Label source sensitivity audit: BH-FDR survivor claims
Joined dataset n=412  is_correct↔llm_label agreement=55.1%

cat              feature                   flex AUC [CI]           llm AUC [CI]            flip?
Conspiracies     fragility_score           0.882 [0.667,1.000]  0.536 [0.262,0.800]  NO
Conspiracies     vulnerability             0.853 [0.647,1.000]  0.560 [0.283,0.833]  NO
Conspiracies     verbalized_confidence     0.176 [0.059,0.282]  0.655 [0.375,0.900]  YES
Law              dissociation_rate         0.548 [0.379,0.719]  0.620 [0.540,0.708]  NO
Law              verbalized_confidence     0.476 [0.280,0.664]  0.840 [0.680,0.960]  YES
Economics        impostor_fragility        0.577 [0.308,0.824]  0.850 [0.650,1.000]  NO
Economics        counterfactual_fragility  0.481 [0.218,0.741]  1.000 [1.000,1.000]  YES
Economics        dissociation_rate         0.591 [0.355,0.816]  0.700 [0.594,0.800]  NO

== How to read ==
flip=YES means AUC crosses 0.5 when label source changes → claim is label-dependent.
Robust claim: both AUC > 0.5 with overlapping CIs.
