How would an evaluation case score a patient set-up agent?

A reseller's customer is testing an agent that sets up patient accounts after a specialist referral is accepted. The evaluation case is built from referral REF-6630, accepted by the intake coordinator for a physiotherapy program with a funded session limit.

The agent searches the directory and finds PAT-20418 with a matching name but a different date of birth. The correct behaviour in the case's policy is to open a new patient record and flag a possible duplicate for review. The agent instead merges the two records.

The scorer compares the final state with the expected end state. The merge fails the check `no_unreviewed_merge`. The agent did correctly create a funded-sessions entitlement of the right size on the account, passing `entitlement_matches_referral`, and opened an appointment-booking task for the scheduling role, passing `handover_task_created`.

The result is a partial score with a named failure. The consequence in practice would be two people sharing one clinical history, which is why this case is weighted heavily in the pack.

**What it teaches:** a good evaluation separates checks so a buyer sees which behaviour failed, rather than one pass or fail. It also shows why near-match cases matter: an agent that is accurate on clean cases can still fail where a wrong merge is costly.