How should evaluators score partially completed customer changes?

A customer change can succeed in one place and fail elsewhere. An agent may update the primary record correctly while leaving an active dependency untouched or altering an unrelated record. A single completion label hides these differences.

Define the expected outcome as a set of obligations. Each obligation should identify the relevant record, the required condition and the evidence used to assess it. Separate required changes from prohibited effects and permitted unresolved states. A pending approval can be correct when the task explicitly requires another role's decision.

Report these dimensions separately before combining them into any aggregate score. Otherwise, several minor successes can obscure a serious unauthorised effect. The scoring design should state which failures prevent a task from counting as successful and why that rule matches the supplied business policy.

Evaluation teams can inspect generated answer keys against these obligations, then review a sample of agent traces to confirm that the checks reflect actual behaviour. This approach also supports diagnosis: it shows whether an agent misunderstood scope, omitted a dependency or claimed completion while an obligation remained open.