How do reset, event logs and graders work in a simulated escalation environment?

For an agentic evaluation, the environment needs to be resettable and the graders need hooks into its state. A workable design:

- **Seed.** Each task loads a snapshot, for example `snapshot_id: esc-claims-014`, containing tables for cases, authority, policy and tasks - **Reset.** A `reset(snapshot_id)` call restores all tables and clears the event log, so runs are independent - **Actions.** Tools such as `get_case`, `get_authority`, `raise_referral`, `record_decision` and `add_note` each write to an append-only `event_log` with actor role, action, arguments and result - **Guards.** Out-of-authority actions return an error and write an `authority_violation` event instead of silently failing - **Simulated reviewer.** A scripted senior role responds to a referral with approve, decline, condition or request-for-information, driven by seeded rules so replies are repeatable - **Graders.** Functions read the final tables and the event log, returning separate scores for boundary, end state, handover content and follow-up

Expected outcomes can be stored in a `gold` table per task with the accepted end states, since some cases have more than one defensible path. Store rubric text for the handover note beside it.

Output of each run can be written as JSONL with `task_id`, `run_id`, `events`, `final_state_hash` and `scores`, which makes comparison across models straightforward. Ask for a sample environment definition and one reset-and-run walkthrough before scoping.