Which evaluation hooks reveal whether a routing action preserved the required handover?

An evaluation harness can subscribe to generated events such as `assignment_submitted`, `assignment_accepted`, `ownership_released` and `handover_read`. Each event should carry a request identifier, action identifier and state revision so the evaluator can reconstruct the transition.

Define assertions over the resulting state. A request might require an accountable owner throughout transfer, a destination acknowledgement before the prior owner is released and a handover containing specified case references. These assertions test different properties from destination selection.

For content transfer, distinguish required information from restricted information. A receiving team may need a delivery exception code without needing unrelated account notes. The evaluator can check field-level inclusion and exclusion against the task's permission policy.

Keep private assertions and expected outcomes outside the agent's accessible tools. Return diagnostic categories after the run, such as missing handover context, premature ownership release or unauthorised field disclosure.

A generated task pack could supply these hooks alongside state schemas. Developers can then compare runs using evidence from actual transitions, even when agents produce equally persuasive final messages about what they believe they completed.