Handling the model being wrong (clinical-adjacent guardrails)

In a health-adjacent product, a wrong answer is a real harm.

v1 guardrails: 1. ground strictly in retrieved context and abstain when coverage/confidence is low; 2. always cite the source so the user (and a clinician) can verify; 3. show the retrieved evidence ("why this answer"); 4. never invent or change a variant classification — report the recorded ACMG/ClinVar value; 5. add a persistent "informational only, not a diagnosis; consult your clinician/genetic counselor" boundary; 6. log every generation for audit; 7. red-team for prompt injection and hallucinated significance. The product stance: assistive and explanatory, not diagnostic.