How should evaluators score selective propagation?

Start with an evaluator-owned applicability set: the work items that require action under the changed rule. Compare it with the agent's selected set before assessing the quality of the resulting updates. This separates failures to find affected work from failures to handle it correctly.

Then inspect unnecessary intervention. An agent that updates every open record may capture the affected cases while disrupting valid work outside the policy's scope. The evaluation should identify those extra changes explicitly, including avoidable holds and unsupported requests for new evidence.

Action quality needs its own assessment. Selecting the correct case does not establish that the agent chose the right transition, preserved valid information or completed the required handover. Conversely, a correctly justified escalation should not be marked wrong merely because the agent did not finish a decision it lacked authority to make.

Use consequence categories suited to the workflow, such as an incorrect release, an unnecessary delay or an unresolved review. Report the categories alongside any aggregate score. This makes the result useful for comparing agent behaviour without allowing strong performance on routine cases to conceal a serious propagation failure.