Which evaluation hooks make interactive evidence tasks reproducible?

Bind each run to a `task_version`, `snapshot_id`, `policy_bundle_version` and `seed`. The snapshot should define the review state at the start, including accessible records and active permissions. If the task introduces later events, schedule them deterministically relative to an action or simulation clock.

Capture a structured action trace with `run_id`, `sequence`, `tool_name`, `arguments`, `result_code` and `state_revision`. Store the final submitted package separately from the trace. This allows evaluators to distinguish a correct final answer reached through prohibited actions from one reached through permitted evidence gathering.

Evaluator hooks can check assertions after individual actions and at completion. An intermediate assertion might detect attachment of an inaccessible record; a final assertion might test whether an unresolved discrepancy remains visible. Use stable error categories so results can be compared across agent implementations.

For debugging, compare the starting and ending state and replay the recorded sequence against the same snapshot. Keep expected outcomes and assertions outside agent-visible tool responses. Reproducibility depends on controlling external changes as well as preserving the prompt, particularly when evidence becomes available during a run.