How should evaluation distinguish complete handling from a correct-looking final record?

A final field comparison answers whether the requested value appears in one place. It does not establish whether the business can continue operating correctly. Evaluation should also inspect the constraints that must hold after the agent's actions.

In a professional services account, changing the billing entity may require future unissued charges to reference the new entity while issued invoices remain attached to their original issuer and recipient. A task can therefore have both positive obligations and preservation constraints. Updating every invoice would satisfy a crude text check while violating the business rule.

A buyer can ask for an outcome rubric that separates request fulfilment, dependent record consistency and unintended side effects. The evaluator can report each dimension independently, making it possible to distinguish an incomplete but contained attempt from a completed change that damaged historical records.

This approach also supports useful training feedback. Instead of saying only that a task failed, the evaluation can identify which invariant broke and which action caused it, giving developers a concrete basis for improving agent behaviour.