We asked two leading AIs to triage the same four emergency patients, over and over. They matched the clinician's level 459 times in 464. None of the rest was wrong — and most of them came from the written rule.

Ask an AI the same question about the same patient again and again. Do you get the same answer every time? In emergency triage, a different answer means a different wait.

• Our clinical practitioner, who reviewed the cases: no variation, other than minor phrasing.
• Ours: the clear cases stay on one level; the borderline case splits between 2 and 3 on the same question; rewording the question moves it by a quarter or more; a written rule does not remove the split.

Four invented emergency department triage notes, reviewed by our clinical practitioner before any run, each scored on the 5-level Emergency Severity Index (ESI):

• A: an 82-year-old on apixaban who fell and hit her head, no loss of consciousness, normal vital signs. Borderline: 2 preferred, 3 defensible, and she gets the CT either way.
• A2: the same patient, unresponsive for about 30 seconds and vomited once. Level 2.
• B: a 22-year-old with a rolled ankle who needs one X-ray. Level 4.
• C: a 45-year-old after a head-on crash, in shock and short of oxygen. Level 1.

Two leading models, asked for the level and a one-sentence reason, each time in a fresh conversation, and in several wordings of the question.

• The clinician's level almost every time. The rest were acceptable by the clinician's review: a defensible call on the borderline patient, and over-triage. None was wrong.
• Every clear case got its level every time on both models. On the borderline patient, both models chose the more cautious level 2 almost every time; model B rarely called her a 3, citing the CT and the laceration repair.
• Rewording the question changed nothing: the intended level every time.

The rule — a rule about following the triage scale's own decision points in order — made no answer better. It produced most of the few departures from the clinician's level, all on model A, all over-triage that the clinician judged acceptable. On the patient who had been unresponsive, it sometimes named the high-risk decision point, which gives level 2, and wrote level 1 above it; one reply ended warranting ESI-2 under a heading of Level 1. It also counted a single ankle X-ray as two or more resources and called a level 4 a 3.

The practitioner's prediction held, and ours mostly did not. Where a written standard defines the answer, the same question gets the same answer, even on a patient experienced clinicians would debate. The little variation there was came from the fix, not the model: asking for the reasoning out loud produced levels the reasoning itself contradicted. None of it was wrong; all of it leaned cautious. Consistency is not where triage AI breaks. The next place to look is where no standard exists.

Four invented cases on one scale with written decision points, and two models used directly, without the chat app around them.