Security
Security, tenancy and responsible-AI enforcement
Where each defence sits, which route is permitted, and which ingress stops at the boundary.
What it shows
The permitted route is a paved road. An authenticated caller's request passes the safety gates in a fixed order, then a server-side filter built from the caller's claims selects the sources of the studies that caller may see.
Retrieved text is treated as untrusted: a shield screens the spans for indirect prompt injection before they reach the draft. The drafting step runs under a least-privilege identity that can only read and retrieve.
Two routes stop at the compliance boundary. A question that tries to widen its own scope in the typed text has no effect, because the filter never reads the text. A preview or remote source that would take data outside the boundary is not on the production path.
This is a reference design for an agentic clinical-trials Q&A assistant. It is a design, not a description of a system built as shown.
Key decisions
- Gate order is a hard requirement, enforced by the graph's topology and not by convention.
- The tenancy filter comes from claims and is applied at the query layer, not as post-processing.
- The checkpoint store is a trust boundary: private, encrypted, and never loaded from an untrusted source.
- Production uses generally available services only; previews are for development and evaluation.