Cross-dataset replication audit — TriviaQA n=200
bc-only CV AUC: 0.745 ± 0.038  (baseline)

feature                   CV(bc+feat)       Δ         z
fragility_score           0.738±0.039    -0.007    -0.13 
adversarial_fragility     0.739±0.038    -0.006    -0.12 
adaptive_fragility        0.749±0.038    +0.004    +0.07 
impostor_fragility        0.753±0.037    +0.008    +0.14 
counterfactual_fragility  0.739±0.039    -0.005    -0.10 
paraphrase_fragility      0.742±0.038    -0.002    -0.05 
vulnerability             0.743±0.039    -0.002    -0.03 
dissociation_rate         0.734±0.039    -0.011    -0.20 

== Regex / length features (pivot candidate) ==
bc + is_hedge             0.741±0.038    -0.004    -0.07 
bc + answer_len           0.738±0.038    -0.007    -0.12 
bc + regex + len        0.734±0.038    -0.010    -0.19 

== Conclusion check ==
If TriviaQA shows significant (|z|>2) features where TruthfulQA did
not, the dataset choice drives the finding, not the feature semantics.
