reading; BCBSM does not define the term). Optum and Evernorth ask for no instrument in a progress note, and the packs say so with their sources.
  - Checks across claims: Two sessions by the same clinician that overlap on the same day, and a note whose wording is nearly identical to another claim's note, are flagged to check by hand, with the other claim named.
  - What it does not do: It does not send, fax or upload anything, or log in to a portal. It does not judge medical necessity or whether a code was right, know your contract or the auditor's sampling, or read scanned charts (text and pages only). It never suggests changing a record.
  - Data retention: Nothing written to disk. The letter, charts and result live in memory; a finished run is kept one hour so its downloads work, then dropped. Logs carry counts only. The signed record holds hashes, not chart text.
  - What leaves the box (hosted demo): The letter and notes go to Qwen3.8-27B through Decosa's gateway, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo takes synthetic cases only; real records go through self-host (confidential access can't take patient data until a business associate agreement is in place).
  - Model calls per audit: One letter read, one call per claim, one cover-letter draft and one grounding call per letter sentence: 18 calls for 12 claims, about 145 for 140.
  - Typical run cost: A few cents for the small sample and more for a full audit at the gateway list price; a fraction of a cent per claim on the held-out sets. Each run shows its own measured cost.
  - Codes: You supply the billed codes. The tool holds no code descriptors or code-set guidelines; it checks the chart against the payer's own published policy text, quoted with its source.
- Samples:
  - 1. Therapist 140 (`therapist-140`): https://decosa.ai/clinics/payer-audit?sample=1&autorun=0
  - 2. Therapist 12 (`therapist-12`): https://decosa.ai/clinics/payer-audit?sample=2&autorun=0
  - 3. Self audit 60 (`self-audit-60`): https://decosa.ai/clinics/payer-audit?sample=3&autorun=0
- Regulatory note: Charts are protected health information under HIPAA: run it on the practice's own hardware (confidential access can't take patient data until a business associate agreement is in place); the hosted demo takes synthetic cases only. Not legal, coding or compliance advice: a person, with their compliance lead or counsel, decides what to send and signs. It never suggests adding to, changing or back-dating documentation (a code guard enforces this) and it flags entries dated after the request instead of counting them. It holds no CPT descriptors or AMA guideline text: the practice supplies its codes, and the check uses the payer's own published policy text, quoted with the source. Payer and auditor portals are not automated (their terms forbid it). In a self-audit, rules on identified overpayments may apply (for Medicare, 42 CFR 401.305, text read on eCFR 28 Sep 2026); ask counsel.

### Check a prior auth before you send it (`prior-auth-check`)

For prior-authorization coordinators and medical assistants, before a request goes out. Give it the payer's own published medical policy (pasted, or the PDF) and the chart. It finds the criteria section for the service, then marks each criterion met with a chart quote, missing with the record to get, or not met with the line that contradicts it, and says plainly: ready to send, get these records first, or not supported on this chart. It answers the payer's form questions with chart quotes for you to copy into the portal yourself, writes a one-page peer-to-peer brief, and drafts a letter of medical necessity only when every criterion is met, each sentence checked against the chart and the policy. Licensed criteria (InterQual, MCG) are never reproduced: a policy that points to them gets a plain "can't check". It never suggests codes or changes to a note, and never touches a payer portal.

- What it is: The payer's own policy and the chart in, before you send: each criterion met, missing or not met with quotes, and a plain call on whether to send.
- Who it's for: Prior-authorization coordinators, medical assistants and physicians at specialty and primary-care practices.
- Page: https://decosa.ai/clinics/prior-auth-check
- Last reviewed: 2026-09-28
- Key numbers (details and caveats: https://decosa.ai/metrics/prior-auth-check):
  - 14 of 16: Decision right, latest fresh set (test split, n = 16)
  - 1 of 11: Requests that should have waited, called ready (test split, n = 11)
  - 24 of 35: Decision right, earlier fresh set (test split, n = 35)

- Status: preview. Models: Qwen3.8-27B. Where: Self-host for real patients (hosted demo: synthetic charts only; confidential access can't take patient data yet).
- Verification: Receipt per model call; every criterion quote found word for word in the policy, every met or not-met answer quoted from the chart, not-met answers re-checked; every letter sentence grounded; signed hash-chained record (receipt coverage: full). Hosted QA 2026-09-28: partial. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the policy reader, the decision clock and the signed record need no GPU
- Output: Send, get records first, or not supported; each criterion with its quote; form values to copy; a peer-to-peer brief; a letter of medical necessity when the chart supports it; a Markdown packet and a signed record.
- Page: https://decosa.ai/clinics/prior-auth-check · JSON: https://decosa.ai/use-cases/prior-auth-check.json
- Prompts: https://decosa.ai/prompts/prior-auth-check-hosted.md · https://decosa.ai/prompts/prior-auth-check-selfhost.md · https://decosa.ai/prompts/prior-auth-check-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B finds the section, maps and checks the criteria, answers the form and drafts the letter; the quote checks, the recommendation, the clock and the brief are plain code. Every model call receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long charts and dense payer policies.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Patient records stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it gives you: A call (ready to send, get records first, not supported on this chart, can't check), each criterion met, missing or not met with the policy and chart quotes, what record to get, form values with their chart lines to copy, a peer-to-peer brief, a letter of medical necessity when the chart supports it, the federal decision clock, a Markdown packet and a signed record.
  - What it does not do: It doesn't submit, fax or touch a payer portal, predict the payer's decision, know which services need authorization, suggest codes, or fetch policies (paste them or upload the PDF). It never includes InterQual or MCG criteria. It never suggests adding to or changing a note. Scanned charts and scanned policy PDFs need OCR first.
  - Data retention: Nothing kept on the server. Policies, charts and results live in memory for the request; logs carry counts only. You keep the signed record with the request.
  - What leaves the box (hosted demo): The policy section and the chart go to Qwen3.8-27B through Decosa's gateway, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for synthetic charts only.
  - Model calls per check: One criteria map (plus one section pick for a long policy), one check per criterion, a re-check per not-met answer and per exclusion answered met, one form call; with a letter, one draft and one grounding call per sentence: 17 to 20 calls at the median, up to about 50.
  - Typical run cost: About a cent or two per check at the gateway list price (held-out runs); more with long policies and a letter. Each run shows its own measured cost.
  - Measured accuracy: On the latest fresh held-out set (16 synthetic charts, 4 real policies): 14 decisions right; 1 of 11 requests that should have waited called ready. Earlier sets: 24 of 35 and 45 of 60. Measured 28 Sep 2026; see the eval.
- Samples:
  - 1. Cpap ready (`cpap-ready`): https://decosa.ai/clinics/prior-auth-check?sample=1&autorun=0
  - 2. Lumbar mri ready (`lumbar-mri-ready`): https://decosa.ai/clinics/prior-auth-check?sample=2&autorun=0
  - 3. Bariatric missing eval (`bariatric-missing-eval`): https://decosa.ai/clinics/prior-auth-check?sample=3&autorun=0
  - 4. Epidural gather (`epidural-gather`): https://decosa.ai/clinics/prior-auth-check?sample=4&autorun=0
  - 5. Cpap not supported (`cpap-not-supported`): https://decosa.ai/clinics/prior-auth-check?sample=5&autorun=0
  - 6. Proprietary criteria (`proprietary-criteria`): https://decosa.ai/clinics/prior-auth-check?sample=6&autorun=0
- Regulatory note: Charts are protected health information under HIPAA: run it on the practice's own hardware (confidential access can't take patient data until a business associate agreement is in place); the hosted demo takes synthetic charts only. Decision clocks read on 28 Sep 2026 (Cornell LII, eCFR text): employer group plans must decide a pre-service claim within 15 days of receiving it, extendable once by 15 days (29 CFR 2560.503-1(f)(2)(iii)(A)), and an urgent one within 72 hours ((f)(2)(i)); ACA individual coverage follows the same rules (45 CFR 147.136(b)(3)(i)); Medicare Advantage: 7 calendar days for a standard request for an item or service subject to its prior-authorization rules from 1 Jan 2026 (42 CFR 422.568(b)(1)(ii)) and 72 hours expedited (422.572(a)); Medicaid managed care: the state's time frame, no more than 7 calendar days for rating periods from 1 Jan 2026 (42 CFR 438.210(d)(1)(i)(B)) and 72 hours expedited ((d)(2)(i)). These are federal maximums; state law, plan documents and provider contracts can be shorter. Payer medical policies are quoted from the payer's published text with its source; licensed criteria sets (InterQual, MCG) are not included and never reproduced. CPT is an AMA code set: the tool takes codes as you enter them, removes descriptor rows from policy text and never suggests codes. Payer portals' terms of use generally forbid automation, so it never logs in to, fills or submits a portal. Not medical, coding or legal advice, and it never predicts the payer's decision.

### Write a note from my jottings (`jottings-note`)

For therapists in private practice with a notes backlog. After the session, give it what you already jot: typed bullets, a photo of your handwritten notes, or a dictation of up to a minute. Out comes a DAP, SOAP or BIRP progress note tied to your treatment-plan goals, where every sentence links to the jotting it came from. Nothing is added: numbers, diagnoses, moods, risk and recommendations must be in your jottings or the sentence is removed. A completeness check lists what payers audit (start and stop times, modality, interventions, goals, response, plan, your credential, and risk when you mentioned it) and marks what is missing instead of guessing. Catch-up mode drafts a backlog in one go. There is no recording, ever. The note is a draft you review, edit and sign.

