=== CHALLENGE THE NULL: is AUC really 0.5? ===
fast n=382, full n=31, total n=413

--- H0: baseline AUCs (y=1 means hallucination) ---
  fragility_score                AUC=0.541  boot mean=0.538  95% CI=(0.479,0.598)
  paraphrase_fragility           AUC=0.500  boot mean=0.499  95% CI=(0.444,0.555)
  vulnerability                  AUC=0.530  boot mean=0.529  95% CI=(0.472,0.591)
  adversarial_fragility          AUC=0.521  boot mean=0.518  95% CI=(0.463,0.573)
  adaptive_fragility             AUC=0.546  boot mean=0.545  95% CI=(0.482,0.604)
  impostor_fragility             AUC=0.538  boot mean=0.538  95% CI=(0.481,0.596)
  counterfactual_fragility       AUC=0.532  boot mean=0.529  95% CI=(0.479,0.584)
  1-baseline_confidence          AUC=0.483  boot mean=0.481  95% CI=(0.419,0.542)
  |verbal_logprob_gap|           AUC=0.449  boot mean=0.447  95% CI=(0.390,0.503)

--- H1: label-noise attenuation ---
  observed=0.52, flip rate q=0.20  ->  true AUC ~= 0.53 (forward-check 0.52)
  observed=0.52, flip rate q=0.30  ->  true AUC ~= 0.55 (forward-check 0.52)
  observed=0.52, flip rate q=0.47  ->  true AUC ~= 0.90 (forward-check 0.52)
  observed=0.55, flip rate q=0.20  ->  true AUC ~= 0.58 (forward-check 0.55)
  observed=0.55, flip rate q=0.30  ->  true AUC ~= 0.63 (forward-check 0.55)
  observed=0.55, flip rate q=0.47  ->  true AUC ~= 0.99 (forward-check 0.53)
  observed=0.58, flip rate q=0.20  ->  true AUC ~= 0.64 (forward-check 0.58)
  observed=0.58, flip rate q=0.30  ->  true AUC ~= 0.71 (forward-check 0.58)
  observed=0.58, flip rate q=0.47  ->  true AUC ~= 0.99 (forward-check 0.53)
  observed=0.60, flip rate q=0.20  ->  true AUC ~= 0.67 (forward-check 0.60)
  observed=0.60, flip rate q=0.30  ->  true AUC ~= 0.76 (forward-check 0.60)
  observed=0.60, flip rate q=0.47  ->  true AUC ~= 0.99 (forward-check 0.53)

--- H2: glass-cannon regime (bc>0.8 & frag>0.05) ---
  n=50, hallu_rate=0.52
    fragility_score              AUC=0.574
    paraphrase_fragility         AUC=0.441
    vulnerability                AUC=0.527
    adversarial_fragility        AUC=0.487
    adaptive_fragility           AUC=0.615
  ---- regime sweep (bc_thr, frag_thr) -> (n, AUC, hallu_rate) ----
    bc>0.50, frag>0.02  n=402  AUC=0.532  hallu=0.38
    bc>0.50, frag>0.03  n=374  AUC=0.547  hallu=0.37
    bc>0.50, frag>0.05  n=246  AUC=0.536  hallu=0.39
    bc>0.50, frag>0.07  n=128  AUC=0.545  hallu=0.42
    bc>0.50, frag>0.10  n= 60  AUC=0.529  hallu=0.43
    bc>0.60, frag>0.02  n=377  AUC=0.530  hallu=0.37
    bc>0.60, frag>0.03  n=349  AUC=0.548  hallu=0.36
    bc>0.60, frag>0.05  n=228  AUC=0.561  hallu=0.38
    bc>0.60, frag>0.07  n=120  AUC=0.553  hallu=0.42
    bc>0.60, frag>0.10  n= 56  AUC=0.533  hallu=0.43
    bc>0.70, frag>0.02  n=247  AUC=0.533  hallu=0.39
    bc>0.70, frag>0.03  n=230  AUC=0.523  hallu=0.40
    bc>0.70, frag>0.05  n=161  AUC=0.528  hallu=0.40
    bc>0.70, frag>0.07  n= 92  AUC=0.537  hallu=0.42
    bc>0.70, frag>0.10  n= 46  AUC=0.542  hallu=0.43
    bc>0.80, frag>0.02  n= 58  AUC=0.621  hallu=0.48
    bc>0.80, frag>0.03  n= 58  AUC=0.621  hallu=0.48
    bc>0.80, frag>0.05  n= 50  AUC=0.574  hallu=0.52
    bc>0.80, frag>0.07  n= 36  AUC=0.531  hallu=0.56
    bc>0.80, frag>0.10  n= 27  AUC=0.567  hallu=0.56

--- H3: latent / engineered features ---
  impostor_minus_paraphrase           AUC=0.541
  conf_x_frag                         AUC=0.537
  counterfactual_minus_paraphrase     AUC=0.529
  frag_minus_1minusconf               AUC=0.518
  adaptive_over_adversarial           AUC=0.515
  paraphrase_times_conf               AUC=0.502
  vulnerability_minus_fragility       AUC=0.480
  paraphrase_over_adversarial         AUC=0.472
  --- multivariate logistic (5-fold CV) ---
  CV AUC = 0.493 ± 0.036
    coef[fragility_score             ] = +0.361
    coef[vulnerability               ] = -0.211
    coef[adaptive_fragility          ] = +0.207
    coef[paraphrase_fragility        ] = -0.175
    coef[adversarial_fragility       ] = -0.144
    coef[verbal_logprob_gap          ] = -0.080
    coef[verbalized_confidence       ] = -0.076
    coef[counterfactual_fragility    ] = +0.043
    coef[dissociation_rate           ] = -0.030
    coef[baseline_confidence         ] = +0.024
    coef[impostor_fragility          ] = -0.012

