docx-format-audit

name	docx-format-audit
description	Use when a `.docx` or Word task requires paragraph-by-paragraph translation QA, format review, bilingual document audit, formula preservation checks, 逐段核对, 翻译审校, or validation before deleting Chinese text.

DOCX Format Audit

Audit bilingual or English-only Word translation outputs before finalization. This skill checks structure and formatting; it does not perform translation by itself.

When To Use

Use this skill when:

1. a translated Word document needs paragraph-by-paragraph review;
2. Chinese and English paragraphs must be compared;
3. table translation coverage must be checked;
4. superscripts, footnotes, formulas, or embedded formula images may have been lost;
5. Chinese text will be deleted after a bilingual review pass.

Do not use this skill for non-DOCX code, finance, or general audit tasks.

Quick Reference

Need	Required action
Bilingual translation QA	Match btx_<id>_src to btx_<id>_en; missing pair markers are blocking
Table-cell review	Require source and English pair in the same w:tc
Formula preservation	Compare per-pair OMML, drawing, VML, and OLE signatures
English typography check	Confirm inserted English resolves to Times New Roman
English-only cleanup	Run audit first, remove source paragraphs, then audit with --mode english-only

Baseline Failures Addressed

The RED baseline capture must show at least one omitted guardrail before this skill is created. This skill closes the observed baseline gaps for DOCX audit work: deleting Chinese text without pair-ID checks, accepting missing markers, auditing only body paragraphs while missing tables, counting formulas at document level instead of per pair, or accepting English-only output that still contains Chinese text. If a future baseline covers one of these without the skill, remove that claim from this section instead of keeping untested guidance.

Blocking Checks

Fail the audit when:

1. a Chinese source paragraph has no pair-ID-linked English translation;
2. a table-cell Chinese paragraph has no pair-ID-linked English translation in the same cell;
3. an unmarked Chinese body or table-cell paragraph remains in a bilingual output;
4. a source formula signature has no corresponding English pair signature;
5. a superscript, subscript, footnote reference, or endnote reference is missing from the English pair;
6. inserted English text does not resolve to Times New Roman through direct formatting or style inheritance;
7. English-only output still contains Chinese characters outside allowed names or citations;
8. the output DOCX cannot be opened by python-docx.

Warning Checks

Warn on:

1. possible terminology inconsistency;
2. duplicate phrases such as the the;
3. large drawings that may be figures rather than formulas;
4. missing LibreOffice when headless conversion was requested.

Pair And Formula Model

Use stable pair IDs. Missing pair markers are blocking because adjacency is not safe enough for automated audit or deletion. Formula preservation should compare per-pair signatures, including object type, relationship target, dimensions, media hash, file name, and paragraph/table-cell context.

Runnable Helper Example

From the package root, this verifies that the audit helper is importable:

python3 - <<'PY'
import importlib.util
import sys
from pathlib import Path

audit_path = Path("skills/docx-format-audit/audit_docx_translation.py")
spec = importlib.util.spec_from_file_location("audit_docx_translation", audit_path)
assert spec is not None and spec.loader is not None
module = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = module
spec.loader.exec_module(module)
assert callable(module.main)
print("PASS: audit helper import")
PY

Works With

• Use after docx-bilingual-translation.
• Use with word-formula-writing for formula-heavy documents.
• Use with word-default-formatting when typography and layout must be validated.
• Use before using any English-only deletion helper.

Common Mistakes

Mistake	Fix
Auditing only marked pairs	Also block unmarked Chinese body and table-cell paragraphs
Treating adjacency as proof	Treat missing pair IDs as blocking; add bookmarks before audit cleanup
Counting document-level formulas	Compare formula signatures for each source/English pair
Deleting Chinese before audit	Run bilingual audit first and refuse deletion on blocking findings
Ignoring English-only output	Re-run audit in English-only mode after removal