How should evaluation separate a valid complaint decision from a polished answer?

Start with the decision the task requires, then identify what makes that decision defensible. A useful evaluator asks whether the agent found the relevant evidence, addressed material contradictions and acted within the authority supplied by the environment. A fluent final message cannot establish those facts on its own.

Avoid requiring one exact sequence when several investigation routes are valid. An agent may inspect the transaction history before reading the contact log, or reverse that order. Both routes can be acceptable if the required evidence is considered before the consequential decision.

A generated evaluation pack can define required findings, prohibited actions and acceptable outcome conditions separately. It can also identify facts that must remain unresolved when the available evidence is insufficient. This supports evaluation of justified uncertainty rather than rewarding confident completion at any cost.

Keep communication assessment as a distinct component. The response should explain the decision accurately and avoid claims that exceed the evidence. Reviewing action traces alongside the final response helps locate failures: evidence retrieval, reasoning, authority selection or explanation. That distinction gives developers a clearer basis for improving the agent.