- What it is: Therapy progress notes from what you already jot after a session: typed bullets, a photo of your handwriting or a one-minute dictation. Every sentence links to the jotting it came from, nothing is added, and what an auditor would miss is marked, never filled in. No session recording, ever.
- Who it's for: Therapists in private practice (LPC, LCSW, LMFT, psychologists) who write their own progress notes and are behind on them.
- Page: https://decosa.ai/clinics/jottings-note
- Last reviewed: 2026-09-30
- Key result: On 48 held-out synthetic sessions a blind reviewer found 25 of 701 draft units not supported by the jottings, against 751 of 1,574 for a plain prompt to the same model; synthetic sessions only.
- Key numbers (details and caveats: https://decosa.ai/metrics/jottings-note):
  - 113 / 117: Risk and safety statements carried, fresh blind split #3 (verbatim rule) (held out, n = 117)
  - 7 / 805: Draft units not supported, fresh held-out split (29 Sep, blind review) (held out, n = 805)
  - 0.0072 USD: Cost per note at list price (mean), held-out split (held out, n = 60)
  - 25.5 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: live. Models: Qwen3.8-27B (reads the photo, drafts, and checks every sentence with the grounding judge) · PaddleOCR-VL-1.6 (second reader for handwriting) · Qwen3-ASR-1.7B (a dictation of up to 60 s). Where: Self-host for real client notes; the hosted demo takes made-up sessions only.
- Verification: Every sentence cites its jottings and passes the number, clinical-claim and grounding checks; receipt per model call; signed record of hashes, no note text (receipt coverage: full). Hosted QA 2026-09-30: pass. Self-host QA 2026-09-28: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) or 1x RTX 5090 (32 GB, estimate) for Qwen3.8-27B; the photo's second reader needs about 3 GB more; typed jottings need nothing else
- Output: A draft DAP, SOAP or BIRP progress note with a cite per sentence, the completeness check, the removed sentences with reasons, plain text to paste, and a signed record.
- Page: https://decosa.ai/clinics/jottings-note · JSON: https://decosa.ai/use-cases/jottings-note.json
- Prompts: https://decosa.ai/prompts/jottings-note-hosted.md · https://decosa.ai/prompts/jottings-note-selfhost.md · https://decosa.ai/prompts/jottings-note-assemble.md
- Tiers:
  - Standard · one GPU for the model (hosted demo) (1x RTX PRO 6000 96 GB (measured) or 1x RTX 5090 32 GB (estimate)): Qwen3.8-27B drafts and checks; code checks times, dates, risk, numbers and clinical words. Photos are read by Qwen3.8 alone. This is what the hosted demo runs, on made-up sessions only.
  - Best · adds the second reader and dictation (1x RTX PRO 6000 96 GB: Qwen3.8 plus about 3 GB for the reader and the ASR loaded on demand (measured on our server)): Everything in Standard, plus PaddleOCR-VL-1.6 reading each handwritten line again (lines the readers disagree on are flagged) and Qwen3-ASR-1.7B for a dictation of up to a minute.
- Buyer facts:
  - Data retention: Nothing stored: the jottings, the photo or the dictation live in memory for the request. The signed record holds hashes of each jotting and sentence and receipt ids, never text; logs carry counts only.
  - What leaves the box: Hosted (made-up sessions only): every model call goes through our gateway to the GPU serving Qwen3.8-27B. Self-hosted on the direct route: nothing leaves the machine.
  - What it will not do: Record or take a session recording; drop, change or paraphrase a risk or safety statement you wrote (a sentence about one quotes your jotting; a draft that would lose one is blocked); add a diagnosis, a mood or mental state, a risk level or a recommendation; fill in a missing time, date or plan; sign or send anything. Your own ? and rule-out lines are held back for you.
  - Input formats: Typed jottings (one per line, up to 6,000 characters), a JPEG, PNG or WebP photo of handwritten jottings (up to 8 MB), or a dictation of up to 60 seconds; DAP, SOAP or BIRP; up to 10 treatment-plan goals; catch-up up to 30 typed sessions.
  - Typical run: One note: a few dozen model calls (each kept sentence gets a grounding check and a meaning check), a fraction of a cent at the gateway list price; under a minute on a busy shared gateway, seconds self-hosted on the direct route. A catch-up: a few minutes.
  - Pricing basis: A flat self-host licence per clinician (price not set). A cold-user therapist said about $25 a month.
- Samples:
  - 1. Cbt panic (`cbt-panic`): https://decosa.ai/clinics/jottings-note?sample=1&autorun=0
  - 2. Dbt photo (`dbt-photo`): https://decosa.ai/clinics/jottings-note?sample=2&autorun=0
  - 3. Couples gaps (`couples-gaps`): https://decosa.ai/clinics/jottings-note?sample=3&autorun=0
  - 4. Catch up (`catch-up`): https://decosa.ai/clinics/jottings-note?sample=4&autorun=0
- Regulatory note: Checked 28 Sep 2026. Illinois HB 1806, the Wellness and Oversight for Psychological Resources Act (225 ILCS 155; signed 4 Aug 2025, effective on becoming law), allows a licensed professional to use AI for administrative or supplementary support, which includes 'preparing and maintaining client records, including therapy notes' (Sec. 10); requires written informed consent when the client's session is recorded or transcribed (Sec. 15(b)), which this tool never does; and forbids letting AI make independent therapeutic decisions, communicate therapeutically with clients, generate recommendations or treatment plans without review, or detect emotions or mental states (Sec. 20(b)); penalty up to $10,000 per violation. Whether a post-session dictation by the clinician counts as transcribing the session is our reading, not settled: ask counsel. The element list follows CGS's LCD L34353 fact sheet (revised 26 May 2021) and First Coast's LCD L33252 and article A57520 (rev. 1 Jan 2025). Therapy notes are PHI under HIPAA; a cloud service that holds encrypted PHI is still a business associate, so real notes run self-hosted until Decosa has BAAs, and the hosted demo takes made-up sessions only. Not legal advice.

### Capture audit evidence from your admin screens (`evidence-runner`)

For the compliance lead who screenshots 30 to 150 admin settings screens every quarter for SOC 2 or CMMC. Name the screens once (2-step verification, password policy, audit logging, your own product's admin pages). The agent finds each one, read-only, and code reads the settings you named; you approve those values as the baseline. Every quarter after that it runs with no model: it opens each screen from its saved route, re-reads every setting, captures it on a fresh load stamped with the URL, the time and who is signed in, and lists what changed since the baseline, all in a signed certificate. The browser session blocks every write, so it cannot change a setting. It runs on your side, in your own signed-in browser or inside your network.

- What it is: Capture SOC 2 evidence screenshots from admin consoles without the quarterly clicking: name the screens once, approve the values, and every quarter it opens each screen, re-reads each setting, stamps the capture and lists what changed.
- Who it's for: Compliance leads at SaaS companies and the consultants and MSPs who prepare SOC 2 and CMMC evidence.
- Page: https://decosa.ai/tools/finance/evidence-runner
- Last reviewed: 2026-09-29
- Key result: On four held-out made-up consoles, quarterly runs gave the right verdict on 33 of 33 screens with no model calls, and setup read no wrong value in any run. No real tenant yet.
- Key numbers (details and caveats: https://decosa.ai/metrics/evidence-runner):
  - 33 / 33: Quarterly verdicts right on held-out consoles (no model) (test split, n = 33)
  - 25 / 26: Setup: screen found and every named setting read right, held-out after fixes (test split, n = 26)
  - 10 / 17: Setup, first look at two held-out consoles (test split, n = 17)
  - 7.5 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: preview. Models: Qwen3.8-27B finds each screen once (setup) and repairs routes when a console is redesigned; reading the settings, the quarterly replay, the stamps and the certificate are plain code. Where: On your side: a local helper attached to your own signed-in browser, or self-hosted in your network. The hosted demo runs on made-up consoles only.
- Verification: A decosa.test-run.v1 certificate per run (every setting re-read by code, stamped screenshots hashed in, signed); receipt per model call at setup (receipt coverage: partial). Hosted QA 2026-09-29: pass. Self-host QA 2026-09-29: pass.
- Hardware: The quarterly run needs only a CPU (headless Chromium). Setup and repairs call Qwen3.8-27B: 1x RTX PRO 6000 (96 GB) self-hosted, or the hosted gateway
- Output: An evidence pack: a stamped PNG per screen, a manifest (screen, controls, URL, UTC time, SHA-256, matches or changed) and a signed decosa.test-run.v1 certificate.
- Page: https://decosa.ai/tools/finance/evidence-runner · JSON: https://decosa.ai/use-cases/evidence-runner.json
- Prompts: https://decosa.ai/prompts/evidence-runner-hosted.md · https://decosa.ai/prompts/evidence-runner-selfhost.md · https://decosa.ai/prompts/evidence-runner-assemble.md
- Tiers:
  - Lite · quarterly runs only, no GPU (Any CPU): Replays an approved baseline spec: saved routes, every setting re-read in code, stamped captures, certificate, evidence pack, offline verify. No model; setup must come from the standard tier or a colleague's approved spec.
  - Standard · setup and repairs with Qwen3.8-27B (1x RTX PRO 6000 96 GB (measured) or the hosted gateway): The agent finds each named screen read-only (two runs must agree) and repairs saved routes after a redesign; code reads the values; a person approves the baseline.
- Alternates (smaller models for the same job):
  - Official read-only API exports (Graph, Policy API, AWS, Okta) (Any CPU), not built: The first route where an API covers a setting; not built into this tool yet (compliance platforms already do it).
- Buyer facts:
  - Data retention: Hosted demo: runs on made-up consoles are kept one hour for the session that made them, then deleted. Your own runs stay on your machine: the spec, the evidence packs and the signing key live where you run the helper.
  - What leaves the box: Quarterly runs: nothing (no model calls). Setup and repairs: screenshots and the element table of the admin screens the agent opens go to the model you choose (your own Qwen3.8-27B, or the hosted gateway with a key). Passwords never do: you sign in yourself.
  - Read-only by construction: The runner's tab blocks every write at the network level and no approval can release one; it refuses toggles, pickers and save-like clicks; every capture is a fresh load of the page.
  - Where it runs: On your side: a local helper attached to the browser you are already signed in to (it opens and gates one tab of its own), or inside your network. Your SSO, conditional access and device checks stay yours.
  - Typical quarter: A quarterly run takes seconds with no model calls; setup takes under a minute and a fraction of a cent per screen, once per console.
  - What it will not do: Say a control is effective, certify, or change a setting. It captures, compares with the baseline you approved, and signs.
- Samples:
  - 1. Keystone quarterly (`keystone-quarterly`): https://decosa.ai/tools/finance/evidence-runner?sample=1&autorun=0
  - 2. Keystone setup (`keystone-setup`): https://decosa.ai/tools/finance/evidence-runner?sample=2&autorun=0
  - 3. Northwind quarterly (`northwind-quarterly`): https://decosa.ai/tools/finance/evidence-runner?sample=3&autorun=0
  - 4. Harbor quarterly (`harbor-quarterly`): https://decosa.ai/tools/finance/evidence-runner?sample=4&autorun=0
- Regulatory note: Read 28 Sep 2026 (public terms only; not legal advice): Google Workspace (Terms, last modified 2 Sep 2026; Acceptable Use Policy, 13 Oct 2025), the AWS Customer Agreement (14 Aug 2026) and Acceptable Use Policy (1 Jul 2021), and the Okta Master Subscription Agreement (Q1 FY26) say nothing against a customer automating read-only views of its own tenant, and each offers read-only APIs (Cloud Identity Policy API, the IAM credential report and a SecurityAudit role, okta.policies.read) that should come first. Microsoft's Product Terms bar using an Online Service "to scrape or use other data extraction methods", so for Microsoft 365 and Entra this tool offers only witnessed capture (a person clicks, it stamps and signs) and Microsoft Graph is the route for settings. Okta's MSA also bars using the service for others as a managed service: run the tool as the customer's own. Auditors ask for a URL and date stamp on screenshots (Secureframe, 16 Jul 2026) and for the completeness and accuracy of system-generated evidence (PCAOB AS 1105.10). This tool captures and compares; it does not say a control is effective, and it is not an assessment.

