Seeds: atomic canonical claims for the #1137 corpora, lint before any registration

[Tailor-AUS]

Refs TailorAU/tailor-group#7 (step 2, data half — the fix the first apply needed). Base rehome-review.

What happened

Under Knox's go, the three topic corpora were applied against pact.tailor.au after #68 deployed. All 30 topic creations were rejected with canonicalClaim must be at most 140 characters (got 155…277) — the #3691 W2 atomic-claim rule (src/lib/claim.ts), which these corpora predate. The GET-only dry-run could not see a POST-side validation, and 30 seed agents (plus 1 for the edge seed) had already been registered when the rule fired. Topics 12 → 12, edges 0 → 0, agents 1 → 32. The 11 legislation documents were byte-identical before and after (sections + SHA-256 of each record); qld/act-2016-025 is not present in this database at all (404 before and after).

What this does

• Rewrites each of the 30 canonicalClaims to one atomic proposition ≤ 140 chars (longest 135). The longer text remains in content. Two claims drop their dotted rule numbers from the claim text — ASX LR 3.1 and DFARS 252.225-7052 — because the server splits sentences on .; the titles still carry the numbers. Please eyeball the 30 rewrites in the diff before the re-run.
• _defence_seed_helpers.py gains lint_atomic_claim / lint_corpus, a Python mirror of lintAtomicClaim (length, single sentence, top-level and/or joining two verb-bearing clauses, hedges). seed_topic_batch lints the whole corpus first — in dry-run and in apply — and exits 2 before registering anyone.

Evidence

• Lint: 0 failures across 12 + 10 + 8 claims; the old wording and a 141-char string both fail the mirror.
• SEED_DRY_RUN=1 after the rewrite: 30 to create, 0 present; 25 edges pending those topics.
• python -m py_compile clean on the four files.

Re-run note

The registration flood backstop is 60/hour per address; this seat spent 32 of those at ~05:05–05:15 UTC. The re-apply needs 31 more, so it runs after ~06:15 UTC or from a different egress.

🤖 Generated with Claude Code

---

Note

Low Risk
Changes are limited to offline Python seed scripts and README; they do not alter production API or app runtime, though operators must re-run seeds against live PACT with rate-limit awareness.

Overview
Fixes failed #1137 topic seeding where all 30 POST /api/pact/topics calls were rejected after agents had already been registered, because canonicalClaim violated the server’s atomic-claim rule (≤140 UTF-16 units, single sentence, no bundled conjunctions/hedges).

Data: Rewrites every canonicalClaim in the AU/US defence and critical-minerals corpora to one short atomic proposition; fuller text stays in content. Some claims drop dotted section numbers from the claim line because the server treats . as sentence boundaries.

Tooling: _defence_seed_helpers adds a Python mirror of lintAtomicClaim, runs lint_corpus before any agent registration (exit 2 on failure), resolves existing titles first and only registers agents for missing topics, and stops instead of round-robinning when registration returns fewer keys than needed (exit 3). Optional PACT_SEED_KEYS_FILE / PACT_SEED_AGENT_KEY persist and reuse keys so seed_topic_dependencies need not burn a 31st registration. create_topic distinguishes exact-title 409s from fuzzy near-duplicate 409s; pact_pow.register retries transient network errors.

Docs: README seed steps document dry-run, keys file, hourly registration budget, and default SOURCE_BASE (https://pact.tailor.au).

^{Reviewed by Cursor Bugbot for commit 5afacba. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_dca0339e-a76a-4f85-936f-297f60e563a0)

[Tailor-AUS]

Adversarial review record (dark seat, 21 Sep 2026) — refs TailorAU/tailor-group#7.

Four independent lenses read this PR at e4ca1e6 against the deployed server code, plus three skeptics per blocker/high finding:

• Lint parity. All 30 rewritten claims pass the real lintAtomicClaim on the route path (sanitise, then lint): 30/30 ok, 0 warnings, max 135 chars, 0 exact or fuzzy (≥ 75 %) title collisions among the 30 (nearest pair 0.727). The Python mirror diverged from the server on 8 of 72 adversarial cases (UTF-16 vs code-point length, Unicode \b/case-folding vs the server's ASCII regexes, no control-char/HTML strip before linting); none fired on this corpus.
• Claim fidelity. 25/30 faithful or defensible narrowings. Confirmed by the refute pass: the DTCA claim introduced "to a foreign person", which neither the topic content nor the cited Act (s 10, not the 2024 s 10A limb) states. Also replaced: BLM 3809 ("beyond casual use" → the Plan of Operations for commercial mining the content describes; the skeptics rated the original a low-severity nit, the replacement is simply the more faithful wording), ASX (restores the reasonable-person test), DFARS (drops "such as antimony", outside 10 USC 4863), Critical Minerals List purpose clause, the antimony "80 percent of mine production" figure the content's own 40,000 t / 83,000 t contradicts, NEPA (4321 et seq, "human environment"), the NSLA EFI Act's real name, and the rare-earth claim reduced to one proposition.
• Helper logic. Every orphan-agent path from the pre-flight was confirmed at e4ca1e6 by all three skeptics and reproduced offline: registration before any existence check, round-robin on shortfall into civic-duty 403s that a fresh agent cannot vote out of, the dependency script registering before resolving titles, any 409 with existingTopicId treated as EXISTS, an unguarded POST in pact_pow.register, and a registration budget exactly saturated (60/hour per address, two hits per registration, 31 registrations).
• Local rehearsal (local Postgres + next dev, production PoW bits, no overrides): the three topic scripts created 30/30 topics (30 × 201, no 422/403/409); the dependency script's registration was hit 61 → 429 → exit 1 with 0 edges; after a limiter reset the same script created 25/25 edges. A re-run of seed_defence_au.py minted 12 orphan agents for 12 × 409 EXISTS.

Commits c4ef569 and 5afacba carry the fixes: the nine claims (re-verified 30/30 on the route path), the mirror hardening, existence check before registration, abort (exit 3) instead of round-robin, dependency registration after resolution, exact-title-only 409, network retries in the PoW client, key reuse for the edge step (PACT_SEED_AGENT_KEY / PACT_SEED_KEYS_FILE, mode 0600) so the 31st registration is never needed, the README seed block, and a dry-run footer that no longer reads "0/12 topics in place". All seven offline harness scenarios pass on the patched scripts.

Merging under the 2026-09-20 finish-authorised-work ruling once the checks are green. The live apply runs from this revision, topic scripts one at a time, then the dependency script with a reused key.

---

Generated by Claude Code