Untreated illness
Relapse, withdrawal, mania, psychosis, suicide, substance recurrence, hospitalization, functioning, nutrition, sleep, and prenatal-care engagement.
Methodology
Traceable evidence before fluent prose.
The system is designed around a versioned evidence release and a constrained clinical reasoning contract. The model may synthesize retrieved content; it may not silently create the evidence base.
Core reasoning frame
Four risks belong in every decision.
Medication exposure
Maternal toxicity, obstetric outcomes, pregnancy loss, congenital outcomes, fetal growth, neonatal effects, and long-term development.
Alternatives
Prior response, comparative effectiveness, reproductive evidence, access, speed, burden, and fit with the patient’s informed priorities.
Treatment change
Withdrawal, relapse, switch failure, adverse reactions, pharmacokinetic change, cross-exposure, and new monitoring requirements.
Decision taxonomy
Five positions stay separate.
Each medication/formulation record carries distinct clinical-position fields. A reassuring continuation position cannot be copied into initiation or switching without evidence and context.
- 1Continuation
- 2Initiation
- 3Switching
- 4Unplanned exposure
- 5Delivery, postpartum & lactation
Retrieval and synthesis
One request, three control layers.
Deterministic retrieval
Medication names, classes, diagnoses, stage, position, outcomes, and modifiers score the local evidence index. The highest-relevance records and claims form the evidence packet.
Structured synthesis
The OpenAI Responses API receives only the case object and packet. JSON Schema constrains the brief’s sections, certainty labels, citations, and verification tasks.
Post-generation validation
The server rejects malformed output and any claim or source citation that was not present in the retrieved packet. Emergency flags take a non-generative path.
Release status
What “v0.95” means.
The release provides a broad, auditable knowledge substrate and passed package, formula, and visual QA. It is not yet a commissioned clinical product.
Complete in artifact286 records · 206 claims · 1,269 links · 147 evidence units
SurveillanceSame-day official-source check; no clinical amendment identified
Open gatesLicensed searches, dual review, formal RoB/GRADE/AGREE II, label reconciliation, five external sign-offs
Limitations
The interface does not upgrade the evidence.
- Currency: Guidance is bounded by the deployed release cutoff. A newer model does not make an older evidence packet current.
- Coverage: A record may be present while important outcome, formulation, dose, or jurisdiction evidence remains sparse.
- Causality: Observational associations may reflect confounding by indication, severity, co-exposure, detection, and live-birth bias.
- Individualization: The structured intake cannot reproduce a complete psychiatric, obstetric, medical, social, or preference assessment.
- Output: Structured generation reduces formatting failure; it does not prove clinical correctness.
- Scope: This is decision support for licensed clinicians—not patient advice, autonomous prescribing, or emergency triage.