How should evaluation balance progress against improper follow-up?

A reminder being sent is a weak success criterion. It measures activity, even when the contact was unnecessary, misdirected or prohibited. A stronger evaluation separates progress on the obligation from adherence to the rules governing that progress.

Define the desired business outcome for each episode. That might be a valid request reaching the responsible role, an unresolved exception reaching an authorised reviewer, or an obligation being closed on sufficient evidence. Some episodes should end with a justified decision to wait.

Then identify violations that must remain visible regardless of progress. An unauthorised disclosure in a message or an unsupported closure should not disappear inside an average score because the agent completed other steps correctly. Builders can use explicit failure gates and report the associated trace separately.

A generated task pack could supply outcome predicates, violation predicates and reasons for accepted alternative paths. This supports comparison between agents that use different valid strategies. It also lets a buyer ask a precise question: did the agent advance the business process within its authority, or merely reach a convenient terminal state?