### Put the visit into your EHR as drafts (`ehr-drafts`)

For clinicians in small practices who finish every visit by re-typing it into the EHR. Take what the visit produced (diagnosis codes, lab and imaging orders, medicine changes, the note) and an agent in your own browser enters it into the chart you have open as drafts: problems marked unconfirmed, prescriptions and orders saved but not sent, the note unsigned. It checks the chart is the visit's patient (name and date of birth read off the screen) before anything and again before every save, reads each medicine against the transcript and holds any that differ, never types a controlled substance, and stops before sign. Signing, transmitting, sending and e-prescribing are held in the browser's network layer whatever the button says. You get a checklist of what went where with the transcript line behind each value, and a copy list for what it could not enter. Where an EHR's terms forbid tools in your session, you get the copy list only.

- What it is: Works today with self-hosted OpenEMR: the visit's codes, orders, prescriptions and note typed in as unsigned drafts. You sign.
- Who it's for: Clinicians and practice staff in small practices who re-type each visit into the EHR.
- Page: https://decosa.ai/clinics/ehr-drafts
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/ehr-drafts):
  - 0 / 419: Typed or picked values wrong (read back from the EHR's database) (synthetic, n = 419)
  - 101 / 104: Items entered, of those the rules allow the agent to enter (synthetic, n = 104)
  - 0 / 20: Wrong-patient attempts that wrote to another chart (synthetic, n = 20)
  - 90.3 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: preview. Models: Qwen3.8-27B (drives the forms) · decosa-note-detail-checker (M17, reads each medicine against the transcript, CPU) · decosa-commit-detector (flags buttons that commit, CPU). Where: Works today with self-hosted OpenEMR, in the clinician's own browser; commercial EHRs through each vendor's official API (not yet available). Hosted demo: made-up patients only.
- Verification: Chart identifiers read off the screen before any entry and before every save; every typed value from the visit output; the form compared with the visit in code before its one save; sign/send/transmit held at the network layer; signed flight record of every step (receipt coverage: full). Hosted QA 2026-09-29: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the medicine check and the commit detector run on CPU
- Output: Drafts in your EHR (unsigned), a checklist of what went where with its transcript line, a copy list for the rest, the requests the gate held, and a signed record.
- Page: https://decosa.ai/clinics/ehr-drafts · JSON: https://decosa.ai/use-cases/ehr-drafts.json
- Prompts: https://decosa.ai/prompts/ehr-drafts-hosted.md · https://decosa.ai/prompts/ehr-drafts-selfhost.md
- Tiers:
  - Standard · one GPU for the model (1x RTX PRO 6000 96 GB (measured)): Qwen3.8-27B drives the forms; the medicine checker runs on CPU. What the eval measured.
- Buyer facts:
  - What it gives you: Drafts in your EHR (unsigned), a checklist of what went where with the transcript line behind each value, a copy list for what it could not enter, every request the gate held, and a signed record of the run.
  - What it never does: Sign, finalise, transmit an order, send, fax, e-mail or e-prescribe; draft a controlled substance; type a value the visit did not give (a quantity or refill count nobody said stays blank); edit or delete an existing entry.
  - Where it may type: Works today with self-hosted OpenEMR (open source). Commercial EHRs (eClinicalWorks, DrChrono, athenaOne and others) come through each vendor's official API, not yet available; until then they get the copy list. The hosted demo uses made-up patients only.
  - Data retention: The hosted check keeps nothing: the visit lives in memory for the request, logs carry counts only. The agent's signed record holds hashes of the patient identifiers, never their text, and stays with the practice.
  - What leaves the box: Self-hosted, nothing: the agent, the model and the checker run in your network. Our hosted model would see chart screens, so it waits for a BAA.
- Samples:
  - 1. Diabetes followup (`diabetes-followup`): https://decosa.ai/clinics/ehr-drafts?sample=1&autorun=0
  - 2. Uti (`uti`): https://decosa.ai/clinics/ehr-drafts?sample=2&autorun=0
  - 3. Insomnia controlled (`insomnia-controlled`): https://decosa.ai/clinics/ehr-drafts?sample=3&autorun=0
  - 4. Wrong dose planted (`wrong-dose-planted`): https://decosa.ai/clinics/ehr-drafts?sample=4&autorun=0
- Regulatory note: Checked 28 Sep 2026; not legal advice. FDA: suggested codes and orders a clinician can review, each with the transcript line behind it and nothing time-critical, fit the non-device clinical decision support criteria (FD&C Act 520(o)(1)(E); FDA guidance of 29 Jan 2026). DEA: controlled-substance e-prescriptions are signed only by the prescriber through their own two-factor authentication (21 CFR 1311.135, 1311.140); this tool never drafts one and never touches a sign screen. HIPAA: hosting or relaying chart data makes Decosa a business associate (45 CFR 160.103); no BAA is signed yet, so the hosted demo is synthetic only and real use runs the model in the practice's network. EHR terms: eClinicalWorks, DrChrono and TherapyNotes forbid tools in the customer's session in their own terms (read 28 Sep 2026); Practice Fusion, Tebra and SimplePractice restrict it or could not be read; those EHRs get the copy list. CPT codes are not shown or typed (AMA licence pending); diagnoses are ICD-10-CM and tests carry LOINC codes. Texas SB 1188 requires telling patients when AI is used for diagnosis or treatment recommendations.

### Read a demand before the deadline (`demand-reader`)

For bodily-injury adjusters at carriers, TPAs and self-insureds. Give it an injury demand package: the lawyer's letter, the medical bills and the records. It shows whether the letter is a time-limited demand and the respond-by date the letter sets, computed in code from its own words and the day you received it, next to the statute's minimum where the state is encoded (California, Georgia, Missouri, Montana, Utah; Florida's safe harbour). Every condition of acceptance is quoted with its page, the state's required elements are checked, each bill line is tied to its page and summed, and the total is reconciled with what the letter claims. It lists providers billed with no records, records with no bill, gaps in treatment, earlier care, and notes on the letter's citations (a statute from another state, a case that isn't at its cite). It never values the claim or recommends paying.

- What it is: A first read of a time-limited demand: it shows the respond-by date the letter sets and the statute's minimum, every condition of acceptance with its page, and the medical specials tied to the bills, and it never values the claim.
- Who it's for: Bodily-injury adjusters and claim leads at carriers, TPAs and self-insured claims units.
- Page: https://decosa.ai/tools/insurance/demand-reader
- Last reviewed: 2026-09-29
- Key result: On a held-out set of 8 synthetic packages it got the respond-by date and the statute's minimum right on all 7 time-limited demands, found all 31 conditions of acceptance and summed the bills to within $1 on all 8; one author wrote the packages and the set is small.
- Key numbers (details and caveats: https://decosa.ai/metrics/demand-reader):
  - 7 of 7: Respond-by date exact (test split, n = 7)
  - 7 of 7: Statute's minimum exact (test split, n = 7)
  - 31 of 31: Conditions of acceptance found (test split, n = 31)
  - 161.0 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: preview. Models: Qwen3.8-27B (reads the letter, the bills and each record page; the grounding judge checks every sentence it writes). Where: Self-host or confidential access for real claim files; the hosted demo takes made-up packages only.
- Verification: Every quote and amount found on the page it cites; dates and totals computed in code; receipt per model call; signed record of page hashes and every item (receipt coverage: full). Hosted QA 2026-09-29: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the deadline rules and the tie-out need no GPU
- Output: The respond-by date and statute check, the conditions with pages, the required elements, the specials tied to the bills, records against bills, gaps, earlier care, notes on the citations, and a signed record.
- Page: https://decosa.ai/tools/insurance/demand-reader · JSON: https://decosa.ai/use-cases/demand-reader.json
- Prompts: https://decosa.ai/prompts/demand-reader-hosted.md · https://decosa.ai/prompts/demand-reader-selfhost.md · https://decosa.ai/prompts/demand-reader-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B reads the letter, the bills and the records with quotes; the dates, the statute rules, the sums and the matching are code. About 20-25 receipted calls per package.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, multi-claimant files.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Claim files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: Time-limited or not; the respond-by and payment dates from the letter's words; the statute's minimum (CA, GA, MO, MT, UT; FL's safe harbour); each condition with its page; the state's required elements; bill lines on their pages, summed and reconciled per provider; records vs bills; gaps; earlier care; citations from another state, cases not at their cite, and misstatements of the law.
  - What it never does: It never values the claim, recommends paying, offering or accepting, or judges the letter's credibility. A sentence that would do so is removed. The adjuster decides.
  - Data retention: Nothing kept on the server: the package lives in memory for the request; logs carry counts only. You keep the signed record with the claim file.
  - What leaves the box (hosted demo): The pages go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Case citations (only the cite) are looked up in the local index and, on a miss, CourtListener. The hosted demo is for made-up packages; real files run on your own box or confidential access.
  - Model calls per package: About 20-25: the letter (three reads), one per bill page, one per record page, and the grounding checks.
  - Cost per package: A few cents or less per package on average at the gateway list price (held-out test). Each run shows its own measured cost.
  - States encoded: California, Georgia, Missouri, Montana, Utah; Florida's safe harbour and disclosure rules. Texas and Louisiana checked with no statute found. Any other state: the letter's own deadline, and 'not encoded: check with counsel'.
- Samples:
  - 1. Ga policy limits (`ga-policy-limits`): https://decosa.ai/tools/insurance/demand-reader?sample=1&autorun=0
  - 2. Ga short deadline (`ga-short-deadline`): https://decosa.ai/tools/insurance/demand-reader?sample=2&autorun=0
  - 3. Ca mailed short (`ca-mailed-short`): https://decosa.ai/tools/insurance/demand-reader?sample=3&autorun=0
  - 4. Mt holiday (`mt-holiday`): https://decosa.ai/tools/insurance/demand-reader?sample=4&autorun=0
  - 5. Fl fabricated case (`fl-fabricated-case`): https://decosa.ai/tools/insurance/demand-reader?sample=5&autorun=0
  - 6. Oh not encoded (`oh-not-encoded`): https://decosa.ai/tools/insurance/demand-reader?sample=6&autorun=0
  - 7. Tx no deadline (`tx-no-deadline`): https://decosa.ai/tools/insurance/demand-reader?sample=7&autorun=0
- Regulatory note: Time-limited demand statutes read 28-29 Sep 2026 and encoded with their sources: California CCP 999-999.5 (SB 1155, demands transmitted from 1 Jan 2023: at least 30 days from transmission by email, fax or certified mail, 33 by mail); Georgia O.C.G.A. 9-11-67.1 as amended by SB 83 (2024: at least 30 days from receipt; payment at least 40 days); Missouri RSMo 537.058 (at least 90 days from receipt); Montana MCA 33-18-251 (2023: at least 60 days from receipt, rolled to the next business day); Utah Code 31A-22-323 (S.B. 74, 2026: at least 30 days); Florida 624.155(4) (90-day safe harbour after notice) and 627.4137 (30-day disclosure). Texas (common-law Stowers) and Louisiana were checked and have no such statute; every other state returns "not encoded: check with counsel". Deadlines are reminders from the letter's own words, not legal advice. The NAIC AI model bulletin asks insurers to govern third-party AI: the signed record lists every model call.

