Healthcare

A clinician asks a question and pastes the chart to save time. The record number, the insurer and the address travel with it. This starter treats that as the normal case rather than the exception. It masks health detail before the model sees it and keeps the assistant out of diagnosis and prescribing.

This is the general health starter, written for a team that works across more than one country. It assumes clinical content will arrive in prompts, and that the safest handling is early masking. Reach for it when a deployment spans several regions and no single national rule governs the workflow.

Injection screening opens at 0.75, and a prompt that reaches the signal is refused. Redaction runs in health mode and masks record numbers, patient identifiers, insurance membership numbers, phone patterns, health card numbers and street addresses. A second pass masks diagnosis and procedure codes, treatment plans, laboratory results, medication schedules and study identifiers. Both passes mask rather than refuse, so a clinician can keep working.

Clinical language limits work on the output. The assistant cannot state a diagnosis, tell a person what they may have, prescribe or change a dose. Two disclaimers are required, pointing the reader to a provider and away from treating the answer as medical advice. A response has to clear a combined 0.80 with accuracy at 0.90 before it is delivered. Citation checks record grounding at a confidence of 0.85 rather than blocking the response.

Interactions are captured in redacted form for clinical audit. The audit marker only shows that an evaluation ran, so storage and retention stay with your own systems. This starter names no jurisdiction, so choose the United States health starter for protected health information rules, or the European one for special category data. Device classification, clinical validation and lawful basis remain decisions for your clinical and legal owners.