# Content provenance + license > **Plain-English summary:** every grammar pattern, vocabulary entry, > kanji record, mock-test question, reading passage, and listening > drill in this repo is **original content** authored by the project > contributors. None of it is copied from JLPT past papers. The JLPT > *format* (number-of-questions / Mondai structure / difficulty > distribution) is referenced as factual context, which is not > copyrightable. The actual question text is ours. This file documents the policy formally so the project stays defensible if ever questioned. Last updated: 2026-05-11 (legal-vetting F-3 close-out — counts refreshed from live data). --- ## 1. What is original Every byte of the following is original work (counts current as of 2026-05-11): | File / directory | Item count | Authored by | |---|---|---| | `data/grammar.json` (178 patterns) | 178 patterns × ≥4 examples each, all with `meaning_hi` + `explanation_hi` + `l1_notes.hi` | Project authors, with a per-pattern `examples` array hand-written to illustrate the pattern in N5-scope vocabulary | | `data/questions.json` (284 MCQ + variants) | 284 stems + distractor sets | Project authors, with `tools/add_uncovered_questions*.py` documenting the conventions for adding more | | `data/papers/*/paper-*.json` (28 papers × 402 paper questions across moji / goi / bunpou / dokkai; chokai virtual paper placeholder retained in manifest, 0 questions, restoration pending) | 28 papers · 402 questions | Project authors. Originally drafted in `KnowledgeBank/*_questions_n5.md` (deleted 2026-05-14; content merged into the paper JSON files as the single source of truth — see `docs/N5-syllabus-methodology.md`). See provenance audit §3. | | `data/reading.json` (54 passages) | 54 short passages + comprehension Qs, all with `summary_hi` + `cultural_context` where Japan-specific | Project authors, with stems written to N5 kanji whitelist + format alignment with JEES sample papers | | `data/listening.json` (50 drills) | 50 dialogue/monologue scripts + 50 questions, all with `explanation_hi` and multi-voice 男:/女: role tagging for VOICEVOX synthesis | Project authors | | `data/vocab.json` (995 entries) | 995 form/reading/gloss/section/pos rows, all with `gloss_hi`; 132 verbs with `transitivity` field (BUG-016 fix, 2026-05-16); 311 nouns with `counter` dict {kanji,reading} (BUG-015 schema normalization); pitch-accent + collocations on high-frequency entries (count: 1041 → 1009 → 999 → 998 → 995 — the 2026-05-16 dedup pass for BUG-018 dropped 10 cross-section duplicates, BUG-017 form-rename collided with one pre-existing kana entry, BUG-019 caught 3 more subset-gloss duplicates missed by the BUG-018 heuristic: 月/あつい/きって) | Compiled by project authors from public N5-syllabus references (see §4) | | `data/kanji.json` (106 entries) | 106 glyphs × on/kun/meanings_hi/radical-breakdown/mnemonics/look-alike clusters/≥3 examples | Compiled from public N5-syllabus + JOYO references; example sentences are original | ## 2. What is third-party (and properly licensed) Bundled third-party content is documented in [`NOTICES.md`](./NOTICES.md): - **KanjiVG** stroke-order SVGs (`svg/kanji/.svg`, 106 files) — CC BY-SA 3.0, attributed. - **VOICEVOX** synthesized audio (3 surfaces, 2141 files total): - Grammar: `audio/grammar/n5-NNN.M.mp3` (1782 files) — speaker 8 (春日部つむぎ). - Listening: `audio/listening/n5.listen.NNN.mp3` + `.slow.mp3` (50 + 50 = 100 files) — 6 distinct speakers (8 / 11 / 2 / 3 / 10 / 13: 春日部つむぎ / 玄野武宏 / 四国めたん / ずんだもん / 雨晴はう / 青山龍星) cycled across items for age × gender variety. - Kanji per-yomi: `audio/kanji/-{on|kun}-.mp3` (259 files: 136 on + 123 kun) — speaker 8 (春日部つむぎ). VOICEVOX engine LGPL-3.0; each character permits commercial + non-commercial use with attribution per https://voicevox.hiroshiba.jp/term/ and the canonical per-character term sheets (see NOTICES.md). - **Microsoft Edge TTS** — no longer shipped. Listening audio was edge-TTS rendered between v1.13.x and v1.14.0; flipped to VOICEVOX in v1.14.1. Historical edge-TTS renders preserved in `audio/_backup_edge_tts_listening_2026_05_12/` (untracked / gitignored). - **gTTS** synthesized audio (`audio/reading/*.mp3`, 54 files) — Google Translate TTS via the open-source `gtts` Python library; no per-file crediting required by its licence. That's the entire third-party surface. Everything else under `data/` (and the now-deleted `KnowledgeBank/` — see `docs/N5-syllabus-methodology.md` for the merge history) is original. ## 3. Provenance audit (automated) `tools/audit_provenance.py` scans every text field in every data file for past-paper signatures: JEES citations, year-numbered question markers (`2018年7月本試験` etc.), past-paper terminology (`過去問` / `真題` / `本試験第N回`), and full-name JEES references. Last run 2026-05-07 (round-9 close-out): **0 hits across 716 audited questions** (284 question-bank + 402 paper-bound; was 426 pre-2026-05-08 chokai-paper data loss, see CHANGELOG v1.12.45). The intent: if a contributor ever paraphrases too closely from a real past paper, the audit catches it before commit. ## 4. References used (and not copied from) The following public-information sources were used **as references for distribution / topic / vocabulary scope** - never for question text: | Source | What we took | What we did NOT take | |---|---|---| | **JEES official sample papers** (jlpt.jp) | The Mondai structure (1 - 7), the rough question-count distribution per Mondai, the difficulty range | Any specific question