14 Grammar Dictionaries in one

1. Download the dictionary ZIP
2. In Yomitan settings, go to Dictionaries → Import
3. Select the downloaded ZIP

You can also browse the grammar library or use the English website. Every dictionary entry links to its page, with all contributing explanations and examples. Search by grammar, reading or meaning, and filter by JLPT level or source. 絵でわかる日本語's explanations show their original illustrations, on the website and in the dictionary popup.

Japanese popup lookup covers all 9,515 source records across the 14 sources. Grammar notation such as Nらしい is indexed under its Japanese text (らしい), with kana aliases and conjugation support for known endings. Reference articles use Japanese topic names. See lookup coverage and examples.

The English Yomitan edition is separately installable. Monolingual explanations and untranslated examples are translated with GPT-6 Luna and labelled as translations. Source-authored English is retained. The English website includes the original source text for reference.

Every grammar point shows a compact card with progressive disclosure — click to expand per-source details, extra examples, and provenance.

• Single lookup: one dictionary, one card per grammar point — not 14 separate popups
• Merged entries: when multiple sources describe the same grammar point, their contributions are merged into one card with per-source attribution
• Deduplication: the keymap stage aligns entries across sources so you don't see the same point repeated
• Progressive disclosure: compact by default, expand any source's details on demand

python3 -m venv .venv
source .venv/bin/activate
pip install .
make

The default build recreates both full 14-source editions from the committed corpus and English translation cache, checks their pinned asset digests, and verifies every dictionary-to-article link. It needs no source downloads, Anki exports, scraper access or model credentials. The ZIPs, update indexes and SHA256SUMS are written to build/site/downloads/.

The raw 501-page IMABI acquisition is also committed under data/sources/imabi/, together with its digest lock, index and coverage report. It contains 495 lessons; six site pages are excluded. The publication snapshot preserves the complete merged corpus from every source, including the source decks that are unavailable in a fresh checkout.

IMABI lessons are also indexed by Japanese grammar forms, so looking up は, に, を, けど or のだ shows the relevant IMABI articles. The committed website/data/imabi-lookups.json maps 495 lessons to 1,400 Japanese lookup terms, using their titles, headings, constructions, readings and topic labels. Multiple lessons keep their own titles and explanations. Original article addresses remain available; the linked website includes every matching lesson. This includes the restored “About” lesson under について, に関して and をめぐって, plus compound constructions such as ことにする and ことがある.

To refresh the corpus from original inputs, acquire any missing sources and run make all. Those pipeline stages are:

1. extract — read each source's locked data into normalized GrammarPoint records
2. keymap — align entries across sources (which rows are the same grammar point?)
3. merge — combine aligned entries into a unified dataset
4. build — render the Yomitan dictionary ZIP
5. validate — verify against Yomitan's JSON schemas

The committed publication snapshot makes both websites and dictionary downloads reproducible without local source decks, network scraping, or model credentials:

make site
python3 -m http.server 8000 --directory build/site

Open http://localhost:8000/ or http://localhost:8000/en/. GitHub Actions builds this snapshot and deploys it to GitHub Pages on main. See the publishing workflow for updating the corpus and translations.

After acquiring all locked sources and running make all, update the publication:

make translate-english  # authenticated Codex CLI; explicitly uses gpt-6-luna
make publish-data      # refuses missing sources or translations
make site

After refreshing the snapshot, bump its revision, run make site, regenerate the release asset lock with python scripts/verify_release.py --write-lock, and run make verify-release. Commit the snapshot, lock and release notes. The release workflow builds and checks those exact inputs before publishing.

Translation results are cached by source-text hash under data/translations/. The original corpus stays unchanged. A changed source text requires a fresh translation; an incomplete English edition fails the build.

1. Place source data in data/sources/<name>/ with a SOURCE.lock.json
2. Write an extractor in src/bugd/sources/<name>.py (see existing extractors for patterns)
3. Register it with @register_extractor
4. Add to SOURCE_PRECEDENCE in keymap.py and _SOURCE_EXPLANATION_LANG in banks.py
5. Run make all