### Check a certificate request (`certificate-check`)

For commercial-lines account managers at independent agencies. When a general contractor, landlord or lender asks for a certificate of insurance and attaches a long contract, it pulls every insurance requirement out of the contract and the request email and checks each against the insured's policy papers: met by the policy (with the wording quoted), needs an endorsement or a policy change from the carrier, a certificate can't do this, or the papers don't say. It drafts the request to the carrier, a note to the holder, and plain text to paste into your own licensed certificate system. A certificate can't add coverage, so it never suggests putting anything on the certificate that the policy doesn't have. No ACORD form is drawn or reproduced.

- What it is: A check of certificate of insurance requirements: it reads the contract's insurance article and the holder's email, checks each requirement against the insured's policy papers with the wording quoted, and drafts the carrier request and the note to the holder.
- Who it's for: Commercial-lines account managers and CSRs at independent insurance agencies.
- Page: https://decosa.ai/tools/insurance/certificate-check
- Last reviewed: 2026-09-30
- Key result: On a test set of 9 synthetic requests (120 requirements found of 124) it got the status right on 109, and 102 of its 103 'met' calls were right; one author wrote all the papers, and the test set was run three times.
- Key numbers (details and caveats: https://decosa.ai/metrics/certificate-check):
  - 109 of 120 (0.91): Status right on found requirements (test split, n = 120)
  - 120 of 124 (0.97): Requirements found (test split, n = 124)
  - 102 of 103: 'Met' calls that were right (test split, n = 103)
  - 33.3 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: preview. Models: Qwen3.8-27B (reads the contract and the policies; the grounding judge checks every reason it writes). Where: Hosted demo on made-up papers; self-host or confidential access for client files.
- Verification: Every requirement quoted from the contract or email and every 'met' quoted from the policy; limits compared in code; receipt per model call; signed record for the E&O file (receipt coverage: full). Hosted QA 2026-09-30: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B
- Output: A requirements table with statuses and quotes, a carrier request, a holder note, text to paste into your certificate system, and a signed record.
- Page: https://decosa.ai/tools/insurance/certificate-check · JSON: https://decosa.ai/use-cases/certificate-check.json
- Prompts: https://decosa.ai/prompts/certificate-check-hosted.md · https://decosa.ai/prompts/certificate-check-selfhost.md · https://decosa.ai/prompts/certificate-check-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B lists the requirements and checks each against the policies with quotes; limits, the umbrella rule and 'met needs a quote' are code. 26-29 receipted calls per request.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, multi-claimant files.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Claim files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: Limits (each occurrence, aggregates, auto, umbrella, employers liability, professional, pollution, deductibles), additional insured for ongoing and completed operations, primary and non-contributory, waivers of subrogation per line, notice of cancellation, per-project aggregate, umbrella follow-form, holder wording, carrier rating, and asks a certificate can't do.
  - What it never does: It never issues a certificate, never draws an ACORD form, and gives no coverage opinion. A 'met' always shows the policy wording it rests on; a licensed agent decides.
  - Data retention: Nothing kept on the server: the contract, request and policies live in memory for the request; logs carry counts only. You keep the signed record for the E&O file.
  - What leaves the box (hosted demo): The papers go to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. Self-hosted, nothing leaves. The hosted demo is for made-up papers.
  - Model calls per request: One to list the requirements, one check per requirement, and one grounding check per reason. Each run shows its own count.
  - Cost per request: Charged at the gateway list price. The page header shows the cost measured on production; each run shows its own.
- Samples:
  - 1. Tx framing gaps (`tx-framing-gaps`): https://decosa.ai/tools/insurance/certificate-check?sample=1&autorun=0
  - 2. Nj umbrella not allowed (`nj-umbrella-not-allowed`): https://decosa.ai/tools/insurance/certificate-check?sample=2&autorun=0
  - 3. Fl condo email asks (`fl-condo-email-asks`): https://decosa.ai/tools/insurance/certificate-check?sample=3&autorun=0
  - 4. Ca city notice (`ca-city-notice`): https://decosa.ai/tools/insurance/certificate-check?sample=4&autorun=0
  - 5. Fl clean (`fl-clean`): https://decosa.ai/tools/insurance/certificate-check?sample=5&autorun=0
- Regulatory note: State certificate-of-insurance laws read 28-29 Sep 2026 (NCOIL Certificates of Insurance Model Act lineage): Texas Ins. Code ch. 1811, N.Y. Ins. Law § 502, Minn. Stat. § 60A.39, 2017 Wis. Act 241, N.H. RSA 412:6-b, La. R.S. 22:890 and Utah Code §§ 31A-22-1701 to -1707 say a certificate may not amend, extend or alter coverage; Georgia, Connecticut and Tennessee from secondary sources. ACORD forms are licensed (acord.org forms FAQ, read 28 Sep 2026): this tool renders no ACORD form and outputs plain text for the agency's own licensed system. Not a coverage opinion and not legal advice; a licensed agent decides and issues.

### Check a bank-detail change before you pay (`bank-change-check`)

For AP clerks, controllers and the bookkeepers who pay vendors for clients. Forward the email that asks to change a vendor's bank details, with your vendor file. It lists the warning signs with their evidence: a look-alike domain, replies routed elsewhere, failed sender checks, a sender who isn't on file, a new phone number, a bank in another country, an account holder who isn't the vendor, pressure, secrecy, "don't call". It always says to hold the change and call the number already on file, never the one in the email, because a real vendor's hacked mailbox passes every check the email allows. You record the call-back and a second person approves, in a signed record. It never says an email is safe and never releases or blocks a payment.

- What it is: A check for vendor bank change fraud: it reads the email that asks to change a vendor's bank details against your own vendor file, lists every warning sign with its evidence, and makes the call-back to the number on file part of a signed record with a second approver.
- Who it's for: AP clerks, controllers and bookkeepers who pay vendors by ACH or wire.
- Page: https://decosa.ai/tools/finance/bank-change-check
- Last reviewed: 2026-09-29
- Key result: On a held-out set of 19 synthetic emails (9 frauds, 8 genuine bank changes, 2 with no change) it flagged 9 of 9 frauds and 0 of 8 genuine changes; the set is small and written by one author, so the confidence interval is wide.
- Key numbers (details and caveats: https://decosa.ai/metrics/bank-change-check):
  - 9 of 9: Fraud flagged (score 3 or more) (test split, n = 9)
  - 0 of 8: Genuine changes flagged (test split, n = 8)
  - 19 of 19: Bank-change requests detected (test split, n = 19)
  - 8.1 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: preview. Models: Qwen3.8-27B (quotes the pressure and secrecy signs; the header, domain and bank checks are code). Where: Hosted demo on made-up emails; self-host or confidential access for your real AP inbox.
- Verification: Header and vendor-file checks in code; every model sign quoted from the email; receipt per model call; signed check and call-back records (receipt coverage: full). Hosted QA 2026-09-29: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the header and vendor-file checks run on CPU
- Output: Warning signs with evidence, the call-back plan to the number on file, and a signed record with the call-back and a second approver.
- Page: https://decosa.ai/tools/finance/bank-change-check · JSON: https://decosa.ai/use-cases/bank-change-check.json
- Prompts: https://decosa.ai/prompts/bank-change-check-hosted.md · https://decosa.ai/prompts/bank-change-check-selfhost.md · https://decosa.ai/prompts/bank-change-check-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same pipeline on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B quotes the content signs; the header, domain and vendor-file checks and the call-back plan are code. One model call per email, receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, multi-claimant files.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Claim files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: Sender, Reply-To and Return-Path against your vendor file; SPF, DKIM and DMARC results; look-alike domains; phone numbers and addresses not on file; the new bank's country (IBAN, SWIFT) and account holder against the vendor; pressure, secrecy, "don't call" and redirected payments, quoted.
  - What it never does: It never says an email is safe, never edits your vendor file, and never releases, blocks or schedules a payment. The call-back must go to the number on file; the record refuses any other number. A confirmed change needs the vendor's last 4 digits of the new account: if they differ from the email, it is signed as "details differ" and stays on hold. The second approver's name must differ from the caller's; that is a name check only (it catches the same name written another way, such as initials, word order or an email address, not two people who sign for each other).
  - Data retention: Nothing kept on the server: the email and vendor file live in memory for the request; logs carry counts only. You keep the signed records.
  - What leaves the box (hosted demo): The email text goes to Qwen3.8-27B through the Decosa API, whose receipts hold hashes, not text. No domain look-ups are made. Self-hosted, nothing leaves. The hosted demo is for made-up emails.
  - Model calls per email: One (none in rules-only mode).
  - Cost per email: A fraction of a cent per email on average at the gateway list price (held-out test). Each run shows its own measured cost.
  - Also used in: Vendor onboarding (HR and procurement): a new vendor's first bank details go through the same check and call-back.
- Samples:
  - 1. Lookalike urgent (`lookalike-urgent`): https://decosa.ai/tools/finance/bank-change-check?sample=1&autorun=0
  - 2. Misspelled domain urgent (`misspelled-domain-urgent`): https://decosa.ai/tools/finance/bank-change-check?sample=2&autorun=0
  - 3. Calm tld swap (`calm-tld-swap`): https://decosa.ai/tools/finance/bank-change-check?sample=3&autorun=0
  - 4. Real mailbox redirect (`real-mailbox-redirect`): https://decosa.ai/tools/finance/bank-change-check?sample=4&autorun=0
  - 5. Genuine bank acquisition (`genuine-bank-acquisition`): https://decosa.ai/tools/finance/bank-change-check?sample=5&autorun=0
  - 6. Genuine german merger (`genuine-german-merger`): https://decosa.ai/tools/finance/bank-change-check?sample=6&autorun=0
  - 7. No change phishing (`no-change-phishing`): https://decosa.ai/tools/finance/bank-change-check?sample=7&autorun=0
- Regulatory note: Nacha Operating Rules, 2024 risk-management amendments (nacha.org rule pages read 29 Sep 2026): non-consumer Originators must have risk-based processes and procedures reasonably intended to identify ACH entries initiated due to fraud, including entries authorized under False Pretenses ("the inducement of a payment by a Person misrepresenting ... that Person's identity"), reviewed at least annually. Phase 1 from 20 Mar 2026 (ODFIs and originators with 6 million or more ACH entries in 2023); Phase 2 from 19 Jun 2026 (all other non-consumer originators). The signed call-back record documents one such step; it is not a compliance determination, and this is not legal advice.

### Turn a family interview into a film (`family-film`)

Record Grandma telling her stories, in her language, and add her old photos. Studio hears who is speaking, finds the chapters of her life and her best lines (each traced to the second she said it), subtitles her in English under her own words, reads the backs of her photos, and cuts a trailer and a film: her real voice over her real photos, with maps, dates and a quiet score. A book view puts a QR code next to every quote that plays her saying it. Her consent is recorded in her own words and her own link takes it back. No model ever draws, animates or changes a photo, and there is no voice clone.

