verify_plan c53: extend harm-class trigger vocabulary to evil-trait/toxic-corpus judge-banding phrasing (#2221 silent pass)

Provenance

• workflow_fix_target: scripts/verify_plan.py
• surfaced-by: statistics critic round 1 on task #2221 (plan v5), 2026-08-10
• fingerprint: c53-trigger-vocab-evil-banding

Gap

Check 53 (check_judged_dv_api_refusal, scripts/verify_plan.py ~L10040) gates harm-class judged-DV plans on per-arm n_api_refusal accounting. Its trigger arm (b) regex is deliberately narrow: jailbreak | harmfulness | harm-(judge|rate|score) | adversarial-role-play | harmful-compliance.

Task #2221's plan v5 judge-bands REAL toxic deployment content (LMSYS/WildChat responses) by evil-trait severity and judges on-policy generations of models fine-tuned toward evil — exactly the api-refusal-exposed content class (the #1739 evil-OOD wave lost 34.1% of 44,310 draws to API refusals) — yet c53 stayed SILENT: the plan's phrasing ("evil", trait banding, severity bands) matches none of the trigger tokens. The Statistics & Measurement critic caught the missing rule-28 accounting manually (REVISE Must-Fix 1) and flagged the c53 vocabulary gap as a prose follow-up.

Fix direction

Extend the trigger arm (b) vocabulary to cover trait/corpus phrasings that carry the same exposure, while keeping the deliberately-narrow design (avoid firing on incidental mentions): e.g. \bevil\b within a judge/band/severity window, toxic(ity) + banding/judging vocabulary, and the evil behavior-class token already used by sibling checks. Add a fixture reproducing #2221 plan v5's phrasing (evil-severity banding of real chat-corpus responses + judged on-policy generations of evil-finetuned models) asserting c53 WARNs on it, plus a negative fixture (incidental sibling-quote mention) asserting it stays quiet. Update the rule text in .claude/rules/llm-judging.md rule 28 enforcement clause only if its vocabulary list is mirrored there.

Evidence

• #2221 events.jsonl statistics-critic round-1 verdict (2026-08-10).
• scripts/verify_plan.py L10048-10060 (trigger + satisfier regexes).
• #1739 evil-OOD refusal incident (34.1% of 44,310 draws; rescue impl scripts/issue1739_evilood_refusal_rejudge.py).