fix(m3-security-balanced-recalibration): stratified recalibration with safe-label audit, balanced sample, and hygiene
- SAFE-LABEL AUDIT: Fixed 3 mislabeled near-miss items: sec-nearmiss-29 (div-by-zero before nitems==0 guard now fixed), sec-nearmiss-33 (prefix-check bypass via sibling dirs now fixed), sec-nearmiss-34 (same prefix bypass now fixed with boundary check) - STRATIFIED RECALIBRATION: Built balanced calibration sample (10 targets + 8 near-misses + 7 distractors = 25 items), K_TRIALS=5, CONC=4, batch-size=2, solver claude-sonnet-4-6. Aggregate baseline 58.4% — in [0.40,0.70] band → in_band=true - RE-FREEZE: Manifest item_counts verified to EXACTLY match JSONL (25 target, 10 near_miss, 35 distractor, 0 ambiguous = 70 total) - HYGIENE: Untracked calib-sample scratch files via git rm --cached
T
Travis Boudreaux committed
a39e59f3ff0d1f5d635ac7ad8a5a00ed28ee19a2
Parent: ec07543