- What it is: A family interview film from one phone recording: Studio finds the chapters and her best lines, subtitles her in English under her own words, and cuts a trailer, a film and a book in her real voice over her real photos.
- Who it's for: Grandchildren and adult children who want a parent's or grandparent's stories kept, in their own voice and language.
- Page: https://decosa.ai/apps/family-film
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/family-film):
  - 46 / 46: Shown quotes inside her own turn, at the right time (dev (tuned on), n = 46)
  - 44 / 46: Shown quotes that match her words (fuzzy 0.9) (dev (tuned on), n = 46)
  - 2 / 48: Quotes dropped by the re-hearing check (dev (tuned on), n = 48)
  - 40.8 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: live. Models: Qwen3-ASR-1.7B · ECAPA-TDNN (speakers) · Qwen3.8-27B · Hy-MT2-7B · PaddleOCR-VL-1.6 (photo backs) · ACE-Step 1.5 (score). Where: Hosted, private by default; self-host for recordings that shouldn't leave the family.
- Verification: Every quote re-heard at its timestamp; meaning check on every English line; receipts per model call; C2PA in every film (receipt coverage: partial). Hosted QA 2026-09-29: pass.
- Hardware: CPU for the edit; speech recognition, translation and the photo-back reader share one GPU (no video model)
- Output: A trailer and a film in her voice with English subtitles under her own words, a book with QR codes that play her quotes, and private links for the family.
- Page: https://decosa.ai/apps/family-film · JSON: https://decosa.ai/use-cases/family-film.json
- Prompts: https://decosa.ai/prompts/family-film-hosted.md · https://decosa.ai/prompts/family-film-selfhost.md · https://decosa.ai/prompts/family-film-assemble.md
- Tiers:
  - Lite · chapters, quotes, subtitles and the film, no photo reading (One GPU for speech recognition and translation (Hy-MT2 served at 18 GB), the text model (about 20 GB, or a remote server), CPU for the edit): Everything except reading the backs of photos: add places and years by hand instead.
  - Standard · the hosted demo, with photo backs read (1x RTX PRO 6000 Blackwell 96 GB (shared) for speech, translation and the reader; the text model on a second card or remote; CPU for the edit): Adds the face guard and the document reader, so handwriting on the backs of photos dates and places each picture.
- Buyer facts:
  - Data retention: The interview, photos, transcript and films stay in your private Studio project until you delete it or she takes her consent back with her link (which deletes everything). Her consent entry expires after three years. Logs hold ids, counts and hashes, never her words.
  - What leaves the box: On the hosted route, speech, translation, photo backs and the text model all run on Decosa's hosted service. Self-hosted, nothing leaves your machine.
  - Her voice, her photos: No voice clone and no generated faces: the film uses her real voice and shows her real photos. Only pictures with no face in them (the backs of photos) are read by a model.
  - Cost per film: A few cents or less per film in the eval, at list prices for the text model and GPU time.
  - Typical time: Chapters and a first trailer arrive within about a minute of upload for a short interview, longer for a longer one (measured).
  - Outputs: A trailer and a film (MP4, 16:9) with English subtitles under her words and a C2PA credential; subtitle files; a book page per chapter with a QR code per quote that plays her saying it; private share links you can turn off.
- Samples:
  - 1. Rosa (`rosa`): https://decosa.ai/apps/family-film?sample=1&autorun=0
- Regulatory note: Not legal advice; written 29 Sep 2026. Consent: the storyteller records her consent in her own words before anything else happens; the meaning of what she said is checked (a refusal or a hesitation stops the project), and the consent is kept in the consent ledger with her own revocation link, which deletes the project. Her voice: the ledger keeps a voiceprint of her consent recording so the film can tell her voice from the interviewer's; some laws treat voiceprints as biometric data (for example the Illinois Biometric Information Privacy Act, 740 ILCS 14, which asks for a written release), so self-host or get the consent your law requires. Photos: fronts of photos are shown, panned and captioned only; nothing is generated from them. Children: children are paused in Studio. A family photo in which anyone looks under 18, or whose caption or writing says so, is refused and not kept, and Studio never makes a child the subject of a generated picture or video. A family mode, where a parent verifies and consents for their children, is planned. Every film carries a C2PA credential; nothing is used for training.

### Make a music video starring you (`music-video-starring-you`)

Record a 15-second consent clip (a selfie reading a sentence that ends in three fresh words), add your track and a line of direction, and get 8 to 12 shots of you cut on the beat: the beat grid and sections of the song, a shot list per section you can edit, a re-roll on any shot, a 16:9 and a 9:16 export and a Spotify Canvas loop. Your face comes only from your consent clip: there is no photo upload, and adults only. Every frame says AI video and the end card credits the music to you. Moving shots run on MiniMax H3 when a render GPU is attached; otherwise the shots are a storyboard of drawn stills with camera moves.

- What it is: Record a 15-second consent clip, add your track, and get a music video of you cut on the beat, in 16:9, 9:16 and a Canvas loop.
- Who it's for: Independent artists releasing a single who want a first video with themselves in it.
- Page: https://decosa.ai/apps/music-video-starring-you
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/music-video-starring-you):
  - 27 / 27: Planned cuts found on the beat in the exports (synthetic, n = 27)
  - 0 / 30: Storyboard frames flagged by the frame safety check (synthetic, n = 30)
  - 88.2 s: Storyboard video done (p50) (synthetic, n = 3)
  - 88.2 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: preview. Models: MiniMax H3 (reference mode, Turbo v4) · FLUX.2 klein 4B (storyboard) · Qwen3.8-27B (shot list, checks) · Qwen3-ASR-1.7B (consent read-back). Where: Hosted (moving shots when a render GPU is attached); self-host on a 96 GB card with your own H3 licence.
- Verification: Consent clip checked (words heard, one live face, an adult); frame safety and likeness checks; cuts measured on the beat; C2PA in every export (receipt coverage: partial). Hosted QA 2026-09-29: pass.
- Hardware: 1x 96 GB card for MiniMax H3 shots (fp8, about 50 GB); the storyboard runs on FLUX.2 klein (about 6 GB) and the edit on CPU
- Output: A music video in 16:9 and 9:16 and a Spotify Canvas loop, labelled AI video, crediting the music to you, with a content credential in each file.
- Page: https://decosa.ai/apps/music-video-starring-you · JSON: https://decosa.ai/use-cases/music-video-starring-you.json
- Prompts: https://decosa.ai/prompts/music-video-starring-you-hosted.md · https://decosa.ai/prompts/music-video-starring-you-selfhost.md · https://decosa.ai/prompts/music-video-starring-you-assemble.md
- Tiers:
  - Standard · the hosted demo, a storyboard cut on the beat (One card with about 6 GB free for the stills, the text model (about 20 GB, or remote), CPU for the edit): A drawn still per shot from your consent clip, with a camera move, cut on the beat: a planning cut and a keepsake, not yet a moving video.
  - Best · moving shots on MiniMax H3, one 96 GB card (1x RTX PRO 6000 96 GB with about 60 GB free (fp8 H3, peak 49.6-52 GiB measured)): The same shot list with each shot moving, from your consent-clip references. Not ready: blind testers rejected the moving shots, and the likeness gate keeps about half. Hosted once a render GPU is attached; self-host with your own H3 licence.
- Buyer facts:
  - Data retention: The consent clip, face references, track and videos stay in your private Studio project until you delete it or the performer takes consent back with their own link (which deletes their references and videos). The consent entry expires after one year. Logs hold ids, counts and hashes.
  - What leaves the box: On the hosted route, speech, the text model and renders run on Decosa's hosted service. Self-hosted, nothing leaves your machine.
  - Whose face: Only your face, from your own live consent clip, adults only. There is no photo upload anywhere, so it can't be pointed at someone else.
  - Cost per video: A few cents per storyboard video in the eval; a moving MiniMax H3 shot costs a few cents of GPU time.
  - Typical time: Every picture is checked before you see it, so the shots appear together with the storyboard video, a minute or two after you start (measured); each moving H3 shot takes about a minute more.
  - Outputs: MP4s in 16:9 and 9:16 and a Spotify Canvas loop, with an AI video label on every frame, an end card crediting the music and a C2PA credential; private share links you can turn off.
- Samples:
  - 1. Artist (`artist`): https://decosa.ai/apps/music-video-starring-you?sample=1&autorun=0
- Regulatory note: Not legal advice; written 29 Sep 2026. Consent: a live consent clip with three fresh words, one moving face and an adult check (fails closed) goes into the consent ledger for the purpose music_video only, with the performer's own revocation link, which deletes their references and videos. There is no photo upload anywhere in the tool, so it can't be pointed at someone else's face. Children: Studio keeps children out of every video: a minor can't be enrolled for this purpose (tested), and memories that describe a child are refused. Disclosure: every frame carries a visible AI video label and every export a C2PA credential; EU AI Act (Regulation (EU) 2024/1689) Art. 50(2) asks providers to mark generated video in a machine-readable way from 2 Aug 2026 (EUR-Lex link under Tools). Use each platform's own AI-content toggle as well. Music: you confirm you have the rights to the track, and the end card credits it; the statement is not verified.

### Make a film of your story together (`our-story-film`)

A 60-second film of the two of you, cut to your song. Each partner records their own consent clip, you both write three memories in your own words, and Studio makes an opening shot, the three memories in order with a line from your own words on screen, and a closing shot of the two of you now, cut on the beat. Both faces come only from your consent clips; adults only; nobody else in focus. Private by default: share it with the people you name, or download it.

