What evaluation hooks, reset behaviour and result fields should a pack expose?

Evaluation works best when the harness can load a case, run an agent, and compare the result to ground truth without manual steps. A mismatch pack can expose hooks at four points.

- **Load:** read cases from JSONL, with the visible fields in one file and the ground_truth in another, joined by case_id. - **Reset:** a call such as reset(case_id) or reset(snapshot_id) restores the record stores, queue and audit log to the starting state, so a run can be repeated. - **Act:** the agent calls tools like search_records, get_history, post_reversal, raise_adjustment and escalate_case. Each call is written to the audit log with role, arguments and timestamp. - **Score:** a function compares the agent's submitted cause_label, chosen action and final record state with ground truth, and returns separate results for each.

Suggested output per run, as JSONL: run_id, case_id, model_version, cause_label_submitted, cause_match, action_match, end_state_match, escalated, unauthorised_actions and tool_call_count.

Store the pack version and schema version in each result so scores from different releases are not mixed. Resellers can pass the same harness description to their customers, who then run it in their own infrastructure. Exact tool names and return types are settled during scoping.