Change your policy. See exactly what changes.
Because every evaluation is captured as a canonical, hashable input, every evaluation can be re-run. Replay turns policy authoring into a change-controlled discipline instead of a leap of faith.
Regression testing for governance
Draft a new policy version. Point the replay engine at the last thirty days of evaluations. See — request by request — which verdicts flip, which gates now fire differently, and which harm bands shift. Ship the policy when the diff matches your intent. Not before.
Parity, not approximation
Replay uses the same engine, same gates, and same judgment path as live evaluation. There is no separate "simulation" code path to drift out of sync with production. What replay says will happen is what will happen.
Fixture library included
Every tenant ships with a starter fixture library covering the common request shapes — identity changes, external communications, infrastructure mutations, financial artifacts, low-consequence reads — so you can validate a new policy against known-good behavior on day one, before it touches a real request.
Parity-tested against the reference engine
The pipeline is parity-tested against a reference implementation across every combination of the shipped policy + request fixtures. When we say "deterministic," we mean the parity harness passes 594 out of 594 snapshots.