How should evaluation handle more than one acceptable destination?

Some routing tasks permit several correct outcomes. A utility’s generated operating rules can allow either of two regional support queues to handle a general account enquiry, provided the selected queue covers the service location and accepts that request category.

An evaluator can express correctness as constraints over the final state. The destination must be eligible, the request must have an accountable owner and any required evidence must exist. A preferred destination can be scored separately if the policy describes a preference rather than an obligation.

The evaluation should also distinguish invalid routing from justified deferral. A request lacking the service location may belong in an authorised information-gathering state. Marking every non-assignment as failure would reward agents for guessing.

A buyer can ask MirrorFirm for generated tasks with admissible outcome sets and the policy reasons behind them. Inspect whether those sets include all supported alternatives and exclude destinations that merely have similar names. This helps builders avoid a brittle answer key and makes disagreements reviewable: the question becomes which constraint was violated, rather than whether the agent selected the evaluator’s favourite queue.