Premise doubt: does bc-chance on TruthfulQA come from label noise?
If so, bc should work on consensus subset (flex == llm).

=== Cerebras-Llama-3.1-8B TruthfulQA (n=412) ===
flex agree-rate with llm: 0.551  (both-correct=114, both-incorrect=113, disagree=185)
  consensus subset (n=227): bc-AUC=0.450 [0.374,0.525] p=0.187  Cohen d=+0.25
  disagree subset (n=185) flex-label:  AUC=0.673 p=0.001
  disagree subset (n=185) llm-label:   AUC=0.327 p=0.001

== Interpretation ==
  If bc-AUC on consensus subset is near 0.5 AND |d| < 0.2:
    → label noise is NOT the cause. bc is truly orthogonal.
  If bc-AUC on consensus subset > 0.6 (CI exclusion 0.5):
    → label noise was the cause — bc works when labels are reliable.
