What evaluation hooks and reset behaviour should a booking environment provide?

An evaluation harness for this workflow benefits from five hooks.

**1. Snapshot reset**

`reset(task_id)` restores the seeded tables for that task, clears the event log and returns the first observation. The same task id and seed should always give identical state.

**2. Scheduled events**

Tasks list events by step number, for example an approval amendment at step 6. The environment applies them on the same step in every run, so differences in outcome come from the agent.

**3. Trace logging**

Every tool call is written as a JSONL line with `step`, `tool`, `arguments`, `result` and a hash of affected rows. Traces support error analysis and can be used as training data in your own pipeline.

**4. State-diff grader**

`grade(task_id)` compares final tables with `expected_state` and returns structured output:

- `passed` as a boolean - `field_diffs` with table, key, expected and actual - `failure_modes`, a list drawn from an agreed taxonomy - `allowed_outcome_matched`

**5. Policy probes**

Flags recording whether the agent read the amended approval or the relevant policy document before the write.

Agree the grader output schema in scoping, then test it against known-bad runs in a pilot.