How do state-diff graders and environment resets work for these tasks?

A simulated environment for this workflow can expose hooks that a harness calls around each run.

**Reset**

- `reset(case_id, seed)` loads the starting state and returns a `state_hash`. - The same case and seed return the same hash, so results can be compared across agents and versions.

**During the run**

- Each tool call is logged with arguments, result and a post-call `state_hash`. - Optional `pre_hook` and `post_hook` callbacks let you inject faults such as a failed lookup or a delayed document.

**Grading**

- `snapshot()` returns all records as JSON. - `diff(expected_ref)` returns a list of `{record_type, record_id, field, expected, actual}`. - Rule checks run over the call log, for example `status_active` must be preceded by `preconditions_checked`. - A `side_effects` list records any notices, payments or access grants triggered.

**Output**

A result line per run can carry `case_id`, `seed`, `outcome_score`, `process_score`, `handover_score`, `gate_failures` and `tags`.

The specific hook names are a proposal to settle in a pilot. The environment holds generated, synthetic state, so also test your harness against deliberately broken runs to confirm the graders fail what they should.