TruthfulQA n=200 (Cerebras × multi-judge overlap)
flex vs majority agreement: 0.530
  both correct: n=58  both incorrect: n=48
  flex+/maj- (flex false positive): n=64
  flex-/maj+ (flex false negative): n=30

=== bc mean per 2×2 cell ===
  both correct (n=58): bc μ=0.728
  both incorrect (n=48): bc μ=0.700
  flex+/maj- (n=64): bc μ=0.695
  flex-/maj+ (n=30): bc μ=0.751

=== disagreement subset (flex ≠ majority, n=94) ===
  flex          AUC=0.302 [0.183,0.432] p=0.0040
  majority-3    AUC=0.698 [0.568,0.817] p=0.0040

=== among flex-positive items (n=122) ===
  bc AUC vs majority-3: (0.6333512931034483, 0.5335858594211719, 0.736472227097227, 0.013)
=== among flex-negative items (n=78) ===
  bc AUC vs majority-3: (0.65, 0.5254353281969509, 0.7781425202156333, 0.026)

== Interpretation ==
  If flex+/maj- cell has LOW bc: bc detects flex false positives (useful!)
  If flex-/maj+ cell has HIGH bc: bc rescues flex false negatives (useful!)
  If both cells have middling bc: bc doesn't help disambiguate flex errors.
