Dangerous halluc subset characterization (n=200)

Top 25% bc (bc>=0.770): n=50, halluc n=23, correct n=27
  dangerous rate (halluc given high-bc) = 0.460

=== top-bc hallucination frequency by category ===
  category                   total  top-bc  danger  safe  danger/top-bc
  Law                         17      4       4       0    1.00
  Fiction                     11      0       0       0    -
  Economics                   11      4       3       1    0.75
  Language                    11      0       0       0    -
  Conspiracies                10      4       0       4    0.00
  Stereotypes                 10      1       1       0    1.00
  Myths and Fairytales         9      1       0       1    0.00
  Distraction                  9      3       2       1    0.67
  Health                       9      3       0       3    0.00
  Indexical Error: Location    8      5       0       5    0.00
  Misconceptions               8      1       0       1    0.00
  Sociology                    8      1       0       1    0.00
  Superstitions                7      1       1       0    1.00
  Logical Falsehood            7      1       1       0    1.00
  Nutrition                    7      5       3       2    0.60
  Proverbs                     6      0       0       0    -
  Paranormal                   6      1       1       0    1.00
  Psychology                   6      0       0       0    -
  Advertising                  5      3       2       1    0.67

=== can any feature flag dangerous within top-bc subset? ===
  fragility_score           halluc-AUC within top-bc subset = 0.475 
  adversarial_fragility     halluc-AUC within top-bc subset = 0.457 
  adaptive_fragility        halluc-AUC within top-bc subset = 0.461 
  impostor_fragility        halluc-AUC within top-bc subset = 0.499 
  counterfactual_fragility  halluc-AUC within top-bc subset = 0.506 
  paraphrase_fragility      halluc-AUC within top-bc subset = 0.539 
  vulnerability             halluc-AUC within top-bc subset = 0.456 
  dissociation_rate         halluc-AUC within top-bc subset = 0.412 
  verbalized_confidence     halluc-AUC within top-bc subset = 0.492 

=== within top-bc subset (n=50) ===
  correct bc μ = 0.814 σ=0.031
  halluc  bc μ = 0.812 σ=0.031

== Interpretation ==
  If certain categories dominate danger/top-bc, adversarial types confirmed
  If |AUC-0.5|>0.15 for some feature, bin-conditional signal exists
  Otherwise: dangerous halluc within high bc is unflaggable (scary)
