What evaluation hooks should a simulated intake environment expose?

An environment is only as useful as the evidence it returns. These hooks make runs comparable.

- **Reset:** `reset(snapshot_id)` restores all stores to a named starting state, so each run begins identically. - **Action log:** every tool call is recorded with `task_id`, `step`, `tool`, `arguments`, `before` and `after` values, in JSONL. - **State export:** `export_state(task_id)` returns final case, party, document and note records for comparison. - **Grader:** a function takes the exported state and the task's `expected` block and returns per-field results, for example `final_status: pass`, `reason_codes: fail`. - **Violation check:** forbidden actions in the log are listed separately from outcome mismatches.

A run report can then carry `task_id`, `agent_version`, `field_results`, `violations`, `steps_taken` and `escalated` as columns. Written to Parquet, the reports can be grouped by exception class, case type or policy to show where an agent version improved or regressed.

For training use, the same action logs become trajectories, with the expected block supplying a reward signal or a filter for keeping only successful runs. Keep graded and training splits separate by `task_id` so results are not contaminated.

Agree the hook names and return formats during scoping so they fit your harness without a wrapper layer.