日英翻訳の用語集チェック（訳語のゆれ・訳し漏れ）

What it does

Keeps terminology consistent when translating between Japanese and English. A glossary CSV is the single source of truth, and the script compares the source and target files read-only.

• Glossary terms in the source whose approved translation is missing from the target
• Discouraged translations (for example "log in" when the glossary says "sign in")
• Do-not-translate terms (brand names, product names, commands) that were translated or dropped
• Untranslated leftovers: Japanese characters in an English target, runs of 4+ English words in a Japanese target, or lines identical to the source
• Candidate terms that recur in the source but are not in the glossary (--suggest)

When both files have the same number of non-empty lines, they are compared line by line, so each finding points at the exact lines. Code, URLs and HTML tags are skipped. A glossary template (ja, en, dnt, avoid, notes) is included.

How to use

python scripts/glossary_guard.py --glossary glossary.csv --source ja/guide.md --target en/guide.md --src-lang ja

The agent confirms the glossary, translates with it, runs the check, fixes findings and re-runs, then reports the remaining items with reasons and proposed glossary additions.

Requirements

• Python 3 (standard library only)
• UTF-8 CSV and text or Markdown files

Caveats

• It checks terminology and leftovers only, not fluency or accuracy.
• Japanese terms match as substrings, so very short terms cause more false positives.
• Add irregular English forms (such as "signed in") to the glossary explicitly.
• No file edits and no network access.