# Yuragi Abstention Analysis (LLM labels as ground truth)
N records: 412
Base accuracy (LLM judge): 0.3859

## Summary: AURC (lower=better), CAA = coverage at target accuracy (higher=better)
signal                           AURC  CAA@50  CAA@60  CAA@70  CAA@80  CAA@90  CAA@95
ideal                          0.2466   0.772   0.643   0.551   0.481   0.427   0.405
fragility_score                0.6032   0.024   0.000   0.000   0.000   0.000   0.000
adversarial_fragility          0.6001   0.078   0.015   0.002   0.002   0.002   0.002
adaptive_fragility             0.6041   0.000   0.000   0.000   0.000   0.000   0.000
vulnerability                  0.6498   0.000   0.000   0.000   0.000   0.000   0.000
dissociation_rate              0.6244   0.010   0.007   0.000   0.000   0.000   0.000
baseline_confidence            0.5618   0.141   0.085   0.000   0.000   0.000   0.000
random                         0.6141   0.000   0.000   0.000   0.000   0.000   0.000

## Selective accuracy at abstention rate alpha
signal                       a=0.0 a=0.1 a=0.2 a=0.3 a=0.4 a=0.5
ideal                        0.386 0.430 0.483 0.552 0.644 0.772
fragility_score              0.386 0.397 0.392 0.389 0.372 0.403
adversarial_fragility        0.386 0.384 0.377 0.368 0.385 0.393
adaptive_fragility           0.386 0.395 0.404 0.410 0.413 0.413
vulnerability                0.386 0.381 0.383 0.389 0.385 0.374
dissociation_rate            0.386 0.389 0.401 0.365 0.364 0.379
baseline_confidence          0.386 0.411 0.426 0.406 0.409 0.422
random                       0.386 0.386 0.386 0.386 0.386 0.386

## Headline
Best signal overall (by AURC): baseline_confidence
Best yuragi-native signal: adversarial_fragility
  Note: base accuracy 0.386 — targets >=60% are more informative than >=90% here
  CAA@60 best=0.085 (random=0.000, ideal=0.643)
  CAA@70 best=0.000 (random=0.000, ideal=0.551)
  CAA@90 best=0.000 (random=0.000, ideal=0.427)
  fragility_score: AURC_norm_improvement=0.030 (1=ideal, 0=random)
  adversarial_fragility: AURC_norm_improvement=0.038 (1=ideal, 0=random)
  adaptive_fragility: AURC_norm_improvement=0.027 (1=ideal, 0=random)
  vulnerability: AURC_norm_improvement=-0.097 (1=ideal, 0=random)
  dissociation_rate: AURC_norm_improvement=-0.028 (1=ideal, 0=random)
  baseline_confidence: AURC_norm_improvement=0.142 (1=ideal, 0=random)
