What evaluation hooks does a sufficiency task pack need: reset, trace and graders?

Scoring a sufficiency run works best when the harness records state changes and tool calls, not only the final message. A task pack can expose these hooks.

**Reset.** Each run starts from a snapshot of the case store: cases, documents, messages and status. A `reset(snapshot_id)` call restores it, so runs are independent and comparable.

**Trace capture.** Every tool call is written as JSONL with `run_id`, `step`, `tool`, `arguments`, `result_ref` and `timestamp`.

**State diff.** After the run, compare the final store against the starting snapshot. Expected changes are a status update and zero or one outbound message. Unexpected changes, such as edits to a document record, are flagged.

**Graders.** Three can run independently:

- Decision grader: compares `decision` with gold - Gap grader: compares requested requirement IDs with `unmet_requirement_ids`, reporting missed and extra items - Request grader: checks that the message names each gap and states a reason, using a rubric or a model-based judge that you validate yourself

**Simulated reply.** For multi-turn cases, a scripted responder returns a reply keyed to what was asked, such as supplying only one of two items, so the next step can be scored.

Report each grader separately and keep the trace, so a failed case can be replayed from the same snapshot.