diff --git a/CLAUDE.md b/CLAUDE.md
index f8fd993..bc38128 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -469,6 +469,20 @@ steps 3-5.
 Full discipline + worked examples in `docs/bench-maxing.md`. Headlines:
 
 - Bench before AND after every change (n=3, signal under 5pp is noise).
+- **When a lever's failure class is sub-noise-floor, fix the
+  instrument, not just the lever.** The curated n=3 audit_mode bench
+  can't resolve a class that is ≤3-5 of 75 q (four 2026-05-18
+  hypotheses died there — round-trip, entity-mask, disambiguation).
+  Mine **ground-truth-carrying** questions from corpus titles
+  (`bench/mine_questions.py`) and grade by **deterministic retrieval
+  recall@k** via `query --dry-run` (`bench/recall_at_k.py`): no LLM,
+  no verifier, no n=3 noise, no 5pp floor, scalable to the corpus.
+  Resolves a single lever to ±1 question (worked example: numeral-
+  fold recall 55→75%, `a3ac653`). Caveats: measures *retrieval
+  surfacing* — necessary-not-sufficient for STRICT; mined fixtures
+  are answerable-by-construction so they **complement, never
+  replace**, the curated adversarial set (the verifier-honesty/trap
+  gate). Conflating the two is itself a bench-maxing error.
 - Avoid negation in prompts (Hermes-3-8B inverts under attention).
 - Bench is the scoreboard; live fixtures are the gates.
 - Self-heal beats retry (preserve partial output, never fabricate).