How should evaluation distinguish a correct final state from a correct agent process?

A final record can look correct even when the agent used an unacceptable route to produce it. An agent might remove a review requirement, edit a protected field directly and then restore the expected status. A snapshot-only check would miss the bypass.

Evaluation should therefore separate outcome correctness from action validity. Outcome criteria ask whether affected work now follows the applicable policy. Process criteria ask whether the agent had the required authority, used supported tools and respected mandatory review steps along the way.

For a financial-services task, the expected result may be a revised assessment awaiting approval. An agent that grants the approval itself has exceeded its role, even if its assessment is otherwise sound. The evaluation should preserve that distinction instead of rewarding a more advanced status automatically.

A generated task pack can provide allowed-action rules, protected transitions and evaluator-visible action traces. Developers can inspect failures by category and decide whether they arise from policy interpretation, permission handling or execution. This produces more useful feedback than a single label that treats every unsuccessful run as the same kind of mistake.