Trustworthy biomedical AI must select reliable actions when evidence is missing, stale, conflicting, or delayed, not merely provide accurate answers on static inputs.
We construct an episode‑level benchmark comprising 400 synthetic healthcare workflow episodes. Each episode is generated from 40 hand‑crafted semantic seeds covering four task classes, two surface parameterizations, and five controlled variants. Model outputs are categorized as ANSWER, ASK CLARIFICATION, ABSTAIN, or ESCALATE and scored under a pre‑declared 5:1 missed‑versus‑unnecessary escalation cost. Reference actions are single‑author study‑policy labels rather than clinician‑adjudicated ground truth.
Across five open‑weight configurations (1.7B‑9.7B), clean‑scenario decision accuracy reaches 0.97, yet in uncertainty‑prompted scenarios models miss 32%‑97% of required escalations. The largest 9.7B model’s accuracy drops from 0.97 to 0.78 under conflict, leaving a substantial gap.
Verbalized confidence is inconsistent: calibration error (ECE) ranges from 0.10 to 0.54, and failure detection AUROC spans 0.48‑0.77. A confidence‑only policy tuned on the development set yields no utility gain.
A transparent workflow‑aware controller improves 5:1 utility for every model, reducing missed escalations to 0.12‑0.24 while maintaining specificity of 0.83‑0.93. However, a 30‑concept lexical stress test drives each rule detector’s recall from 1.00 on trigger phrases to 0.00 on unseen paraphrases, limiting the controller’s claim to this benchmark.
The paper packages seed‑clustered statistics, ablations, cross‑task transfer results, raw outputs, and manifests for audit. The findings motivate explicit workflow‑state evaluation and demonstrate that seemingly effective routing controls require their own generalization tests.
Review