Biotech RAG assistant: controls before the model
- Problem
- A document assistant will happily answer from a retired SOP if it is the closest match.
- What I did
- I built the trust layer first: a status gate before ranking, a refusal when nothing approved supports the answer, exact-span citations, and tests that remove each control to show it matters.
- Evidence
- Live demo, public code and tests, and the decision record.
- Limits
- Synthetic 10-document corpus. Deterministic: no language model writes answers yet. Not validated, not in production.