Regex hedge validity audit: TruthfulQA n=412
has_uncertainty rate=0.133   is_hedge rate=0.413

== label: flex(is_correct) ==
  bc only:              AUC=0.443 ± 0.029
  bc + has_unc               AUC=0.427 ± 0.028  Δ=-0.016 z=-0.41 
  bc + is_hedge              AUC=0.448 ± 0.030  Δ=+0.005 z=+0.12 
  bc + has_unc + is_hedge    AUC=0.426 ± 0.029  Δ=-0.018 z=-0.42 
  bc + regex + len           AUC=0.561 ± 0.030  Δ=+0.118 z=+2.81 ***

== label: llm_label ==
  bc only:              AUC=0.574 ± 0.029
  bc + has_unc               AUC=0.563 ± 0.028  Δ=-0.011 z=-0.28 
  bc + is_hedge              AUC=0.600 ± 0.028  Δ=+0.026 z=+0.65 
  bc + has_unc + is_hedge    AUC=0.593 ± 0.028  Δ=+0.019 z=+0.47 
  bc + regex + len           AUC=0.600 ± 0.028  Δ=+0.026 z=+0.64 

== Interpretation ==
*** (|z|>2) = regex feature adds statistically significant value over bc-only
If all non-sig, then regex hedges are NOT robust on TruthfulQA either
