bc ↔ surface form correlations

  bc vs answer_len   pearson r=+0.029 p=0.559   spearman ρ=-0.073 p=0.137
  bc vs ngram1       pearson r=+0.040 p=0.418   spearman ρ=+0.025 p=0.608
  bc vs ngram2       pearson r=+0.112 p=0.023   spearman ρ=+0.070 p=0.157
  bc vs ngram3       pearson r=+0.115 p=0.019   spearman ρ=+0.052 p=0.291

AUC of surface features against labels
  answer_len   flex AUC=0.580   llm AUC=0.553
  ngram1       flex AUC=0.695   llm AUC=0.611
  ngram2       flex AUC=0.743   llm AUC=0.634
  ngram3       flex AUC=0.706   llm AUC=0.629

== Disagreement subset n=185 (key from prior audit) ==
  bc           flex AUC=0.327   llm AUC=0.673
  answer_len   flex AUC=0.509   llm AUC=0.491
  ngram1       flex AUC=0.650   llm AUC=0.350
  ngram2       flex AUC=0.654   llm AUC=0.346
  ngram3       flex AUC=0.613   llm AUC=0.387

== bc with surface form residualized ==
  label=flex   raw -bc AUC=0.517   surface-residualized -bc AUC=0.542
  label=llm    raw -bc AUC=0.423   surface-residualized -bc AUC=0.427

== Interpretation ==
  High |r|(bc, ngram) indicates bc tracks surface overlap with gold.
  If bc AUC drops markedly after residualization, bc signal was surface-form.
  If surface features alone match bc AUC on disagreement subset,
    then the disagreement signal IS surface-form based.
