How would you audit AI decisions and actions without leaking sensitive data?

The schema supports reconstruction across asynchronous runs, provider calls, retrieval, agent delegation, approval, and external effects. Sampling never drops required high-risk actions or security failures, while privacy controls prevent broad analyst access to raw content. Integrity checks, clock discipline, export controls, and legal holds preserve credible evidence. Periodic tests start from an incident question and verify that the audit trail can answer it without resorting to unrestricted database access.