Trust
Status
What we host, and whether it is up. Every model and service behind the hosted API: live status, latency, load and uptime, the exact model and its licence, and where it runs. Each one is open weights, so you can also run it yourself.
Machine-readable: GET /status on the API. Per-tool checks and costs: metrics. How we measure: below.
- What this page proves
- The latest end-to-end run of each tool's sample on the hosted service, with when it ran and how long it took.
- What it doesn't
- Uptime between checks, or how your own runs went.
Loading live status…
Models
Qwen3.8-27B (text, image, video in)
No recent check
The main language model behind every tool: chat, extraction and judging (hosted video input is off at launch). Every call gets a signed receipt that names who served it.
- Model
- Qwen3.8-27B (open weights, FP8 at the provider)
- Licence
- Apache-2.0
- Runs on
- Outside provider: NEAR AI through OpenRouter, with Reka AI as the only fallback
Capability blocks
Document reader
No recent check
Reads PDFs and scans into text, tables and charts with page coordinates.
- Model
- Docling (Heron layout) + PaddleOCR-VL-1.6
- Licence
- MIT (Docling) · Apache-2.0 (weights)
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Evidence retrieval
No recent check
Finds the passages that support a claim: embeddings plus a reranker, with the index hash on every receipt.
- Model
- Qwen3-Embedding-0.6B + Qwen3-Reranker-4B
- Licence
- Apache-2.0
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Machine translation
No recent check
Translation for the language pack. The other EU languages go through Qwen3.8-27B.
- Model
- Hy-MT2-7B
- Licence
- Apache-2.0
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Translation quality estimate
No recent check
Scores each translated line so weak ones get a human look.
- Model
- decosa-qe-xlmr-large (ours, on XLM-RoBERTa-large)
- Licence
- MIT base · our weights unreleased
- Runs on
- Decosa server · CPU
Speech: transcription and voice
No recent check
Batch speech-to-text and text-to-speech for the language pack.
- Model
- Qwen3-ASR-1.7B + VoxCPM2
- Licence
- Apache-2.0
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Live transcription
No recent check
Streaming speech-to-text for the live consoles: scribe, calls and depositions.
- Model
- Voxtral Mini 4B Realtime
- Licence
- Apache-2.0
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Speaker diarization
No recent check
Who spoke when, for recorded calls and hearings.
- Model
- MOSS-Transcribe-Diarize 0.9B
- Licence
- Apache-2.0
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Studio renders
No recent check
Image and video renders for the creative studio, one job at a time.
- Model
- ComfyUI with Wan 2.x workflows
- Licence
- GPL-3.0 (ComfyUI) · Apache-2.0 (Wan)
- Runs on
- Decosa's hosted service (model provider under Decosa's account)
Music similarity
No recent check
Compares a new track against a rights-cleared catalogue before release.
- Model
- LAION CLAP larger_clap_music
- Licence
- Apache-2.0
- Runs on
- Decosa server · CPU
Beat and lyric alignment
No recent check
Beats, sections and lyric timing for the music-video studio.
- Model
- wav2vec2-large-960h-lv60-self + librosa
- Licence
- Apache-2.0 · ISC
- Runs on
- Decosa server · CPU
Gateway and privacy tiers
Metering and signed receipts
No recent check
Meters every hosted model call and signs its receipt. If it is down, hosted calls fail rather than run without a receipt.
- Model
- Decosa API
- Licence
- Decosa code
- Runs on
- Decosa server · CPU
Sealed tier (paused at launch)
No recent check
End-to-end encrypted chat. Paused at launch: hosted calls run on outside providers, so there is no Decosa-operated server to decrypt on. It comes back as a confidential (TEE) tier.
- Model
- Qwen3.8-27B (end-to-end encrypted chat)
- Licence
- Apache-2.0 (model)
- Runs on
- Paused at launch
- Pilot, time-limited
Confidential tier (pilot)
No recent check
The model and the decryption run inside a hardware-attested enclave that you can verify offline. A time-limited pilot, not a standing service.
- Model
- Qwen3.8-27B FP8 in an attested enclave
- Licence
- Apache-2.0 (model)
- Runs on
- Phala Cloud TEE · 1× H200 (Intel TDX + NVIDIA confidential computing)
How we measure
- What is checked
- Every 60 seconds a probe on our server sends one health request to each service, and reads the model server's own counters where it has them. No model is called and nothing is billed. A service is down if it does not answer within 4 seconds, answers with an error, or says it is not ready. An endpoint that needs the gateway is also down when the gateway is.
- Degraded
- Still serving, but more than 8 requests are waiting on a model server, more than 3 renders are queued, a health check took over 2.5 seconds, or the gateway is holding receipts. Degraded counts as up in the uptime figures.
- Uptime
- The share of probes in which the service answered and was ready, over 24 hours, 7 days and 30 days. The bar shows one day per cell in UTC: green at 99.9% or better, amber from 95%, red below, grey where we have no data. History starts when the probe was switched on; we don't backfill.
- Latency
- For model servers: the median time to first token and to the full answer over real requests in the last 24 hours, from the server's own histograms. For the other services it is only the health check's round trip, which shows the service answers, not how long a real job takes. Per tool timings are on the metrics page.
- What is not shown
- No addresses, ports, keys or error text. A failure is shown as one of a few fixed reasons.