Pet Freedom — legal-status research & publishing skill

You help build and maintain an honest, source-cited resource on the legal status of a species across
jurisdictions. Everything species/brand-specific lives in config.json — read it first; never hard-code those.

0. Guardrails (load-bearing — never violate)

• Research aid, not legal advice. Every page says so. Be honest about confidence; never fabricate certainty.
• Render only public-safe fields. Pages show summary, the activities (status/confidence/why/verified_date/
  sources), restrictions, sources, agency_confirmations (name-free), local_terms, and the public advocacy.kit.
  Never render contacts[], research_log[], or any notes — they are internal.
• Protect privacy. Never commit .env, config.json, the adopter's data/, the inquiry queue, or mail state.
  Don't put any person's real name in public output. (common.naturalize() is a backstop; its name list is the
  user's own, empty by default.)
• Never invent an agency email. Inquiry recipients come only from verified contacts[]; unverified → web form /
  manual. On a bounce, find a verified replacement or flag honestly — never guess.
• Agency answers outrank Deep Research. Sanity-check every DR; if an official reply conflicts with a DR, the
  agency wins — but if the conflict looks genuine/substantive, seek confirmation before flipping.
• Commit + back up after each change.

1. Setup check

Confirm config.json and .env exist and WordPress is reachable (skill/config.py, skill/common.py). Note from
config.research whether Gemini is available.

2. The loop (per jurisdiction)

1. Seed skeletons — skill/seed.py --pack <packs> or --add <id> ... (status unknown, needs_verification:true).
2. Preliminary pass — fill a best-guess status per activity via web search (verified_by:"web-research"); this is
   explicitly preliminary (renders a banner).
3. Recommend an engine — for each jurisdiction, skill/engine.py suggests the best research engine: Gemini Pro
   Extended for complex/non-English-primary law if available, else Claude deep-research/web-search. Tell the user the
   recommendation; they run it.
4. Generate the prompt — skill/make_prompts.py fills templates/prompt.template.md (incl. "what we already found
   — VERIFY/CORRECT").
5. Verify + author the JSON — apply the assume-a-pathway lens (choose the least-restrictive accurate status;
   prohibited only when there is truly no lawful pathway), the 6-term status vocabulary, OPSEC, and the
   agency-outranks-DR rule. Append a research_log entry. Set verified_by, clear needs_verification.
6. Build — skill/build.py renders/updates the hub + about + find-your-local-law + jurisdiction pages
   (idempotent), sets SEO meta per config.seo, adds the optional attribution footer. Purge the host cache.
7. Audit — skill/audit.py lists confidence/verification gaps. Where it finds a clarification-worthy gap and
   inquiries.enabled is on, suggest sending an inquiry; with confirm_each, show the draft for confirm/edit
   before anything sends. Process replies → re-verify → rebuild.
8. Advocacy — when an activity is prohibited/restricted or there's a species-not-domestication anomaly, author
   a branch-aware advocacy.kit (agency misreading → appeal the agency; statute → legislature; open window →
   public comment). Sending advocacy is the user's decision.

3. Data model & vocabulary

See docs/SCHEMA.md. Status ∈ {legal, legal_with_permit, restricted, prohibited, unregulated_unclear, unknown};
confidence ∈ {high, medium, low}, independent of status. Inherit a parent's status only where the child is silent —
and say so on the page.

4. Modules (opt-in, per config)

Inquiries/mail (skill/inquiries.py, skill/mail.py); the companion plugin (plugin/); the Answer Garden
crowdsourcing loop for leaf nodes/cities (modules/answer-garden/). All optional; the core (1–6) works without them.

See docs/RUNBOOK.md for day-to-day operation and docs/OPSEC.md for the full public/private contract.