audit-translations

name	audit-translations
description	Audit Scinoephile standard, gapped, or guided translation JSON against the exact source, target-context, and guide SRT files used to generate it. Use when inspecting translation, gap_translation, or guided_translation JSON; judging translated, empty, unnecessary, or unanswered outputs; filtering by verification state; or correcting and verifying translation cases.

Audit Translations

Run commands from the repository root. Select the workflow from the JSON and the inputs that produced it; do not infer it from language alone.

Select the inputs

Workflow	Required SRT inputs
Standard translation	source track
Translation of missing target positions	gapped target and complete guide
Translation using target-language guide context	source and guide

The scinoephile audit translation command infers the workflow from the SRT inputs.

Required report file

Save the complete Markdown report under local/, audit every row, and write each judgment into its Notes cell.

Standard and guided reports have exactly: Indexes, Case / block, Difficulty, Source, Guide, Translation, Notes, Verified. Gapped reports have exactly: Indexes, Case / block, Difficulty, Guide, Target context, Translation, Notes, Verified.

• In standard and guided reports, S is the global source index and Q is its query-local JSON index; guided G labels are query-local guide indexes. In gapped reports, G is the global guide index and Q is its query-local JSON index. Always use Q when editing an answer.
• C is the one-based JSON case position and B the reconstructed block.
• ✓ means the entire JSON case is verified, not merely one displayed row.
• Translation JSON has no note field, so generated Notes cells begin blank. If a future schema populates one, read it as context and replace the entire cell with your own independent judgment; never append to or merely endorse generated note content.
• Keep findings beside rows, validate the edited report, and provide its link. Do not paste the table inline unless the user requests it.

Protect source data

• Never edit files under test/data/<dataset>/input/.
• For an audit-only request, do not modify input SRTs, JSON, generated translations, or downstream review artifacts.
• When corrections are requested, edit JSON answers and verification metadata, then regenerate outputs. Never hand-edit a generated translated SRT.
• Preserve every JSON query exactly.

Locate exact inputs

Standard translation uses the source SRT passed to the regular translator and its translation.json. Guided translation uses the source and guide SRTs passed to guided translation and its guided_translation/<device>.json.

Gapped translation uses:

• target: the gapped target before translation, commonly transcribe_clean_review.srt or transcribe_review.srt
• guide: the complete source-language SRT used to identify gaps
• JSON: <target-output>/lang/<target-source>/gap_translation/<device>.json

Never pass the already gap-translated output as --target; doing so removes the gaps the audit must reconstruct.

Generate the report

Always set UV_CACHE_DIR=/tmp/uv-cache.

Standard:

UV_CACHE_DIR=/tmp/uv-cache uv run scinoephile audit translation \
  --source <source.srt> \
  --json <translation.json> \
  --filter all \
  --outfile local/<dataset>_translation_audit.md \
  --overwrite

Gapped:

UV_CACHE_DIR=/tmp/uv-cache uv run scinoephile audit translation \
  --target <gapped-target.srt> \
  --guide <complete-guide.srt> \
  --json <gap_translation/device.json> \
  --filter all \
  --outfile local/<dataset>_gap_translation_audit.md \
  --overwrite

Guided:

UV_CACHE_DIR=/tmp/uv-cache uv run scinoephile audit translation \
  --source <source.srt> \
  --guide <target-language-guide.srt> \
  --json <guided_translation/device.json> \
  --filter all \
  --outfile local/<dataset>_guided_translation_audit.md \
  --overwrite

Use inclusive --first-index and --last-index. They refer to global source indexes in standard and guided workflows, and global guide indexes in the gapped workflow. Use --first-block and --last-block for inclusive, one-based workflow block ranges: source blocks in the standard workflow and paired blocks in guided or gapped workflows. Subtitle-index and block ranges are mutually exclusive. Omit either bound for an open-ended range. Use --filter unverified to resume verification; the default is all.

(unanswered) means the case has no answer. (empty) means an answered output emitted empty text; in the gapped workflow this may intentionally mean that no target subtitle is needed.

Audit every translation

Judge whether each output accurately and idiomatically expresses its source in the intended target language and script.

• Check omissions, additions, mistranslations, polarity, modality, pronouns, names, numbers, register, terminology, continuity, and meaning-changing punctuation.
• Preserve natural target-language grammar rather than copying source syntax.
• In the guided workflow, use the guide as target-language context and semantic evidence, not as text that must be copied.
• In the gapped workflow, treat existing target subtitles as context, not as text to rewrite. Accept an empty output only when the guide position genuinely needs no target subtitle.
• Treat every unanswered case as incomplete.

Write exactly OK for a clearly appropriate translation. Otherwise begin with:

• Incorrect translation; for wrong, unsupported, unidiomatic, or wrong-script output
• Missing translation; for an empty or unanswered gap that needs text
• Unnecessary translation; for a generated gap that should remain empty
• Uncertain; when the source, guide, target context, and optional audio do not establish the correct output

Audit every displayed row. Do not mark an unanswered row OK.

Correct and verify cases

When corrections are requested:

• For standard and guided cases, keep one output for every source Q index in ascending order.
• For gapped cases, keep one output for every guide Q index absent from query targets; use an empty string only for an intentional no-subtitle result.
• Mark verified: true only after auditing and correcting every output in the case. Do not verify a case when the selected range contains only some rows.
• Never verify unanswered or unaudited cases.

Regenerate translated and downstream artifacts through the dataset workflow. Rerun the corrected range, confirm canonical JSON, and confirm corrected cases leave --filter unverified before linking the interpreted report.