Which evaluation hooks make follow-up scores reproducible and auditable?

An evaluation harness can record events such as `observation_returned`, `decision_committed`, `action_attempted`, `action_applied` and `episode_closed`. Each event should include `episode_id`, `step_id`, `state_version` and the relevant record references. Store policy references as clause identifiers tied to a versioned policy document.

Ask the agent for a concise decision record containing its selected action, cited evidence and unresolved conditions. This is an auditable output, not a requirement to expose private internal reasoning. The evaluator can compare the record with the observations actually returned by tools.

Deterministic checks can verify invariants such as whether a contact was submitted while a hold was active. A separately calibrated language assessment can evaluate whether the message misstates the obligation. Keep those results in distinct fields, including `check_id`, `result`, `evidence_refs` and `evaluator_version`.

MirrorFirm could supply generated episodes with compatible reference traces and expected state constraints. Buyers should version their scoring code and retain tool outputs needed for replay. If a score changes after a rubric update, the trace should allow reviewers to identify the changed judgment without rerunning the agent against a different environment.