review-ugc-render

The QC gate every UGC video recipe MUST clear before it publishes. Not an eyeball /watch — a deterministic transcript-vs-script diff that exits non-zero on a defect so the recipe can hard-stop pinning the final render (video_project_upsert patch.final_render_id).

In short: it compares what the render actually SAYS with the script the user approved. Correct speech that is only written differently passes (numbers, units, URLs, contractions, fused brand names, confirmed pronunciations). A wrong number, a flipped negation, a mis-voiced word or brand, a dropped phrase or silence fails.

Why this exists

Seedance generates the audio natively. It sometimes mis-voices a word — the approved line human-vetted comes back spoken as human witted; documented siblings: Hume→Hune, Alitu→al-too. The defect lives in the render's audio, so an eyeball /watch ("dialogue matches the script") slips it through, and a downstream caption pass then bakes the wrong word in verbatim. Nothing was comparing the actual spoken audio against the script the user approved.

This gate does exactly that, deterministically, and refuses to publish on a miss.

When to run

• MANDATORY in every remix-ugc-*-from-sample and create-ugc-*-video-from-refs recipe, in the QC phase, after the master render exists and before pinning it as the final render (video_project_upsert patch.final_render_id).
• Re-run after every fix / re-roll until it PASSES.

Contract

Before rendering, persist the exact approved spoken lines (the verbatim utterance, no beat notes) to working/approved-script.txt. Then, after render:

python3 <pack>/review-ugc-render/scripts/review_render.py \
  --video working/final.mp4 \
  --script-file working/approved-script.txt \
  --json working/review-verdict.json

When the voice-over used confirmed pronunciations, add them (recommended): --pronunciations working/brand-rules.json — the same file create-vo-elevenlabs read (its --rules brand-rules.json, or the read_pronunciations.py output; both carry a pronunciations: [{term, say_as}] list). Leave the flag out when there is no such file: a missing file is an ERROR (exit 3).

• exit 0 → PASS — proceed: pin the final render (video_project_upsert patch.final_render_id).
• exit 2 → FAIL — do NOT pin it. Read the report, fix, re-run.
• exit 3 → ERROR — the check could not run (see below); fix the environment or the input, do not publish blind.

Transcription backend (in priority order): the GooseWorks whisper-proxy (CLI credentials or the sandbox token) → OPENAI_API_KEY (honors OPENAI_BASE_URL) → local whisper CLI. ffmpeg must be on PATH.

What counts as the same speech

Both the script and the transcript are put in one canonical spoken form before the diff. The rules are bounded — each is an exact rewrite, never a fuzzy match.

Guards that keep the rules honest:

• A unit word not after a quantity is left alone — a brand "MG" never becomes "milligrams".
• Letters spelled one by one ("A G") are not fused into a word. That needs a confirmed pronunciation.
• Fusion is exact concatenation. "Braxly Bands" is not "Braxleybands".
• A join never swallows a negation: "no table" is not "notable" ("no thing" is "nothing").
• A number word joins a word only when every part has 2+ letters: "every one" is "everyone", but "G one" is not "gone".

What still FAILS

The verdict passes only when similarity ≥ --min-ratio and there is no HIGH issue. Numbers, units, negations and declared brand names fail on their own (HIGH). Any other dropped or extra word is medium/low: it lowers the similarity, and fails the gate only when the similarity drops below --min-ratio (so one dropped ordinary word in a long line can pass). --expect-music is advisory only; it does not by itself fail the gate.

Brand names and confirmed pronunciations

Brand words are never removed from the diff. (Before 2026-10-06, --brand-term fuzzily stripped brand-like words, which let "Hune" pass for "Hume". That is gone.)

• --pronunciations PATH (recommended when the voice-over used one) — the confirmed pronunciations file: create-vo-elevenlabs's brand-rules.json or the output of its scripts/read_pronunciations.py ({"brand_id", "basis", "pronunciations": [{"term", "say_as", "fact_id"}]}). Every entry becomes an alias and its term a brand term. An entry that cannot be used (see alias rules) is skipped with a WARNING, never fatal. A missing or unreadable file is an ERROR (exit 3), so pass the flag only when the file exists.
• --alias "TERM=SPOKEN" (repeatable) — one confirmed spoken form, e.g. --alias "AG1=A G one". The term also becomes a brand term. A bad --alias is an ERROR (exit 3), checked before any transcription is spent.
• --brand-term TERM (repeatable) — marks a brand name. Exactly:
  1. Where the term's words appear in the script, a substitution or drop there is HIGH (a brand mis-voicing). Its fused/split forms count as equal.
  2. A differing span is accepted as the same brand (reported [low], counted as a match) only when all of these hold: it has no negation, number or unit change; every word on both sides is an alphabetic word of some --brand-term; and the two sides look alike (character similarity ≥ 0.6). This keeps the older calling pattern working: a script written in the spoken form (Try ak-mee today) with --brand-term Acme --brand-term ak --brand-term mee passes when Whisper writes Try Acme today ("ak mee" vs "acme" = 0.67). Everything else still fails HIGH: a heard word that is not a declared term (Hume heard Hune), two different declared names (Hims vs Hers, Hume Body Pod vs Hume Band), a truncation (Acme heard ak), or a number inside a term (Pod 4 vs Pod 5, 7-Eleven: 7 days vs 11 days). Prefer --pronunciations over passing say_as words as --brand-term.

Rules for aliases:

• Only pass spoken forms the user confirmed (saved brand pronunciations, or a form they confirmed in chat). Never invent one from the transcript to make the gate pass.
• An alias may respell a name, digits and number-like syllables included ("AG1" = "A G one", "Tenzing" = "ten-zing", "Notion" = "NO-shun"). It is refused when the written term has a number, unit or negation that the spoken form changes ("AG1" = "A G two"), when the spoken form adds not, never, without, none, nothing or nobody ("Hume" = "never Hume"; no and nor are fine as syllables, "Nomad" = "no-mad"), or when the spoken form is only numbers, units or negations ("Decagon" = "five").
• If you listened and the audio is right but Whisper spelled a coined brand name in a new way, ask the user to confirm that spelling, save it as a pronunciation, and re-run.

Inputs

• --video PATH (required) — the rendered master mp4.
• --script-file PATH or --script "text" — the approved spoken script. Omit both only for a genuinely script-free clip (the drift check is then skipped and the gate is advisory).
• --brand-term TERM, --alias "TERM=SPOKEN", --pronunciations PATH — see above.
• --min-ratio FLOAT (default 0.90) — transcript↔script similarity to pass, measured on the canonical spoken form.
• --captions-srt PATH — optional SRT to check for caption-text defects.
• --json PATH — write the machine verdict for the app's review panel. Each issue carries the canonical tokens (script_words / heard_words) and the original wording (script_text / heard_text).

Tests

python3 tests/test_review_render.py    # or: python3 -m pytest tests/

Pure Python, no audio, network or paid call. Covers the QA-71 audit fixture table (units, URL, numbers, percent, contractions, fused brands, wrong price, negation, Hume→Hune), number words, units (including added/dropped units in long lines), URLs, negation flips, fused/split words, CLI-style brand terms, confirmed AG1 aliases and saved pronunciations with number-like syllables, omission, extra speech, the report wording, and the CLI exit codes with transcription stubbed out.

Relationship to the content-goose review engine

This is the shipped, single-file, gating slice of the fuller coworkers/video/molecules/review/review-loop (18-axis rubric). Here we enforce the one axis that catches audio-vs-script defects at publish time (review-transcript-integrity / brand_text_accuracy). Deeper multi-axis review stays in the content-goose lab.