We decompose physician disagreement in the HealthBench medical AI evaluation dataset to understand where variance resides and what observable features can explain it. Rubric identity accounts for 15.8% of met/not-met label variance but only 3.6-6.9% of disagreement variance; physician identity accounts for just 2.4%. The dominant 81.8% case-level residual is not reduced by HealthBench's metadata labels (z = -0.22, p = 0.83), normative rubric language (pseudo R^2 = 1.2%), medical specialty (0/300 Tukey pairs significant), surface-feature triage (AUC = 0.58), or embeddings (AUC = 0.485). Disagreement follows an inverted-U with completion quality (AUC = 0.689), confirming physicians agree on clearly good or bad outputs but split on borderline cases. Physician-validated uncertainty categories reveal that reducible uncertainty (missing context, ambiguous phrasing) more than doubles disagreement odds (OR = 2.55, p < 10^(-24)), while irreducible uncertainty (genuine medical ambiguity) has no effect (OR = 1.01, p = 0.90), though even the former explains only ~3% of total variance. The agreement ceiling in medical AI evaluation is thus largely structural, but the reducible/irreducible dissociation suggests that closing information gaps in evaluation scenarios could lower disagreement where inherent clinical ambiguity does not, pointing toward actionable evaluation design improvements.

翻译：暂无翻译