On-premise clinical AI agent grades its own uncertainty

An on-premise clinical AI agent reached 90% accuracy on a seven-disease diagnostic benchmark and can grade its own confidence well enough to defer roughly half of cases to a physician, according to a study in Nature Medicine. The system, designed to run entirely inside a hospital rather than call out to a cloud API, pairs a 'Physician Agent' with a simulated 'Patient Agent' and scores each answer across five independent runs of the same case.

On the MIMIC-IV-derived MIRA-v2 (551 cases, seven conditions), the open-weight Qwen-3.5 hit 90.0%, within 0.7 percentage points of a GPT-5.2 cloud baseline. On the four-condition CDM benchmark of 2,400 abdominal cases it reached 83.8%. Accuracy dropped to 72.22% on the harder external VivaBench, which spans 990 cases across ten specialties.

The reliability layer is the pitch. The team scored each answer with four signals: internal probability, linguistic certainty, concept density, and behavioral consistency across repeated runs. Consistency was most discriminative (AUC = 0.860, rising to 0.875 under stress testing with degraded information). Setting a diagnostic-consistency threshold of 0.90 handled 49.4% of cases at 98.9% accuracy, with only three autonomous errors among 272 retained cases.

'Safety, therefore, depends on reliable decision-time signals indicating when to trust, verify or defer,' the authors write. Physician review judged 81.8% of the 181 reviewed cases clinically valid for both agent and reference diagnoses, and flagged 4.4% as valid agent diagnoses that disagreed with the underlying EHR label.

The cost is compute. The five-run consistency analysis 'increased token use approximately five-fold relative to single-pass inference,' with substantially higher serial wall-clock latency. The paper was already circulating among two of the clinician-AI researchers we follow this week.