#46 replication on secondary datasets
Target: does 'perturbations hurt AUC' generalise beyond TruthfulQA n=412?

-- NQ-Open (n=75)  n=75  pos_rate=0.547  5-fold CV  seed=42 --
config                 n_feat                LogReg                    GB              CatBoost
A4_full                    54  0.4240 [0.288,0.560]  0.5029 [0.368,0.632]  0.4763 [0.343,0.611]
A1_no_pert                 18  0.4907 [0.352,0.628]  0.4656 [0.336,0.596]  0.4419 [0.315,0.574]
A2_pert_only               39  0.4096 [0.283,0.545]  0.4742 [0.341,0.610]  0.4448 [0.315,0.580]
A3_top3                     3  0.4491 [0.315,0.586]  0.4742 [0.341,0.604]  0.5065 [0.370,0.639]

A4 LogReg (full)          = 0.4240
A1 LogReg (no-pert)       = 0.4907
Δ (A4 - A1)              = -0.0667  <-- perturbations HURT
A3 best AUC (3 features)  = 0.5065
-- HaluEval (n=46) [EXCLUDED: all perturbation features = 0.0, see note]  n=46  pos_rate=0.543  5-fold CV  seed=42 --
config                 n_feat                LogReg                    GB              CatBoost
A4_full                    11  0.8095 [0.660,0.931]  0.5619 [0.375,0.735]  0.5981 [0.416,0.766]
A1_no_pert                 14  0.8076 [0.656,0.931]  0.5695 [0.388,0.741]  0.6057 [0.425,0.773]
A2_pert_only                0                  SKIP                  SKIP                  SKIP
A3_top3                     3  0.6571 [0.492,0.815]  0.4286 [0.263,0.603]  0.4324 [0.267,0.610]

A4 LogReg (full)          = 0.8095
A1 LogReg (no-pert)       = 0.8076
Δ (A4 - A1)              = +0.0019  <-- tie
A3 best AUC (3 features)  = 0.6571
-- TriviaQA (n=200) [no verbal_logprob_gap — A3 skipped]  n=200  pos_rate=0.280  5-fold CV  seed=42 --
config                 n_feat                LogReg                    GB              CatBoost
A4_full                    48  0.7457 [0.661,0.823]  0.7752 [0.701,0.845]  0.7467 [0.666,0.820]
A1_no_pert                  9  0.7517 [0.674,0.828]  0.7566 [0.667,0.836]  0.7753 [0.695,0.848]
A2_pert_only               39  0.6559 [0.573,0.738]  0.6715 [0.584,0.752]  0.6589 [0.570,0.742]
A3_top3                     0                  SKIP                  SKIP                  SKIP

A4 LogReg (full)          = 0.7457
A1 LogReg (no-pert)       = 0.7517
Δ (A4 - A1)              = -0.0061  <-- tie
A3 best AUC (3 features)  = n/a
