How can evaluation detect routing loops and premature completion?

An evaluation harness can consume assignment events in sequence and reconstruct the owner of each request after every action. Useful fields include `episode_id`, `request_id`, `event_sequence`, `source_queue_id`, `destination_queue_id` and `transition_reason`.

A loop detector can identify a return to a previously visited queue without new evidence or a policy-supported reason. Repeated visits are not automatically wrong: the evaluator needs an explicit rule distinguishing legitimate reconsideration from circular referral.

Terminal checks can inspect whether the request has a valid owner, whether required acceptance occurred and whether the final status matches the task objective. An agent that reports completion while the request remains rejected should fail that condition, even if it previously selected the reference destination.

MirrorFirm could provide generated event trajectories and evaluator configuration for this purpose. Resetting to the same initial snapshot allows alternative agents to encounter the same starting conditions. The evaluation output can retain a failure event reference, enabling reviewers to inspect the exact transition that introduced a loop, abandoned ownership or an unsupported completion claim.