How to add an ethics / regional / operational rule¶

Field rules (a retired NTP server, a licensing drift, a jurisdiction's data rules) live as sections under _shared/ethics/, registered in its REGISTRY.yml. The design goal: a new rule must not move the questionnaire, the witness leaf space, or anyone's render — it accumulates as documentation first and is promoted only when a distribution bundle forms. Who decides what goes in the generated projects is stated in GOVERNANCE.md: contributors draft and research; the maintainer owns the defaults and the promotions. The mechanics behind this page are explained in Authoring Template Sources ("Accumulate ethics/regional/operational rules as sections first").

Add a rule as a draft¶

A draft is documentation only: nothing includes it, no render changes, and every guard accepts it without any other file being touched.

1. Copy _shared/ethics/_template.md.jinja to the matching bundle directory — baseline/ (rules any project kind can trip), region/ (a jurisdiction's rules), sector/ (an industry's), domain/ (a purpose the questionnaire's domain_traits select) — as <id>.md.jinja.
2. Fill in every field. The section opens with the matching header and a scope blockquote; the first five lines are the LLM/human entry point: if none of the triggers apply, the reader skips the section.
3. Register one row in _shared/ethics/REGISTRY.yml with status: draft and rendered_from: []. Keep the id, version (YYYY-MM-DD.rev), and status in agreement with the file header — the guards check both directions.
4. Check your work:

uv run --locked pytest tests/test_ethics_registry.py

The guards pin: row shape (id, version, dates, primary sources, scope, scale, audience, enforcement), header agreement, the body naming its scale and every audience entry, a documented presence trigger (設定側トリガー) that compiles and stays unique across rows, and — for drafts — that no template file includes you.

Write the rule as a one-line summary plus a link to the primary source; never paste statutes or license texts verbatim, and keep the not-legal-advice footer. Name concrete shapes in 禁止パターン and copy-pasteable settings in 推奨設定, not slogans — back the rule with at least one bad/good code-or-config pair, the way the active sections do. Keep snippets free of {{ }} / {% %}: the section file is rendered as jinja, so double braces would be evaluated, not printed. And keep ```python examples ruff-format-clean: ruff 0.16+ formats Python code blocks inside markdown, so a stylistically-off example fails the generated project's task check (the registry test catches this before any render — a "bad" example is bad semantically, not stylistically).

Promote a draft to active¶

A draft promotes only when a bundle forms: three sections sharing one distribution condition, or one section needing a distinct code/test gate (enforcement L2). A promotion changes generated renders, so it wires the verification stack as it goes — each step below has a check that catches its omission.

The audience in the registry row is the human-readable word for the gate (iot may map to micropython, as baseline-pki-chain does); the gate and the selector spell the mapping, the row keeps the word.

Enforcement levels¶

New sections start at L0. A L1 lift is free — the trigger span is already parsed for every row; drafts warn through check_ethics even while they ship nowhere. A L2 lift needs a real gate with an offline contract (see how license-check stays out of type-check/check).

The sections this template currently ships¶

The full inventory lives in _shared/ethics/REGISTRY.yml; the two most recently promoted rows are:

• domain-scraping-law — L0, gate scraping_effective (the AGENTS.md include is {% if scraping_effective %}). The scraping-law map: copyright and the EU database right (Directive 96/9/EC art. 7) sit beside the contract question (ToS) and the unauthorized-access line (CFAA / 不正アクセス禁止法), and robots.txt is evidence of the operator's intent, not an authorization (hiQ Labs v. LinkedIn, 9th Cir. 2022). Its generated counterpart is LEGAL.md — the per-source pre-crawl checklist every scraping layer ships, asserting what the toolchain enforces vs what stays the operator's duty.
• domain-contest-rules — L0, gate oj_code (the code-submission judges; kaggle and ctf are deliberately outside it). Competition integrity: AI-use rules differ per judge and change (AtCoder bans generative AI during ABC/ARC/AGC except a whitelisted translation prompt), the submission is the entrant's own responsibility, and rating manipulation / multi-accounting are banned.

Both are status: active with rendered_from naming the AGENTS.md appendix, and the ethics-appendix predicate in tests/test_render_invariants.py holds each to the leaf classes its gate selects — the mutation that drops a gate fails the render sweep, not a human review.

Structured data: the lang/ dictionaries¶

Rules whose substance is a table (protection terms per jurisdiction, per-country filing deadlines) keep the prose summary in the section and the data in _shared/ethics/lang/<name>.yml — see copyright-terms.yml, the first one. The table names the section it serves (serves:), matches its review_by, carries primary sources per row, and the section references it back; a guard in tests/test_ethics_registry.py holds the two together.

Keeping sections alive¶

Every row's review_by is the date its 制度変更ウォッチ items must be re-checked by. A passed date fails the registry test, so CI reddens until someone re-checks the primary sources, updates the body, and bumps the row's version and review_by (or supersedes the row via superseded_by). The re-check is a good first contribution; the bump is a one-line registry edit.