A Step-Level Reliability Gateway for Agentic Clinical Decision Support: Architecture and Edge Deployment

Large language model (LLM) agents now retrieve patient records, call risk models and draft alerts in clinical workflows, yet their safety checks usually inspect only the final output. We present a reliability gateway that checks each action a clinical agent proposes, before it runs, and allows it, defers it to a clinician or blocks it. The decision fuses five run-time signals: calibration among similar patients, subgroup calibration gaps, group-conditional conformal sets, drift state and alert timing, plus a check that alerts quote tool outputs correctly. We generated agent trajectories with a local 7-billion-parameter model on emergency department (MIMIC-IV-ED) and intensive care (eICU-CRD) cohorts, injected six fault types at known steps, froze thresholds on a tuning split and replayed held-out trajectories. In the emergency setting, the gateway stopped 62% of faulty trajectories at the fault step, against 31% for an output-only check, with false intervention on 7.0% of clean trajectories. On held-out intensive care hospitals, its drift signal deferred every request. On a Raspberry Pi 5 the 99th-percentile latency per gated action was 25.8 ms, almost all of it from the nearest-neighbour signal, which added no measurable detection in the emergency setting. Step-level gating is feasible on embedded hardware; which signals are worth their cost depends on the deployment.