- What it is: A 60-second film of the two of you, told through three of your own memories and cut to your song.
- Who it's for: Couples and partners who want a private anniversary or wedding gift told through their own memories.
- Page: https://decosa.ai/apps/our-story-film
- Last reviewed: 2026-09-29
- Key numbers (details and caveats: https://decosa.ai/metrics/our-story-film):
  - 30 / 30: Planned cuts found on the beat in the exports (synthetic, n = 30)
  - 0 / 33: Storyboard frames flagged by the frame safety check (synthetic, n = 33)
  - 142.5 s: Storyboard video done (p50) (synthetic, n = 3)
  - 142.5 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: preview. Models: MiniMax H3 (reference mode, two subjects) · FLUX.2 klein 4B (storyboard) · Qwen3.8-27B · Qwen3-ASR-1.7B. Where: Hosted (moving shots when a render GPU is attached); self-host with your own H3 licence.
- Verification: Two consent clips checked; frame safety and likeness checks; cuts measured on the beat; C2PA in every export (receipt coverage: partial). Hosted QA 2026-09-29: pass.
- Hardware: 1x 96 GB card for MiniMax H3 shots; the storyboard runs on FLUX.2 klein and the edit on CPU
- Output: A 60-second film of the two of you in 16:9 and 9:16, labelled AI video, with a content credential.
- Page: https://decosa.ai/apps/our-story-film · JSON: https://decosa.ai/use-cases/our-story-film.json
- Prompts: https://decosa.ai/prompts/our-story-film-hosted.md · https://decosa.ai/prompts/our-story-film-selfhost.md · https://decosa.ai/prompts/our-story-film-assemble.md
- Tiers:
  - Standard · the hosted demo, a storyboard cut on the beat (One card with about 6 GB free for the stills, the text model (about 20 GB, or remote), CPU for the edit): A drawn still per shot from your consent clip, with a camera move, cut on the beat: a planning cut and a keepsake, not yet a moving video.
  - Best · moving shots on MiniMax H3, one 96 GB card (1x RTX PRO 6000 96 GB with about 60 GB free (fp8 H3, peak 49.6-52 GiB measured)): The same shot list with each shot moving, from your consent-clip references. Not ready: blind testers rejected the moving shots, and the likeness gate keeps about half. Hosted once a render GPU is attached; self-host with your own H3 licence.
- Buyer facts:
  - Data retention: The consent clip, face references, track and videos stay in your private Studio project until you delete it or the performer takes consent back with their own link (which deletes their references and videos). The consent entry expires after one year. Logs hold ids, counts and hashes.
  - What leaves the box: On the hosted route, speech, the text model and renders run on Decosa's hosted service. Self-hosted, nothing leaves your machine.
  - Whose face: Only the two faces from your own consent clips, adults only. There is no photo upload anywhere, so it can't be pointed at someone else.
  - Cost per video: Several cents per storyboard video in the eval; a moving MiniMax H3 shot costs a few cents of GPU time.
  - Typical time: Every picture is checked before you see it, so the shots appear together with the storyboard video, a few minutes after you start (measured); each moving H3 shot takes more than a minute.
  - Outputs: MP4s in 16:9 and 9:16 and a Spotify Canvas loop, with an AI video label on every frame, an end card crediting the music and a C2PA credential; private share links you can turn off.
- Samples:
  - 1. Couple (`couple`): https://decosa.ai/apps/our-story-film?sample=1&autorun=0
- Regulatory note: Not legal advice; written 29 Sep 2026. Consent: a live consent clip with three fresh words, one moving face and an adult check (fails closed) goes into the consent ledger for the purpose music_video only, with the performer's own revocation link, which deletes their references and videos. There is no photo upload anywhere in the tool, so it can't be pointed at someone else's face. Children: Studio keeps children out of every video: a minor can't be enrolled for this purpose (tested), and memories that describe a child are refused. Disclosure: every frame carries a visible AI video label and every export a C2PA credential; EU AI Act (Regulation (EU) 2024/1689) Art. 50(2) asks providers to mark generated video in a machine-readable way from 2 Aug 2026 (EUR-Lex link under Tools). Use each platform's own AI-content toggle as well. Music: you confirm you have the rights to the track, and the end card credits it; the statement is not verified.

### Make a settlement video (`settlement-video`)

For plaintiff's personal-injury lawyers taking a case to mediation. Start from the matter's medical chronology (every entry already cited to a page and box), the itemized bills and the client's signed statement and photos. An open model drafts a 7-scene script and names the entries each line rests on; the cites are attached in code. Every line is checked against what it cites, every date, count and amount is checked against the record, and the bills are added up in code, with a statement's wrong total, duplicate charges, charges before the injury and charges with no matching record flagged. The lawyer edits, reorders and approves each scene and picks the narration: their own recording, their voice with their recorded consent, or a stock house voice. Frames are drawn from the exhibits (no video-generation model), and the MP4 ships with content credentials and a cite sheet. A drafting aid; the lawyer decides.

- What it is: A settlement video tool for personal-injury lawyers: it turns the matter's chronology, bills, the client's statement and photos into a narrated video where every fact on screen is cited to its exhibit page, and nothing renders until the lawyer approves each scene.
- Who it's for: Plaintiff's personal-injury lawyers and paralegals preparing a case for mediation.
- Page: https://decosa.ai/legal/settlement-video
- Last reviewed: 2026-09-30
- Key result: On six held-out synthetic matters, 376 of 376 cites of rendered lines pointed at the right page and the bills tied out to the cent on all six; real records are not measured yet.
- Key numbers (details and caveats: https://decosa.ai/metrics/settlement-video):
  - 143 / 144: Cites on the right page, matters written blind by another agent (held out, n = 144)
  - 376 / 376: Cites on the right page, rendered lines (test split, n = 376)
  - 53 / 53: Dates in rendered lines right (test split, n = 53)
  - 103.5 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: live. Models: Qwen3.8-27B (script and per-line grounding) · document reader (bills and statement) · Kokoro-82M house voices or Chatterbox (the attorney's consented voice), CPU · Pillow + ffmpeg renderer (no generative video model). Where: Self-host or the confidential API for real client files; the hosted demo takes synthetic matters only.
- Verification: Receipt per model call; every line's cites attached in code and checked; bills tied out in code; consent-ledger decisions for the voice, the photos and the statement; C2PA on the MP4; signed hash-chained record (receipt coverage: full). Hosted QA 2026-09-30: pass. Self-host QA 2026-09-29: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B and the document reader (as for the medical chronology); the voices, the checks and the renderer run on CPU
- Output: An MP4 settlement video with content credentials, a cite sheet PDF listing every line with its exhibit, page and box, and a signed decosa.record.v1.
- Page: https://decosa.ai/legal/settlement-video · JSON: https://decosa.ai/use-cases/settlement-video.json
- Prompts: https://decosa.ai/prompts/settlement-video-hosted.md · https://decosa.ai/prompts/settlement-video-selfhost.md · https://decosa.ai/prompts/settlement-video-assemble.md
- Tiers:
  - Lite · captions or your own recording, no voice models (1x GPU with about 26 GB free for Qwen3.8-27B and the reader, plus 8+ CPU cores for the frames): The same script, checks, tie-out, renderer, cite sheet and record, narrated by the attorney's own recording or captions only; no Kokoro or Chatterbox runtime to install.
  - Standard · reader, model, house voices and the consented clone (hosted demo) (1x RTX PRO 6000 96 GB (measured on shared cards) and 32 CPU cores): What the hosted demo runs on synthetic matters: the document reader for the bills and statement, Qwen3.8-27B for the script and the per-line checks, Kokoro-82M house voices and Chatterbox for the attorney's consented voice on CPU, and the renderer.
- Buyer facts:
  - Data retention: Drafts, page images, photos and videos are kept for two hours on the server that made them, then deleted (Delete now removes a render at once). The signed record holds hashes and receipt ids, never record text, names or photos; logs carry counts, timings and ids only. The hosted demo takes synthetic matters only.
  - What leaves the box: Self-hosted on the direct route: nothing. The model, the document reader, the voices and the renderer run on the same machine. Hosted: record text and page images go through the Decosa API, and only synthetic matters are accepted.
  - What every line on screen carries: The exhibit, page and box it rests on, attached in code, and a check against the cited text; bill amounts are sums in code. Lines that fail are held back until the lawyer fixes or removes them; only the closing scenes may carry the lawyer's own words, labelled as such.
  - What it will not do: Generate a person, a face, a reenactment or 'what the crash looked like'; edit or generate from the client's photos; show a child's photo (refused for now); write a demand amount; render before every scene is approved.
- Samples:
  - 1. Whitlock v. Harmon (fictional): rear-end collision, 5 providers, bills and a signed statement (`whitlock`): https://decosa.ai/legal/settlement-video?sample=1&autorun=0
- Regulatory note: Written 29 Sep 2026. Settlement communications are generally inadmissible to prove or disprove a disputed claim's validity or amount under Federal Rule of Evidence 408 (https://www.law.cornell.edu/rules/fre/rule_408, read 29 Sep 2026); states have their own versions, and whether and how to label a video is the lawyer's call. Medical records are protected health information in a covered entity's hands (45 CFR Parts 160 and 164, https://www.hhs.gov/hipaa/for-professionals/privacy/index.html); a firm holds them under its own duties of confidentiality, so real matters run self-hosted or on the confidential API, never the hosted demo. The client's photos and statement are shown only with the client's recorded consent (consent ledger), and any generated narration voice is disclosed in the video and its content credential. It is a drafting aid, not legal advice, and it never writes a demand amount.

### Find the themes in your interviews (`interview-themes`)

For PhD students, academic researchers, UX researchers and program evaluators with 10 to 60 recorded interviews. Recordings come in as transcripts with every speaker kept apart; a voice check moves segments credited to the wrong person and marks them, and interviewer turns are never coded or quoted. An open model proposes a codebook of 8 to 12 codes that you rename, merge, split and approve; then every passage of participant talk is coded against it. Themes come with participant counts worked out in code and three quotes each, every quote checked word for word against the transcript and the person it is credited to, and playable from its timestamp. Export a REFI-QDA project, CSVs, a methods paragraph and a signed record of every model call; after you review a few interviews, a small model of your own codes the rest your way.

