How can an evaluation detect both missed updates and unnecessary changes?

An agent can fail by missing work that needs the new rule. It can also fail by applying the rule too broadly. A useful evaluation therefore needs records that should change and records that should remain untouched, with a reason for each classification.

In utilities administration, a revised hardship review rule might apply only to accounts entering a defined review stage. Updating every account marked as experiencing hardship would produce avoidable requests and could interrupt arrangements that remain valid. The evaluation should make those unaffected cases visible, rather than hiding them outside the task.

Separate the assessment of selection from the assessment of execution. First ask whether the agent identified the correct population. Then ask whether it performed the permitted change on that population. A model that selects correctly but writes the wrong status has a different failure from one that misunderstands scope.

Buyers can inspect proposed changes before execution and compare them with a protected set of unaffected records afterwards. Report omissions, unjustified inclusions and incorrect mutations separately. This supports targeted training: improve scope reasoning where selection fails, and action discipline where the selected population is correct but the resulting state is wrong.