Which reset and evaluation hooks make customer change episodes reproducible?

Define an episode manifest containing `scenario_id`, `schema_version`, `policy_version`, `initial_snapshot_id` and `seed`. The seed should control documented randomness such as response selection or worker scheduling. It should not stand in for missing records of external effects.

A reset hook should restore mutable records, pending jobs, tool session state and the simulated clock. An evaluation hook can return named checks with `passed`, `evidence_refs` and an explanation. Record each tool call with its inputs, result, event sequence number and relevant state version.

MirrorFirm could generate manifests, initial snapshots and expected checks for implementation in a buyer's runtime. Builders should keep evaluator-only assertions outside the agent's tool responses. A replay mode can feed recorded external responses back into the same action trace, while a fresh run tests the agent under the original conditions.

Verify reset by running a known sequence, resetting and confirming that no earlier mutation remains observable. When a failure occurs, compare the replay with the fresh run. This helps locate nondeterminism or adapter defects before attributing the result to agent reasoning.