Δ(A4 − A1) bootstrap significance test
N_boot=2000  seed=42  target: does perturbation feature set improve AUC?

variant                                 n    pos      A4      A1   Δ_point                  95% CI    p_2s
TriviaQA (objective)                  200  0.280  0.7457  0.7517   -0.0061  [-0.0634, +0.0429]   0.811
NQ-Open (objective)                    75  0.547  0.4240  0.4907   -0.0667  [-0.1835, +0.0502]   0.276
TruthfulQA single_judge               200  0.635  0.6915  0.7183   -0.0268  [-0.0852, +0.0347]   0.346
TruthfulQA majority_3_judges          200  0.560  0.6607  0.6417   +0.0190  [-0.0421, +0.0774]   0.547
TruthfulQA unanimous                  140  0.571  0.6746  0.6965   -0.0219  [-0.0927, +0.0505]   0.552

== Significance rules ==
  p < 0.05 AND CI excludes 0  => perturbations have a genuine effect (positive or negative).
  p >= 0.05                  => cannot distinguish Δ from zero → no evidence of effect.

== Calibration reminder ==
  Bootstrap CIs overstate precision when folds are correlated; still they flag effects
  that are swamped by sampling noise.
