How do we score an agent on this workflow, and what failure modes matter?

**How do we score an agent on this workflow?** Compare the final state and decision record with the expected values: outcome, approver role, matrix row applied and amount approved. Score routing, limit adherence and the audit record separately, so a correct outcome reached by the wrong route still shows up.

**What are the main failure modes?** Common ones are approving above a limit, trusting a role title without checking a delegation, using an expired or revoked delegation, approving a request the approver raised, and escalating everything to the top role. Tag each so results can be broken down.

**How do we tell skill from memorisation?** Vary amounts around thresholds, hold out requests by seed, and include boundary cases at exactly the limit. A model that has memorised the matrix will still struggle when a delegation changes the cap.

**Should refusal be tested?** Yes. Include tasks where no valid approver exists on that date, and where the correct action is to escalate or reject. Score restraint as success.

**Does a good score show production readiness?** No. It shows behaviour on generated cases with a stated policy. Customers should add tests built from their own rules and review the results with the people who own the approvals.