Guideline-as-Oracle: Rule-Derived Labels Teach a Triage Agent When to Escalate

Can a published clinical guideline replace per-dialogue expert labels when training an interactive triage agent? We study ophthalmic telephone triage: a guideline, compiled by hand into a 70-row rule table, labels every synthetic training dialogue with the most severe tier among the rows its caller profile cites, so no dialogue is labeled by a person. On a 201-case development reference adjudicated by the authors, fine-tuning a 9B agent raises emergent recall in simulated calls from 9.5% to 69.0% and halves under-triage, and this escalation gain holds in every evaluated setting (not fully crossed): five training runs, two software environments, two caller simulators, and five reference-robustness views, including a clinician's review of the flagged cases. The rule-derived labels carry the escalation gain: controls trained on the same dialogues with the final tier words, or entire final messages, permuted raise the same topics but call 0 and 2 of 42 emergent cases emergent, while the first control's next-step recommendations, whose training text was not permuted, still track the rule-derived tier; given the complete case, the untuned model still calls only 7 of 42 emergent cases emergent. Training shifts the balance between under- and over-triage, by amounts that vary across runs of one configuration. Agreement rises from 61.7% to 74.1% in the pre-specified comparison, but over-triage ranges from 26 to 54 calls across runs, so the agreement difference ranges from +12.4 to −5.5 points. Every run costs less than its base model once an under-triaged call counts more than 1.41 over-triaged ones, and significantly less in all five once it counts three. On 40 external vignettes the trained model escalates more but does not agree more. The main results measure agreement with a development reference, not clinical validity.