Reasoning Before Disposition: A Model-Agnostic Cannot-Miss Discipline for Quiet Emergencies and the Case for Deterministic Enforcement

Key Findings

• — Patient minimization did not degrade escalation when danger was overt: across 40 unambiguous emergencies in five framings (neutral, mild minimization, strong denial, third-party reassurance, benign distractor) sensitivity was at or above 97.5% in every arm and flat across framings.
• — Atypical presentation was the failure mode: across 76 subtle or atypical emergencies the base model failed to escalate 13.2% (95% CI 7.3 to 22.6) to the top acuity tier, cut to 3.9% (1.4 to 11.0) by a cannot-miss governance prompt (McNemar 7 to 0, p = 0.016, relative risk 0.30, number needed to treat 11).
• — The effect replicated in Spanish (27.8% to 11.1%), in a second harness, and across eight models from two families, where baseline under-triage ranged from 2.6% to 34.2% by model and elicitation mode.
• — Across more than 5,900 triage classifications no emergency was ever routed home; the failure mode was uniformly emergency-to-urgent, and the prompt added zero false emergency escalations on a 44-stem non-emergency panel.
• — The prompt left medical knowledge (MedQA 95.8% in both arms), adversarial robustness and medication-hazard detection unchanged, raised named-guideline citation from 65.6% to 96.9%, and lowered HealthBench by 3.6 points on one seed.
• — Forcing an immediate single-shot disposition raised baseline hazard by about ten percentage points in two of four Claude models.

Methodology Notes

Prompt-level ablation on Claude Opus 4.8 followed by replication across eight models (Claude Opus 4.8, Sonnet 5, Haiku 4.5, Fable 5; Gemini 3.1 Pro Preview, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite) under two elicitation modes; constructed vignette cohorts (40 overt, 76 subtle, 44 non-emergency stems), each subtle stem mapped to a published guideline naming a cannot-miss diagnosis. Ground truth was a blinded LLM-simulated three-physician panel rather than human adjudication. Posted to medRxiv 2026-09-08 under CC BY-NC; the authors are employees of Certuma, a triage-governance vendor, and the paper argues for deterministic enforcement without evaluating the vendor's product. Verified via the medRxiv landing page and the bioRxiv details API.

Sources