Why not just test agents with a policy document and a yes/no approval check?

**Short answer:** a policy document tells an agent the rules, but it does not say who currently holds which authority, and a yes/no check cannot tell a correct approval from a lucky one.

**What a static policy misses**

- **Who holds authority today.** Limits attach to people and roles, and holders change through leave, acting arrangements, promotion and revocation. - **How limits are measured.** Aggregate exposure, cumulative spend over a period or linked requests may count together, which is only visible in the records. - **Conflicts and exclusions.** A requester may not approve their own item, and a reviewer may be excluded for a related interest. - **Evidence of routing.** Earlier approvals, returns and comments change what the next approver should do.

**What a binary check misses**

A grader that only checks the final answer rewards an agent that approved the right request for the wrong reason. It also cannot distinguish a correct escalation from an unnecessary one, or a justified refusal from a timid one.

**What helps instead**

A task pack can pair each request with the authority records valid at the time, the expected route and a rubric for the reasoning. A simulated environment can add tool actions and log every state change, so a grader scores the path as well as the result.

As with all generated material, buyers should check how well the records are grounded and whether any disclosure risk remains.