MedSafe-Dx separates safe triage from diagnosis

MedSafe-Dx tested 12 frontier models on 250 simulated DDxPlus cases and put GPT-5 Chat first at a 72.4% Triage Success Rate. The medRxiv benchmark paper says GPT-5.2 reached a 97.6% Safety Pass Rate but over-escalated 71% of non-urgent cases, showing why diagnostic accuracy alone cannot size a clinical deployment. Builders shipping clinical decision support should gate pilots on missed escalations, false reassurance, and alert burden—not exam scores—while Forus’s medication-access agents show the adjacent workflow where humans still own treatment decisions.