PRD, agent description, claim, or complaint in: relevant regulations and applicable obligations out, scoped to that agent. Runs at claims-volume, not just at design time.

Each obligation maps to a concrete control. Two agents at the same company can carry different obligation sets, so mapping is never generic.

Each mapped obligation compiles into a test case: an eval suite built automatically as obligations are discovered, not hand-written after the fact.

Carver runs the eval suite against agent behavior continuously, surfacing gaps before incidents. This tests what an agent does, separate from regulatory intelligence, which reports what a regulator published. Adversarial red-teaming is in active development.

Carver reads existing agent logs, tags interactions for legality, and alerts on non-compliant behavior as it happens. No new monitoring stack.

Flagged interactions route to the right team with tracked status: open, in review, resolved.

When a regulation changes, agents mapped against the old version are stale. Carver flags exactly which agents need re-mapping.