Japanese text consistency check with line numbers and unified forms

What it does

Checks Japanese documents (Markdown, text, HTML) read-only and reports notation inconsistencies with line and column numbers.

• Mixed spellings such as ichiban in kanji vs "1-ban", dekiru and kudasai in kana vs kanji, okotowari in kanji vs kana, "please forgive" vs "please acknowledge" phrasing, and okurigana variants of inquiry and application
• Katakana words with and without the trailing long vowel mark (server, user and more), including words that are not in the rules file but appear both ways in the same document
• Full-width and half-width digits and letters mixed together, and half-width katakana
• Polite desu/masu sentences mixed with plain da/dearu sentences (estimated from sentence endings)
• Code blocks, inline code, URLs and HTML tags are skipped

All rules live in an editable JSON file (references/rules.json), so you can set the preferred form and choose between reporting only when variants are mixed or always enforcing a house style. The agent reviews each hit in context, removes false positives and reports the result.

How to use

python scripts/jp_consistency.py docs/ README.md --out report.md
python scripts/jp_consistency.py docs/ --format json --out report.json --rules my-rules.json

Ask Claude, Codex, Hermes Agent or Cursor to check a document for notation variants. The agent runs the script and lists confirmed findings, proposed unified forms and items that need a human decision. It edits files only when asked.

Requirements

• Python 3 (standard library only)
• UTF-8 text (Shift_JIS files can also be read)

Caveats

• Detection is mechanical: ranking numbers, wording with different meanings, quotations and proper nouns need human judgment.
• Sentence style is estimated from sentence endings; sentences ending in a noun are not classified.
• Counts are per file; check the JSON output for variants spread across files.
• Input files are never modified.