Family interview film

Eval results

Not held outRun 29 Sep 2026Eval write-up (decosa-api, access required)

• Shown quotes inside her own turn, at the right time46 / 46dev (tuned on)n = 465 languages; a quote counts when it overlaps her true turn by at least 80% of its span.
• Shown quotes that match her words (fuzzy 0.9)44 / 46dev (tuned on)n = 46Both misses write the year as digits where the script spells it out; one also has "né" for "née" (a real slip).
• Quotes dropped by the re-hearing check2 / 48dev (tuned on)n = 48Both had speech-recognition slips; dropped quotes are never shown.
• Speaker labels right79 / 80dev (tuned on)n = 80The miss: a 0.66 s interviewer segment labelled hers; no quote came from it.
• Meaning check: lines flagged / clear mistranslations27 / 46 flagged; 2 cleardev (tuned on)n = 46The builder read every flag: 2 clear mistranslations, 25 nitpicks or misreadings. Misses not measured.
• Consent refusals and hesitations refused2 / 2syntheticn = 2
• Photo backs read3 / 3syntheticn = 3Synthetic handwriting on the sample's photo backs.
• First trailer after upload (p50)40.8 sdev (tuned on)n = 537-71 s; interviews of 56-207 s, pre-release server.
• Cost per film$0.0085-0.02dev (tuned on)n = 5List prices for the text model and GPU time.
• Blind granddaughter: keeps it / shares it as is / would payyes / no / $39syntheticn = 1An Opus sub-agent on the Italian sample: about 40 min to fix in the tool against 6+ hours by hand (its estimate). Two of its problems fixed since: held error lines, no borrowed photos.

Nightly smoke check

Loading the nightly status…

Result

pass

Run

29 Sep 2026

Latency, this run

n/a

p50 over passed runs

41 s

Receipts

11

Model calls

n/a

Tokens

n/a

Cost per run

$0.009

Self-host verification

Not yet verified on a fresh self-host setup.

Models and licences

Standard tier. Licence posture: permissive (Apache, MIT or BSD).

• Speech recognition in her language, one call per speech segment (so every word keeps its time)Qwen3-ASR-1.7B (language pack speech service)Apache-2.0
• Who is speaking: voice activity by energy, then ECAPA-TDNN voice embeddings per segment in two clusters; the cluster nearest her consent recording is hersECAPA-TDNN (speechbrain/spkrec-ecapa-voxceleb, ONNX export)Apache-2.0
• Chapters of her life and her best lines (copied exactly from the transcript, then re-found in it by code), film titles, and the translation fallbackQwen3.8-27B (NVFP4)Apache-2.0
• English subtitles under her own words, sentence by sentence, then a meaning check (back-translation compared with the source) on every lineHy-MT2-7B (the language-pack block)Apache-2.0
• Face detection only (boxes): a picture is read as the back of a photo only when no face is found; photo pans drift toward facesUltra-Light-Fast-Generic-Face-Detector-1MB (version-RFB-320)MIT
• Reads the handwriting on the backs of photos (place, year, names) to date and place each pictureDecosa document reader (Docling layout + PaddleOCR-VL-1.6)Apache-2.0
• A quiet score under the film: a pre-rendered cue from the cleared music library (no model runs per film)ACE-Step 1.5 cue (music-gen-cleared library)MIT
• The film (CPU): her voice over her photos with slow pans, chapter cards, maps (Natural Earth) and dates, subtitles, the trailer and the book with a QR code per quote; C2PA credential per filedecosa-api family_film module + FFmpeg + c2pa-pythonAGPL-3.0-or-later

All quality evidence

Every sourced number on the tool’s Stack tab, by tier. Some are proxies from another task; their labels say so.