Fragility features within low-bc bins (multi-judge label)

=== bc < 0.65 (n=40, halluc rate=0.725) ===
  fragility_score           halluc-AUC=0.561 [0.363,0.747] p=0.5480 
  adversarial_fragility     halluc-AUC=0.586 [0.392,0.774] p=0.4180 
  adaptive_fragility        halluc-AUC=0.514 [0.330,0.699] p=0.8710 
  impostor_fragility        halluc-AUC=0.455 [0.257,0.663] p=0.6640 
  counterfactual_fragility  halluc-AUC=0.646 [0.459,0.815] p=0.1230 
  paraphrase_fragility      halluc-AUC=0.643 [0.437,0.824] p=0.1660 
  vulnerability             halluc-AUC=0.273 [0.123,0.449] p=0.0120 ★
  dissociation_rate         halluc-AUC=0.370 [0.183,0.554] p=0.1650 
  verbalized_confidence     halluc-AUC=0.375 [0.198,0.566] p=0.1970 

=== bc < 0.7 (n=80, halluc rate=0.688) ===
  fragility_score           halluc-AUC=0.540 [0.409,0.677] p=0.5470 
  adversarial_fragility     halluc-AUC=0.545 [0.413,0.689] p=0.4980 
  adaptive_fragility        halluc-AUC=0.515 [0.386,0.651] p=0.8030 
  impostor_fragility        halluc-AUC=0.473 [0.336,0.617] p=0.7560 
  counterfactual_fragility  halluc-AUC=0.498 [0.363,0.635] p=0.9930 
  paraphrase_fragility      halluc-AUC=0.573 [0.430,0.705] p=0.3040 
  vulnerability             halluc-AUC=0.386 [0.267,0.506] p=0.0650 
  dissociation_rate         halluc-AUC=0.420 [0.301,0.542] p=0.1760 
  verbalized_confidence     halluc-AUC=0.319 [0.214,0.432] p=0.0020 ★

=== bc < 0.75 (n=134, halluc rate=0.634) ===
  fragility_score           halluc-AUC=0.507 [0.402,0.615] p=0.8830 
  adversarial_fragility     halluc-AUC=0.495 [0.391,0.603] p=0.9450 
  adaptive_fragility        halluc-AUC=0.509 [0.398,0.608] p=0.8390 
  impostor_fragility        halluc-AUC=0.531 [0.427,0.629] p=0.5570 
  counterfactual_fragility  halluc-AUC=0.451 [0.349,0.555] p=0.3600 
  paraphrase_fragility      halluc-AUC=0.556 [0.451,0.654] p=0.3010 
  vulnerability             halluc-AUC=0.399 [0.303,0.497] p=0.0440 ★
  dissociation_rate         halluc-AUC=0.450 [0.370,0.533] p=0.2470 
  verbalized_confidence     halluc-AUC=0.404 [0.307,0.503] p=0.0590 

=== combined bc vs fragility_score within bc<0.65 ===
  -bc alone:   AUC=0.690  p=0.0970
  frag alone:  AUC=0.561  p=0.5480
  z(-bc)+z(frag)/2: AUC=0.649  p=0.1720

== Interpretation ==
  If any frag feat has AUC > 0.65 within low-bc bin: bin-conditional signal!
  Otherwise: bc alone suffices, frag adds nothing even in hardest bins
