Evidence Is Not a Vote: Categorical Stability and Confidence Sensitivity Under Conflicting Clinical Evidence

Clinical language models may encounter evidence sources that differ in methodological strength and direction, yet it remains unclear how repeated lower-tier contradiction affects model behavior when an evidence hierarchy is explicitly supplied. We developed a controlled hierarchy-maintenance benchmark that holds an explicitly designated higher-tier systematic-review anchor fixed while adding one, two, or four synthetic lower-tier items that contradict it, together with a matched aligned condition and an anchor-last condition. Categorical performance was near ceiling under the anchor-only condition in an independently frozen 260-case Gemini-3.6-flash replication cohort (259/260; 99.62%). Against this high-performing baseline, contrary to our prespecified hypothesis, no dose-dependent increase in Evidence Dilution Rate (EDR) was observed: EDR was 2/259 (0.77%), 1/259 (0.39%), and 1/259 (0.39%) after one, two, and four conflicting items. The four events arose from two cases and were all shifts to insufficient; no opposite-direction reversals were observed. Model-reported confidence nevertheless responded systematically to evidence conflict. The prespecified anchor-to-four-conflict shift was −4.64 points on the 0–100 confidence scale (95% paired bootstrap CI, −5.13 to −4.16). In a post-hoc matched comparison, confidence was 6.51 points lower under four conflicting than four aligned lower-tier items (95% CI, −7.07 to −5.95), with lower confidence in 224/260 paired cases. A post-hoc GPT-OSS-120B sensitivity analysis showed the same qualitative pattern, with no EDR events through four conflicting items in 40 baseline-correct cases and a conflict-versus-aligned difference of −5.75 points (95% CI, −7.45 to −3.60). Thus, within an explicitly instructed evidence hierarchy, categorical judgments were highly stable across the tested conflict range while model-reported confidence remained sensitive to evidence polarity. These findings motivate evaluating behavioral responses to evidence conflict alongside categorical correctness, without implying autonomous evidence appraisal or calibrated uncertainty.