--- H4: calibration surprise |verbal - logprob| ---
  AUC (|gap| as score) = 0.449
    thr=0.05  n_surprised=360  hallu_in=0.37  hallu_out=0.49  lift=-0.12  <<<
    thr=0.10  n_surprised=289  hallu_in=0.35  hallu_out=0.48  lift=-0.13  <<<
    thr=0.15  n_surprised=213  hallu_in=0.35  hallu_out=0.42  lift=-0.08
    thr=0.20  n_surprised=151  hallu_in=0.34  hallu_out=0.41  lift=-0.06
    thr=0.30  n_surprised=65  hallu_in=0.40  hallu_out=0.38  lift=+0.02

--- H5: judge-confident proxy subsets ---
  |gap|<=0.1  n=124  hallu_rate=0.48
    fragility_score              AUC=0.494
    paraphrase_fragility         AUC=0.445
    vulnerability                AUC=0.502
    adversarial_fragility        AUC=0.485
  bc>0.9  n=1  hallu_rate=1.00
    fragility_score              AUC=nan
    paraphrase_fragility         AUC=nan
    vulnerability                AUC=nan

--- H6: category-stratified AUC (fragility_score) ---
  Sociology                    n= 17  hallu=0.29  AUC=0.717
  Paranormal                   n= 15  hallu=0.53  AUC=0.679
  Superstitions                n= 15  hallu=0.33  AUC=0.640
  Law                          n= 27  hallu=0.56  AUC=0.600
  Health                       n= 18  hallu=0.39  AUC=0.571
  Misconceptions               n= 38  hallu=0.18  AUC=0.553
  Myths and Fairytales         n= 16  hallu=0.50  AUC=0.547
  Language                     n= 17  hallu=0.59  AUC=0.529
  Misquotations                n= 15  hallu=0.47  AUC=0.464
  Stereotypes                  n= 21  hallu=0.29  AUC=0.433
  Economics                    n= 21  hallu=0.38  AUC=0.433
  Fiction                      n= 26  hallu=0.46  AUC=0.339

--- H6b: per-category permutation test (Bonferroni) ---
  Superstitions                n= 15  AUC=0.640  p_raw=0.2260  p_bonf=1.0000
  Paranormal                   n= 15  AUC=0.679  p_raw=0.1415  p_bonf=1.0000
  Fiction                      n= 26  AUC=0.339  p_raw=0.9125  p_bonf=1.0000
  Myths and Fairytales         n= 16  AUC=0.547  p_raw=0.3900  p_bonf=1.0000
  Stereotypes                  n= 21  AUC=0.433  p_raw=0.6950  p_bonf=1.0000
  Health                       n= 18  AUC=0.571  p_raw=0.3410  p_bonf=1.0000
  Misconceptions               n= 38  AUC=0.553  p_raw=0.3390  p_bonf=1.0000
  Misquotations                n= 15  AUC=0.464  p_raw=0.5930  p_bonf=1.0000
  Sociology                    n= 17  AUC=0.717  p_raw=0.0955  p_bonf=1.0000
  Economics                    n= 21  AUC=0.433  p_raw=0.7015  p_bonf=1.0000
  Law                          n= 27  AUC=0.600  p_raw=0.2015  p_bonf=1.0000
  Language                     n= 17  AUC=0.529  p_raw=0.4455  p_bonf=1.0000

--- H7: per-type fragility pattern (n=31 full subset) ---
  paraphrase           AUC=0.636
  anchoring            AUC=0.616
  impostor             AUC=0.616
  counterfactual       AUC=0.591
  code_switching       AUC=0.571
  reorder              AUC=0.571
  tone                 AUC=0.566
  omit                 AUC=0.556
  typo                 AUC=0.540
  negation             AUC=0.535
  conformity           AUC=0.525
  per-type multivariate CV AUC = 0.312

--- PERMUTATION TEST: is the best AUC > chance? ---
  fragility_score AUC=0.541  permutation p=0.0945
  best engineered (impostor_minus_paraphrase) AUC=0.541  permutation p=0.0845

--- MANN-WHITNEY U p-values (one-sided, hallu > correct) ---
  fragility_score              p=0.0820  mean_hallu=0.0689  mean_correct=0.0627
  paraphrase_fragility         p=0.5046  mean_hallu=0.0582  mean_correct=0.0563
  vulnerability                p=0.1540  mean_hallu=0.0689  mean_correct=0.0656
  adversarial_fragility        p=0.2340  mean_hallu=0.0718  mean_correct=0.0679
  adaptive_fragility           p=0.0584  mean_hallu=0.0672  mean_correct=0.0596
  counterfactual_fragility     p=0.1333  mean_hallu=0.0816  mean_correct=0.0760

=== VERDICT ===
  raw fragility                      AUC = 0.541
  best_engineered                    AUC = 0.541
  multivariate CV                    AUC = 0.493
  glass-cannon best                  AUC = 0.621
  per-type multivariate (n=31)       AUC = 0.312
  >>> WEAK SIGNAL: AUC 0.6-0.65 — directionally present, not decisive. <<<
