TriviaQA: bc ↔ surface form audit

  bc vs answer_len   pearson r=-0.377 p=0.000   spearman ρ=-0.339 p=0.000
  bc vs ngram1       pearson r=+0.375 p=0.000   spearman ρ=+0.407 p=0.000
  bc vs ngram2       pearson r=+0.202 p=0.004   spearman ρ=+0.213 p=0.002
  bc vs ngram3       pearson r=+0.043 p=0.548   spearman ρ=+0.047 p=0.507

feature AUC vs is_correct (TriviaQA)
  bc            AUC=0.748
  answer_len   AUC=0.411
  ngram1       AUC=0.946
  ngram2       AUC=0.708
  ngram3       AUC=0.537

raw bc AUC: 0.748
surface-residualized bc AUC: 0.612

== Interpretation ==
  TriviaQA flex label is basically n-gram overlap with gold.
  If bc-ngram r is high AND bc-resid AUC <<< raw bc AUC,
    TriviaQA bc success is a surface artifact.
  If bc-resid AUC ≈ raw bc AUC, bc carries independent signal.
