How do I wire scoring and reset into a condition-gated task?

Scoring a condition-gated task works best when the harness records state changes as they happen, not only the final answer.

**Hooks.** Add a pre-action hook on release_step and any other action that moves a case forward. It logs the status of every blocking condition at that moment. Add a post-action hook on status writes that stores the actor, the old status, the new status and the evidence_id cited.

**Run record.** Write one JSONL line per action: run_id, task_id, step_index, action, arguments, open_blocking_conditions and timestamp. You can then compute premature release by filtering for release_step lines where open_blocking_conditions is not empty.

**Grader inputs.** Compare the final state to the task's expected_end_state: each condition status, the downstream step state and any escalation raised. Compare reason_code values against an accepted set, because some cases allow more than one defensible reason.

**Reset.** Each task points to a start_state_ref, a snapshot of records and inbox contents. Reset restores the snapshot and a fixed seed for any arrival timing, so reruns and model comparisons start from identical conditions.

**Splits.** Keep a held-out set of decisions, not just held-out rows, so related conditions and evidence do not leak between training and test.

**Metrics.** Report premature release, over-blocking, wrong-link evidence, unsupported waiver and correct escalation as separate columns, so one score does not hide the failure type.