- What it is: A thematic analysis tool for interview studies: speaker-separated transcripts with the interviewer left out, a codebook of 8 to 12 codes you edit and approve, every passage coded against it, and themes with participant counts and quotes checked word for word and playable at their timestamp.
- Who it's for: PhD students, academic researchers, UX researchers and program evaluators with 10 to 60 recorded interviews.
- Page: https://decosa.ai/tools/research/interview-themes
- Last reviewed: 2026-09-30
- Key result: On 1,000 passages of published human-coded interviews (28 codes) the coder agreed with the published coding at Cohen's κ 0.52 (0.49 on the test split); a blind second coder reached 0.62 on a 120-passage sample. One dataset, one domain.
- Key numbers (details and caveats: https://decosa.ai/metrics/interview-themes):
  - 0.05% (22 of 41,624): Words credited to the wrong speaker (12 public-domain interviews) (test split, n = 41624)
  - 0.49: Agreement with published human coding (Cohen's κ, 28 codes, test split) (test split)
  - 15 of 15: Quotes passing the word-for-word and speaker check (test split, n = 15)
  - 69.4 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: preview. Models: MOSS-Transcribe-Diarize 0.9B (speech and speakers) · ECAPA-TDNN voice check (CPU) · Qwen3.8-27B (codebook, coding, themes) · bge-small-en-v1.5 (your own model, CPU). Where: Self-host, or a confidential enclave on request, for interviews under an ethics or IRB protocol; the hosted demo takes the public-domain sample interviews only.
- Verification: Every quote re-checked word for word against the transcript and against its speaker; participant counts computed in code; interviewer turns excluded; signed record of every model call (receipt coverage: full). Hosted QA 2026-09-30: pass. Self-host QA 2026-09-29: partial.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B and the diarizer; the voice check, the quote check and your own model run on CPU
- Output: An approved codebook with its edit history, coded passages, themes with participant counts and checked, playable quotes, a REFI-QDA project, CSVs, a methods paragraph and a signed record.
- Page: https://decosa.ai/tools/research/interview-themes · JSON: https://decosa.ai/use-cases/interview-themes.json
- Prompts: https://decosa.ai/prompts/interview-themes-hosted.md · https://decosa.ai/prompts/interview-themes-selfhost.md · https://decosa.ai/prompts/interview-themes-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same codebook, coding and themes on a smaller mixture-of-experts model. Faster and cheaper; agreement on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B proposes the codebook, codes every passage and groups themes; the diarizer and voice check make the transcripts; counts, the quote check and your own model are code on CPU.
  - Wanted · a much larger coder on your own hardware (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash (estimate)): DeepSeek-V4-Flash codes the study; it may come closer to a careful researcher on fine codebooks. Interviews stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What you get: Speaker-separated transcripts with interviewer turns excluded; an approved codebook with its edit history; every passage coded; themes with participant counts and three checked quotes each, playable at their timestamp; a REFI-QDA project (.qdpx) and codebook (.qdc), which most QDA programs import; CSVs; a methods paragraph; a signed record.
  - What it never does: It never decides your findings (codes and themes are proposals you approve), never quotes a line it credits to the interviewer, never edits a quote, and never trains a model on what you send.
  - Data retention: Nothing kept on the server: audio and text live in memory for each request, and logs carry counts and timings only. The study lives in your browser until you export it; the recording is dropped after transcription and only its hash is kept in the record.
  - Where it runs: Self-host on your own server or laptop for interviews under an ethics or IRB protocol, or a confidential enclave on request. The hosted demo takes the public-domain sample interviews only.
  - Your own model: After you review a few interviews, a small classifier (MIT-licensed sentence embeddings plus one logistic-regression head per code, trained on CPU in seconds) codes the rest your way. It is returned to you and not kept.
  - Cost: Cents per interview hour at list price (model time plus GPU-minutes of transcription), measured on a sample of interviews and scaled.
- Regulatory note: Not a replacement for ethics or IRB approval, and it does not decide findings: codes and themes are proposals the researcher reviews. Interviews under a protocol belong on your own hardware or in a confidential enclave (on request); the hosted demo takes public-domain sample interviews only.

### Find the songs in your mix (`mix-cue-sheet`)

Drop in a DJ mix or any long recording, with the tracklist if you have it and the track files you played if you want exact starts. Every song start appears on the timeline; click one and the player jumps there. You get a timestamped tracklist to paste under your YouTube, Mixcloud or SoundCloud upload, a chaptered .m4a your phone skips song by song, and a signed cue sheet that names the audio it describes. It runs on CPU with no model call, and it only names songs you give it a title or a file for.

- What it is: Make a DJ mix tracklist with times from the audio: upload the mix with the titles you played, and optionally the track files, and get every song start, a chaptered .m4a and a signed cue sheet.
- Who it's for: DJs who publish mixes, radio and podcast editors who cut long shows, and developers who need song starts in long audio.
- Page: https://decosa.ai/tools/media/mix-cue-sheet
- Last reviewed: 2026-09-29
- Key result: 105 of 105 song starts within 10 s (92 within 5 s) when you add the track files you played; 84 of 105 within 10 s from titles alone. Held-out synthetic mixes of 60-second generated songs; real DJ mixes with long blends are harder (61 of 190 within 10 s from titles alone on eight public mixes).
- Key numbers (details and caveats: https://decosa.ai/metrics/mix-cue-sheet):
  - 105/105 (92/105 within 5 s, 24/105 within 1 s): Song starts within 10 s, with your track files (test split, n = 105)
  - 84/105 (51/105 within 5 s, 15/105 within 1 s): Song starts within 10 s, titles only (the hosted default) (test split, n = 105)
  - 3/105 / 10/105 / 19/105: Song starts within 1 / 5 / 10 s, audio only (test split, n = 105)
  - 7.0 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: live. Models: No model call: the decosa-cue engine (numpy, Apache-2.0) and a landmark matcher for your own track files, both on CPU. Where: Hosted or self-host.
- Verification: Signed cue sheet naming the audio's SHA-256, checkable at /record/verify (receipt coverage: none). Hosted QA 2026-09-29: pass.
- Hardware: Any CPU; no GPU
- Output: A timestamped tracklist (text), a chaptered .m4a with one chapter per song, and a signed cue sheet (JSON, verifiable at /record/verify).
- Page: https://decosa.ai/tools/media/mix-cue-sheet · JSON: https://decosa.ai/use-cases/mix-cue-sheet.json
- Prompts: https://decosa.ai/prompts/mix-cue-sheet-hosted.md · https://decosa.ai/prompts/mix-cue-sheet-selfhost.md · https://decosa.ai/prompts/mix-cue-sheet-assemble.md
- Tiers:
  - Lite · titles and times only, any CPU (Any CPU): Your tracklist and the audio's change points: times you give are used as written, other titles are placed at the strongest change and marked estimated. No track files.
  - Standard · adds your own track files (hosted) (Any CPU): Everything in Lite, plus up to 30 of your own track files found in the mix from the audio, so their songs start where the music does, not where a guess puts them.
  - Best · adds the learned transition detector (self-host only) (Any CPU): Everything in Standard, with the transition detector in place of the novelty curve for titles without a time or a file. Off on the hosted tool: its training audio (DJ Mix Dataset, YouTube-sourced) has no dataset licence, so turn it on only with weights you are comfortable using (DECOSA_CUE_DETECTOR=onnx). A licence-clean retrain is planned.
- Buyer facts:
  - Data retention: Your mix and track files are deleted when the run ends. The chaptered .m4a is kept for one hour for download, or until you delete it. The song list and the signed cue sheet (no audio) stay in memory for 24 hours.
  - What leaves the box: Nothing. There is no model call and no lookup service on the hosted tool; the audio is processed on Decosa's own server. Self-hosted: nothing, unless you set your own AudD token for song lookup.
  - Limits: A mix up to 150 MB and 2 hours; up to 30 track files of up to 40 MB and 15 minutes each; a tracklist of up to 400 lines.
  - What it can't do: Name a song you gave it no title or file for. Without times or files, starts on long beat-matched blends are rough: they are marked estimated.
  - Output: Every song's start and how it was found; a tracklist (0:00 Title per line); a chaptered .m4a, one chapter per song; and a signed cue sheet naming the audio's SHA-256, checkable at /record/verify.
- Samples:
  - 1. Demo 1 (`demo-1`): https://decosa.ai/tools/media/mix-cue-sheet?sample=1&autorun=0
  - 2. Demo 1 files (`demo-1-files`): https://decosa.ai/tools/media/mix-cue-sheet?sample=2&autorun=0
- Regulatory note: Checked 29 Sep 2026. It marks where songs start in audio you upload; it does not identify, license or clear music. Whether you may publish a mix of other people's recordings depends on your rights and the licences of the platform you post to, which this tool does not check. The optional song lookup on a self-hosted server sends a compressed copy of the mix to AudD (audd.io), a paid third-party service under its own terms, only when you set your own token; it is off by default and never used by the hosted tool. The optional transition detector is off on the hosted tool because its training audio (the DJ Mix Dataset, sourced from YouTube) carries no dataset licence; a self-hosted server can turn it on with its own weights file. Engine licence: Apache-2.0 (decosa-cue, ported from an MIT-licensed original); the track-file matcher is decosa-api code (AGPL-3.0-or-later).

### Reply to a review without breaking patient privacy (`review-reply`)

Paste a Google or Yelp review. The open model drafts a short, warm reply, and code checks it before you see it: no offers or refunds in public, no contact details or numbers you didn't give, no arguing, and for a dentist, clinic or therapist never a word that confirms the reviewer is a patient: no visit, treatment, wait, bill, family member or name. A second model check reads health and care replies on their own. A blocked draft is redrafted once, then a fixed safe reply. Nothing is posted and the review isn't stored.

- What it is: A hipaa review response has to thank or apologise without saying the reviewer was ever a patient. This drafts the reply and checks it in code before you see it.
- Who it's for: Dentists, clinics, therapists and the office managers and agencies who answer their reviews.
- Page: https://decosa.ai/tools/operations/review-reply
- Last reviewed: 2026-09-29
- Key result: On a held-out set judged blind, no healthcare reply confirmed a patient; the price is a generic fixed reply for many of them.
- Key numbers (details and caveats: https://decosa.ai/metrics/review-reply):
  - 0 / 40: Healthcare replies that confirm a patient (blind judge) (test split, n = 40)
  - 93 / 100: Replies a business could post as written (blind judge) (test split, n = 100)
  - 24 / 40: Healthcare replies that fell back to the fixed safe reply (test split, n = 40)
  - 1.8 s: Median end-to-end run, hosted (QA sweep 2026-09-29)

- Status: live. Models: Qwen3.8-27B (drafts, and reads health and care replies for patient privacy); the checks are code from the open-source site kit. Where: Hosted demo on made-up reviews; self-host for your practice's real reviews.
- Verification: Checks in code (Apache-2.0 site kit); receipt per model call; signed record with the review's fingerprint only (receipt coverage: full). Hosted QA 2026-09-29: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) for Qwen3.8-27B; the checks run on CPU
- Output: A reply to read and post yourself, the words that were blocked, and a signed record that keeps only the review's fingerprint.
- Page: https://decosa.ai/tools/operations/review-reply · JSON: https://decosa.ai/use-cases/review-reply.json
- Prompts: https://decosa.ai/prompts/review-reply-hosted.md · https://decosa.ai/prompts/review-reply-selfhost.md · https://decosa.ai/prompts/review-reply-assemble.md
- Tiers:
  - Lite · one 48 GB card (1x L40S or RTX 6000 Ada 48 GB (not measured)): The same guard and flow on a smaller mixture-of-experts model. Faster and cheaper; accuracy on this task unknown.
  - Standard · the hosted demo, one 96 GB card (1x RTX PRO 6000 Blackwell 96 GB): Qwen3.8-27B drafts; the guard is code; health and care replies also get the model's privacy check. One to four model calls per reply, each receipted.
  - Best · DeepSeek-V4-Flash on two more cards (2x RTX PRO 6000 96 GB): A larger model for long, multi-claimant files.
  - Wanted · two large judges from different families (Your own hardware: 2x 96 GB cards for DeepSeek-V4-Flash plus 2x 96 GB for GLM-5.3-Flash, or one Mac Studio with 512 GB holding both 4-bit builds (156 + 165 GB, sizes from our Mac; not run together yet). Estimate.): DeepSeek-V4-Flash and GLM-5.3-Flash each read the file, and a finding stands when they agree; disagreements go to the reviewer. Claim files stay on your own hardware, never on community providers. Not served yet.
- Buyer facts:
  - What it checks: Offers and refunds, contact details or links you didn't give, numbers not in the review or your facts, placeholders, arguing with the reviewer, and that negative reviews are taken offline. For health and care businesses: anything that confirms a patient, a visit, a treatment, a condition, a wait, a bill or a family member's care, and any provider's or the reviewer's name.
  - What it never does: Post a reply, ask for a rating change, or store the review.
  - Data retention: None: the review and the reply live in memory for the request. The signed record holds the review's SHA-256, never its text.
  - What leaves the box (hosted demo): The review goes to the open model on Decosa's hosted service through the receipted gateway. Nothing goes to a third party.
  - Model calls per reply: 1.55 on average on the held-out set (a draft, sometimes a redraft, and for health and care businesses a privacy check of each draft that passes the code)
  - Cost per reply: A fraction of a cent per reply on average at list price, held-out set.
  - Also used in: ElmoSEO's review replies run the same guard (the TypeScript kit).
- Samples:
  - 1. Dentist named (`dentist-named`): https://decosa.ai/tools/operations/review-reply?sample=1&autorun=0
  - 2. Therapist angry (`therapist-angry`): https://decosa.ai/tools/operations/review-reply?sample=2&autorun=0
  - 3. Chiro mixed (`chiro-mixed`): https://decosa.ai/tools/operations/review-reply?sample=3&autorun=0
  - 4. Bakery happy (`bakery-happy`): https://decosa.ai/tools/operations/review-reply?sample=4&autorun=0
  - 5. Plumber refund (`plumber-refund`): https://decosa.ai/tools/operations/review-reply?sample=5&autorun=0
  - 6. Injection (`injection`): https://decosa.ai/tools/operations/review-reply?sample=6&autorun=0
- Regulatory note: HIPAA Privacy Rule, 45 CFR 164.502(a): a covered entity may not use or disclose protected health information except as permitted; confirming in a public reply that someone is a patient is a disclosure. HHS OCR resolutions over online review replies (read 29 Sep 2026 on hhs.gov): Elite Dental Associates, 2019, $10,000; New Vision Dental, $23,000; Manasa Health Center, $30,000, replies to negative Google reviews (hhs.gov/hipaa/for-professionals/compliance-enforcement/agreements: elite, new-vision, manasa). The FTC's rule on consumer reviews and testimonials (16 CFR Part 465, in force 21 Oct 2024) bars buying or suppressing reviews; this tool only drafts replies and never asks for a rating change. It checks common failure patterns; it is not legal advice.

