How should reset, state diffs and grader hooks work for approval tasks?

Approval tasks are stateful, so reproducible resets and clear grading hooks matter more than usual.

**Reset**

- Load a snapshot per task: tables for people, roles, rules, delegations, requests and the decision log, plus a clock value. - Make the snapshot immutable and identified by `snapshot_id` and a content hash. Each episode starts from a fresh copy. - Seed any simulated actors, such as an approver who replies after a delay, so replays match.

**Observation and actions**

- Return records the actor is entitled to see. Hide validity labels and the answer key. - Log every tool call with arguments, simulated time and the resulting state version.

**State diff**

- At episode end, compute a diff between the starting snapshot and final state, limited to `request`, `decision_log`, `assignment` and downstream tables such as `payment_hold`. - Express it as added, changed and removed rows keyed by primary key, so graders compare structure and not text.

**Grader hooks**

- `on_step(state, action)` runs the policy checker and appends a `valid` flag to a private log. - `on_end(diff, trace)` returns outcome, process and penalty scores as separate fields. - Task metadata lists accepted end states, so alternate valid routes need no code change.

Keep graders deterministic and free of model calls where possible. If you use a model-based judge for free-text escalation reasons, keep it separate from the state score and spot check it by hand.

Ask in a pilot whether the release ships reference grader code or only the answer key, and which format suits your harness.