How can a run that reaches the right queue still be a failure?

An evaluation lead at a hospital network's referral office replays a run on referral REF-3301. The task: capture a GP referral, check completeness, and route it to the correct specialty clinic.

The agent reads the referral, sees all required fields are filled, and routes it to the cardiology queue. The final queue matches the expected one, so a simple grader scores a pass.

The lead then checks the audit events. The patient party record PTY-7720 had a near-duplicate, PTY-7721, with the same date of birth and a transposed street number. The agent never called the party search tool and started a second active referral thread. The expected end state included a duplicate flag on REF-3301, and that field-level assertion fails.

**Consequence:** the run is rescored as a failure on the duplicate-handling criterion, and the lead adds a per-criterion breakdown to the report so that a correct queue no longer hides a missed check.

**What it teaches:** grade intake on process evidence as well as the final queue. Useful checks include which tools were called, which fields changed, which fields stayed untouched, and whether flags were set. Seed each task with at least one trap tied to a named criterion, and confirm the grader can see it.