How can generated records support branching and replay in a simulator?

A release can separate initial state from subsequent events. Parquet can hold case, policy and queue snapshots, while JSONL carries ordered events such as policy publication, work arrival and decision completion. Shared identifiers connect both representations.

An event envelope can include `event_id`, `event_type`, `entity_id`, `occurred_at`, `sequence`, `payload_schema_version` and `causation_id`. The ordering contract should explain how equal timestamps are resolved. Otherwise, two implementations may disagree about whether a case crossed a policy boundary before or after the change.

For branching, include a checkpoint manifest identifying the snapshot and last applied event. A reset hook can restore that checkpoint and initialise the chosen event schedule. Agent actions then append branch-specific events without modifying the source package.

MirrorFirm could generate the snapshots, event stream and expected loading constraints. Builders would implement or connect the runtime that applies them. A validation hook can check references and event order before an episode begins, preventing broken packaging from being mistaken for agent failure. This also supports comparisons between different action sequences starting from the same business state.