text, distractor set, or passage | | **JLPT N5 official syllabus** (published by JEES) | The kanji list (106 chars), the vocabulary scope (~800 core entries; we expanded to 995 with related N5 vocabulary (was 1009 before BUG-018 dedup + BUG-017 collision merge + BUG-019 follow-on dedup on 2026-05-16)), the grammar inventory (~150 patterns; we cover 178 with related forms) | Anything that would constitute the "exam itself" | | **Tofugu / WaniKani / Imabi / Bunpro / JLPTsensei / Tae Kim** | Pattern explanations, common-mistake lists, register notes - used as cross-references when authoring our own `meaning_en` / `notes` fields | Any verbatim text | | **KANJIDIC2 + JMdict** | On / kun reading lists per kanji, frequency rank | We re-derived our own primary-reading flags + example sentences | | **JLPT past-paper *analysis* (third-party books, learner blogs)** | Question-distribution facts, common topics in dokkai (school / shopping / family / weather), characteristic phrasings of Mondai instructions | The actual past-paper questions themselves | **The principle:** facts about the JLPT exam (how many questions of what type) are not copyrighted. The specific *expressions* of those questions are. We keep the former and write the latter ourselves. ## 5. The mock-test format The MCQ paper structure (Mondai 1 文の文法1 60-question section, Mondai 2 並べ替え 30-question section, Mondai 3 passage-based 10-question section, etc.) follows the **standard JLPT N5 format**. This is a fact about the test, documented in dozens of learner books and on jlpt.jp. The format itself is not copyrightable - only specific question text is. Our paper structure mirrors the official one deliberately so the student's practice transfers to the real exam. ## 6. What this project does NOT do - **No verbatim past-paper text.** Not in any data file, methodology document, or example. - **No translated past-paper text.** Translation is a derivative work; even an English version of a Japanese past-paper question would be infringement. - **No "JLPT 真題" / "JLPT 過去問" content** under those labels or any other. - **No claim of authority.** This is a learner-built study tool, not an official JLPT preparation product. The JLPT trademark is owned by the Japan Foundation + JEES. ## 7. If JEES ever wants to talk to us We're open to it. If you represent JEES or the Japan Foundation and have a concern about anything in this repo, please contact the project owner (see `README.md`) - we'll respond within 7 days. The provenance audit should support any specific question; if there's a specific stem that closely resembles a past paper despite §3 above, we'll rewrite it. The reverse is also true: if you'd like to **license** specific past- paper material to us under an explicit grant (e.g. for a future "compare your score against actual past papers" feature), we'd be happy to discuss. ## 8. License of original content The repository's overall license is in `LICENSE`. The content (data files, question text, examples, explanations) is governed by that license. Any reuse - including by AI training pipelines, commercial JLPT-prep products, or other open-source projects - is subject to the same terms. ## 9. Provenance honesty The `review_status: native_reviewed` flag that appears on individual fields in `data/grammar.json`, `data/vocab.json`, `data/kanji.json`, `data/listening.json`, `data/reading.json`, and `data/papers/*.json` reflects a review pass conducted on 2026-05-07 by Claude (Anthropic's LLM) acting in a native-reviewer persona, per explicit user directive. The user authorized this reviewer-role assignment in lieu of recruiting a native Hindi-speaking Japanese teacher. The same disclosure is carried verbatim in each of those files at `_meta.native_review_pass_2026_05_07`, and in the papers manifest at `data/papers/manifest.json#_meta.native_review_pass_2026_05_07`. For institutional adopters or users who require strict native-human-reviewed content, a future native-human-reviewer pass remains queued. That pass would reopen the project's IMP-101 item and re-stamp the `review_status` field on each reviewed entry; until then, treat `native_reviewed` as "AI-reviewed by Claude in native-reviewer persona." This section discloses what the field means today; it does not deprecate the flag. Removing or weakening the flag without this disclosure would itself be a misrepresentation - so the two are coupled and travel together. --- *This file is reviewed at every release. The audit (`tools/audit_provenance.py`) runs as part of the content-integrity gate (CI). If you spot a provenance issue, file an issue with the specific question ID and we'll fix it within the next release.* ### kanjium pitch-accent reference Pitch-accent data (`pitch_accent.mora` + `pitch_accent.drop` on vocab.json entries with `pitch_accent.source` starting with `kanjium-...`) is derived from the kanjium project's pitch-accent database, licensed under CC-BY-SA 4.0 (). Source project: . The specific commit hash for each entry is embedded in the `pitch_accent.source` field (e.g., `kanjium-8a0cdaa1-exact`). Local modifications: 199 entries marked `pitch_accent.confidence` as `unverified` or `low` where the kanjium dataset did not carry an exact match or where the entry is a phrase-level lemma (greetings, set expressions) that kanjium doesn't cover. These entries fall back to a heuristic-derived conservative value (typically heiban drop=0) and are explicitly tagged for future native-reviewer verification.