#Reviewing Translations

Review is where translation quality is finalized. A good review checks meaning, terminology, readability, and exported formatting — in that order. In CatScribe, review happens in the CAT Editor, driven by chunk statuses, QA flags, and quality scores.

#The Review Unit: Chunks

CatScribe splits documents into chunks, and each chunk carries an editor status:

• "Draft (Machine)" — untouched machine output.
• "Edited (Human)" — you changed the text (set automatically on any edit).
• "Approved" — your explicit sign-off. Approved chunks also feed the local translation memory used for consistency in later jobs.

The editor sidebar shows "Review progress" (approved/total and percent), so "done" is a measurable state, not a feeling.

#Efficient Review Flow

1. Open the job from the Translations screen ("Open CAT editor").
2. Triage first: filter with "Show only failed chunks" in the job's chunk grid. Failed chunks show the banner "This chunk failed. Type your translation below and Save, or use Re-run chunk with a different engine." Fix them by hand or via the "Chunk Override" panel.
3. Score-guided pass: work through low-scoring chunks next. The chunk score blends quality estimation and glossary compliance ("Score = (2×Quality + Glossary) / 3"). Treat short segments with skepticism — the editor marks them "⚠ Short segment" because quality estimates on little context are unreliable.
4. Terminology pass: resolve glossary QA flags, then bulk-fix repeated decisions with "Find and Replace" rather than chunk by chunk.
5. Style pass: read for fluency. Use "Improve with AI" on awkward chunks — you can add an instruction like "make the text more natural but keep names unchanged". Do this only after terminology is stable, so the AI works within your decisions. See Improve With AI.
6. Approve as you go: Ctrl+Enter approves the current chunk and jumps to the next unreviewed one. This keyboard loop is the fastest way through a long job.

#Keyboard Review Loop

Note: undo history is per chunk and lost when you switch chunks. For persisted rollback, use the "Version History" panel — every save creates a version you can "Revert" to.

#Reading QA Flags

The "QA Flags" panel reports mechanical problems the machine can catch so you can spend attention on what it cannot. Flags include:

• Empty translation, or translation identical to the source.
• Tag count mismatch between source and target (broken <i>, <b>, <p> pairs).
• Placeholder count mismatch.
• Numbers or units that differ between source and translation.
• Suspicious repetition, or punctuation counts that differ significantly.
• Glossary terms not translated as one of the approved targets, plus an overall glossary success rate.

Clear every flag deliberately: fix it, or consciously decide it is a false positive (for example, a number legitimately reformatted for the target locale).

#Quality Scores: What They Mean

• Quality / COMET — with "Enable COMET quality scoring" turned on in Settings (a ~1–2 GB one-time download), chunks get a neural quality estimate that needs no reference translation. Scores are 0–100: 85+ Excellent, 70+ High, 50+ Medium, below that review carefully.
• BLEU — only meaningful when you can paste a human reference translation in the same target language; it is never computed from source vs translation alone.
• "Recalculate score" in the editor toolbar refreshes a chunk's score after edits (preview only).
• The "Layers" button opens the "Translation Layers Inspector", showing what each pipeline stage (base translation, AI validation, rewrite, technical check, your overrides) did to the chunk — useful when you want to know why a chunk reads the way it does.

Use scores to prioritize, never to approve. A high score means "probably fine"; approval means a human read it.

#Review Passes

#Release Checklist

Before delivering:

• Review progress shows "All chunks approved" (or you have consciously accepted the remainder).
• No unresolved failed chunks.
• QA flags cleared or explicitly waived.
• Names spot-checked at the start, middle, and end of the document.
• Exported file generated ("Generate Document" in the Translations screen) and opened in the target reading app.
• Search the export for a few distinctive source-language words to catch untranslated passages.

#Common Pitfall

Avoid correcting isolated chunks without context. Read the neighboring chunks before final edits — especially in fiction and subtitles, where meaning often depends on the previous line. The Source pane shows a paragraph-count badge when a chunk packs multiple paragraphs; read the whole chunk, not just the flagged sentence.