### See what studies found for a supplement (`what-studies-found`)

Type a supplement and an outcome, such as probiotics and antibiotic-associated diarrhea. It searches PubMed for trials, meta-analyses and systematic reviews and gives you a table of what they found: one row per study with its design, how many people, who they were, whether the result was better, no different or worse, the finding in numbers where the abstract gives them, and a short quote. Each row links to its PubMed record, each quote is found word for word in the abstract by code, and each finding is judged against its own abstract. A published rubric in code then grades the body of evidence, a separate reading of every abstract flags possible harm, and the studies it left out are listed with why. It reports what studies found: no doses, no advice on what to take.

- What it is: A supplement evidence table in seconds: type a supplement and an outcome, and get one row per trial or review from PubMed with its design, size, population, what it found and a short quote, each checked against its own abstract, then a grade from a rubric anyone can read.
- Who it's for: Evidence-based supplement sites, supplement brands' regulatory teams, health journalists and researchers who need a first table of the trials with PMIDs.
- Page: https://decosa.ai/tools/life-sciences/what-studies-found
- Last reviewed: 2026-09-30
- Key result: On 289 held-out pairs the table's verdict matched the systematic review's in 231 (80%). It surfaced 36 of 40 harms and called a benefit on 1 of 61 reviews that found no clear difference. Read the rows, not only the verdict.
- Key numbers (details and caveats: https://decosa.ai/metrics/what-studies-found):
  - 231 / 289 (80%): Verdict matches the systematic review's (benefit, harm or no claim) (held out, n = 289)
  - 1 / 61 (2%): Calls a benefit where the review found no clear difference (held out, n = 61)
  - 36 / 40 (90%): Harms surfaced: the verdict is Worsens, or the table carries its harm flag (held out, n = 40)
  - 26.3 s: Median end-to-end run, hosted (QA sweep 2026-09-30)

- Status: preview. Models: Qwen3.8-27B (reads each abstract, reads it again for harm, then checks each finding against it) · Qwen3-Reranker-4B (picks the studies, evidence retrieval block). Where: Hosted or self-host.
- Verification: Receipt per model call; quotes checked word for word in code; the rubric grades in code; signed hash-chained record of the search, each study and the grade (receipt coverage: full). Hosted QA 2026-09-30: pass.
- Hardware: 1x RTX PRO 6000 (96 GB) or 1x RTX 5090 (32 GB) for Qwen3.8-27B; about 8 GB for the reranker; the search, checks and grade run on CPU
- Output: An evidence table (Markdown or CSV) with a PubMed link per row, the grade with the rubric's reasons, the studies left out with why, and a signed record verifiable at /record/verify.
- Page: https://decosa.ai/tools/life-sciences/what-studies-found · JSON: https://decosa.ai/use-cases/what-studies-found.json
- Prompts: https://decosa.ai/prompts/what-studies-found-hosted.md · https://decosa.ai/prompts/what-studies-found-selfhost.md · https://decosa.ai/prompts/what-studies-found-assemble.md
- Tiers:
  - Lite · one 32 GB card, no reranker (1x RTX 5090 32 GB (estimate)): The same model and checks on one smaller card, with PubMed's own relevance order instead of the reranker. Fewer of the picked studies may be on point.
  - Standard · one 96 GB card (measured; hosted demo) (1x RTX PRO 6000 96 GB, plus about 8 GB for the reranker): Qwen3.8-27B reads every abstract, reads it again for harm and checks each finding; the reranker picks the studies; the rubric grades in code.
- Buyer facts:
  - Data retention: The hosted service keeps the PMIDs each search returned for 7 days, keyed by a SHA-256 of the search term, so repeat searches are fast. Titles and abstracts live in memory for the request; the signed record holds each abstract's SHA-256, not its text.
  - What leaves the box: The search words and PMIDs go to NCBI's public PubMed service. Abstracts and search words go to the model and the reranker, which run on Decosa's hosted service (hosted) or your machine (self-host).
  - What it will not do: No doses, no advice on what to take, no treat, cure or prevent wording: a dosing or advice question is refused, and the guard rewrites such wording in findings.
  - Input: A supplement (up to 80 characters) and an outcome (up to 120), or a sample. Optional: how many studies to read, primary trials only, or only studies published before a year.
  - Output: A table with one row per study (PMID and link, design, n, population, direction, finding, a quote of at most 15 words, the grounding verdict, a harm flag), the grade with its reasons, the table's harm flag with the studies behind it, the studies left out and why, and a signed record verifiable at /record/verify.
- Samples:
  - 1. Probiotics antibiotic diarrhea (`probiotics-antibiotic-diarrhea`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=1&autorun=0
  - 2. Melatonin sleep onset (`melatonin-sleep-onset`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=2&autorun=0
  - 3. Beta carotene lung cancer (`beta-carotene-lung-cancer`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=3&autorun=0
  - 4. Caffeine blood pressure (`caffeine-blood-pressure`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=4&autorun=0
  - 5. Vitamin d asthma (`vitamin-d-asthma`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=5&autorun=0
  - 6. Omega3 depression (`omega3-depression`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=6&autorun=0
  - 7. Cranberry uti (`cranberry-uti`): https://decosa.ai/tools/life-sciences/what-studies-found?sample=7&autorun=0
- Regulatory note: Written 29 Sep 2026. This tool reports what published studies found. It is not medical advice, not a recommendation and not a dose, and it says nothing about any product. For supplement sellers: under the Dietary Supplement Health and Education Act of 1994 (DSHEA), a supplement label may carry a structure/function statement with the FDA disclaimer in 21 CFR 101.93(c) ("This statement has not been evaluated by the Food and Drug Administration. This product is not intended to diagnose, treat, cure, or prevent any disease."), but not a disease claim, which 21 CFR 101.93(g) describes as a claim to diagnose, mitigate, treat, cure or prevent disease (https://www.law.cornell.edu/cfr/text/21/101.93, read 29 Sep 2026; official text at https://www.ecfr.gov/current/title-21/chapter-I/subchapter-B/part-101/subpart-F/section-101.93). The FTC's Health Products Compliance Guidance (December 2022, https://www.ftc.gov/business-guidance/resources/health-products-compliance-guidance, read 29 Sep 2026) expects competent and reliable scientific evidence for health claims, which it says will generally need to be randomized, controlled human clinical testing, and the evidence must fit the claim. A table of what trials found is input to that judgement, not substantiation by itself and not wording you can put on a label: the product, amount and population in the trials must match, and counsel decides. The tool's wording guard (studies-guard-1) rewrites or blocks treat, cure, heal, prevent and reverse wording, personal recommendations, and doses or schedules, and a dosing question is refused. The search words go to NCBI's public E-utilities, whose usage policy applies [not re-read for this note]. PubMed abstracts can be under publisher copyright [unverified: it varies by journal], so the tool keeps PMIDs, its own structured findings, quotes of at most 15 words and a SHA-256 of each abstract, not abstract text. Model licences: Apache-2.0 (Qwen3.8-27B, Qwen3-Reranker-4B). Not legal or regulatory advice.

## API routes (from the contract)

Base: https://api.decosa.ai. Hosted keys: `Authorization: Bearer dk_…`. The markdown contract is the source of truth: https://decosa.ai/api/contract.md.

- `POST /demo/session` (Auth: demo sessions)
- `GET /healthz` (Auth: demo sessions)
- `WS /ws/live` (Live audio verticals (clinical, sales, field, translate))
- `POST /demo/replay` (Live audio verticals (clinical, sales, field, translate))
- `POST /v1/chat/completions` (Code assistant)
- `GET /studio/gallery` (Studio)
- `POST /studio/jobs` (Studio)
- `GET /studio/jobs/{id}` (Studio)
- `GET /demo/recordings` (Watch (recorded sessions))
- `GET /receipts/{id}` (Receipts)
- `GET /demo/scripts` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /v1/models` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /v1/keys` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /v1/keys/{key_id}` (Changes (decosa-api v0.1.0, 2026-09-23))
- `DELETE /v1/keys/{key_id}` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /studio/policy` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /studio/report` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /studio/report` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /attest/signing-key` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /attest/models` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /record/verify` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /provenance/status` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /provenance/signing-key` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /provenance/check` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /provenance/receipts/{rr_id}` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /provenance/samples` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /provenance/consent-check` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /provenance/consents` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /provenance/consents/{cr_id}` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /provenance/consents/{cr_id}/revoke` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /audit/signing-key` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /audit/targets` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /audit/references` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /audit/references/{id}` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /audit/runs` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /audit/reports/{id}` (Changes (decosa-api v0.1.0, 2026-09-23))
- `GET /audit/reports` (Changes (decosa-api v0.1.0, 2026-09-23))
- `POST /audit/verify` (Changes (decosa-api v0.1.0, 2026-09-23))