What evaluation hooks should a sufficiency task pack expose to my test harness?

A harness needs to feed the agent a case, capture what it does and compare that to a key it cannot see.

Useful hooks in a task pack:

- **Case loader:** `load_case(case_id)` returns the agent-visible records only. - **Gold labels:** a separate file with `expected_outcome`, `expected_gap_ids[]` and `acceptable_substitutes`. - **Clock control:** `set_now(timestamp)` so age checks are repeatable. - **Tool trace:** every tool call recorded with arguments, result and step index. - **Request parser:** a function that extracts item references from the agent's outbound message and maps them to `gap_id` values. - **Reply simulator:** `next_reply(case_id, request_id)` returns a scripted full, partial or contradictory reply. - **Scorer:** a function returning outcome match, gap precision and recall, request coverage and a count of redundant requests.

Store results as JSONL with `run_id`, `agent_version`, `case_id` and each score, so runs can be joined and compared. Where free-text judgement is needed, such as request tone, state the rubric explicitly and sample for human review instead of trusting a model grader alone.

Scoring code and trace formats are scoped per pilot, so agree them against your own harness before the first release.