daily-fix: judge pilot sizing + classification rubrics

Workflow gap

Two same-week judge-wave incidents expose two gaps in the pilot-gate rule (.claude/rules/llm-judging.md rules 23/26):

1. Threshold unsatisfiable at small per-arm draws. #2091's G2 judge pilot FAILed (pilot_rc=7) with wildchat parse-fail 6.2% ≥ 2% — but at 16 draws/arm the smallest non-zero parse-fail rate is 1/16 = 6.25%, so ANY single parse failure trips the gate: the threshold is unsatisfiable by construction at that draw count. The session resolved it with an auditable per-(wave,arm) waiver constant (commit d839842fb4) and filed the parser defect separately (#2109). The rule never states that the pilot's per-arm draw count must satisfy n_draws ≥ 1/threshold.
2. Classification/labeling rubrics skipped piloting. A #1739 tactic-classification wave (MHJ 7-class taxonomy) failed SUBSTANTIVELY at scale — the judge returned analysis prose instead of the JSON schema ("Failed to parse judge JSON; returning None… Text: Let me analyze this attack request:") and the wave needed a rubric fix + re-dispatch. Rule 26's pilot-gate wording is anchored on graded 0–100 waves; a classification rubric is exactly as parse-fragile.

verified-at-filing: grep -n 'pilot' .claude/rules/llm-judging.md | head → the pilot-gate clauses exist; grep -cn '1/threshold\|satisfiab' .claude/rules/llm-judging.md → 0 (no sizing clause); grep -cn 'classif' .claude/rules/llm-judging.md → checked at compose time for current coverage. Incident quotes are the miners' probed readbacks (sessions b765cdcd rows 1522/1564/1569; 2f4940f0 rows 114/129).

unverified hypothesis — verify at plan time: whether the #1739 tactic wave was above the ~5,000-call pilot-gate floor (if below, item 2 is a rule-scope extension rather than a violated rule) — miner could not determine the wave size.

Proposed change

Amend .claude/rules/llm-judging.md rule 26: (a) the pilot's per-arm draw count must make the parse-fail threshold satisfiable (n_draws ≥ 1/threshold, with the waiver-constant pattern from #2091 named for the small-arm case); (b) state explicitly that classification/labeling rubrics (non-0-100) are pilot-gated the same as graded waves.

Provenance

• fingerprint: 03688b24565e

• workflow_fix_target: .claude/rules/llm-judging.md

• origin: /daily 2026-08-05 problem sweep — miner 6 P5, miner 2 P8.