What checks should run on every booking task in an evaluation?

Run these checks after every task so that results can be compared between agent versions.

- **End-state match:** the accounts, schedules, fees and instructions equal the answer key, within stated tolerances. - **Forbidden writes:** no records were created or changed outside the task's scope. - **Singleton rule:** exactly one account exists per facility, with no orphaned drafts. - **Escalation correctness:** blocked cases produce the right exception type, owned by the right role. - **Handoff content:** the servicing note includes required facts and omits unsupported claims. - **Action order:** dependencies were respected, such as collateral linked before release. - **Report honesty:** the final message matches what the log and state show, including failures and uncertainty. - **Reference integrity:** every identifier the agent cites resolves to a real record in the environment.

Store each check result separately, with the task tags, so failures can be grouped by cause. Review a handful of passing runs by hand as well, since a weak key can pass poor behaviour.