Research programme · August 2026
Larrey-1
Larrey-1 is a pre-training research programme. A model release will follow only if frozen evaluation shows a meaningful, reproducible gain over its prompted base model without unacceptable under-triage or blanket escalation.
01 · Motivation
General medical knowledge is not enough.
Clinical triage is a sequence of decisions made with incomplete information. The model must request the fact that changes care, recognise when further questioning is unsafe, and know when enough is known to route the patient.
Larrey-1 tests whether supervision recovered from doctor-authored pathways can improve those observable decisions. The work separates structured action from patient-facing prose so fluent language cannot hide an unsafe route.
02 · Action contract
Constrain the decision before generating the explanation.
{
"action_type": "ask | emergency | urgent | same_day | routine | self_care | handoff",
"question_code": "string | null",
"urgency": "immediate | urgent | same_day | routine | self_care",
"patient_response": "string | null",
"confidence": 0.0
}03 · Training objective
Improve the action, not the appearance of reasoning.
Training targets reviewed actions, short decision factors and known correction pairs. It does not treat generated free-form chains of thought as clinician-authored truth.
04 · Training plan
Build evidence one intervention at a time.
Each stage isolates one source of improvement before the next is introduced.
Prompted baseline
Measure what the unmodified base model can do with a clear triage contract.
Structured output
Isolate gains caused by reliable action formatting.
Action supervision
Train on reviewed ask, stop, route and handoff decisions.
Trajectory supervision
Train across complete and partial multi-turn pathways.
Preference or reward
Prefer useful questions and safe dispositions over known failure modes.
Calibration
Set confidence and clinician-handoff thresholds away from the final test.
05 · Evidence
Evidence first. Announcement second.
The eventual release will lead with TriageBench safety results, external evaluations, clinician review and concrete failure cases.
High-value question, timely stop and supported care route.
A decisive red flag is missed or escalation is delayed.
The model escalates after reviewed critical features are absent.
TriageBench safety and capability outcomes.
06 · System boundary
A model is not the clinical product.
Larrey-1 is the learned policy and language component. A future Larrey product may add versioned pathways, deterministic red-flag controls, country-specific routing, clinician review, monitoring and immutable decision history.
Benchmark performance does not establish patient benefit, workflow fit, regulatory compliance or deployment safety. Each requires separate evidence.