How should evaluation teams score review quality beyond the final decision?

A final decision is only one part of review quality. An agent can choose the expected outcome using the wrong evidence, or produce a sound assessment and then perform an unauthorised action. Combining everything into one score can hide these differences.

An evaluation rubric can score evidence selection, policy application, uncertainty handling and action validity separately. Evidence selection asks whether the agent used relevant records. Policy application asks whether those facts satisfy the stated rule. Uncertainty handling asks whether unresolved questions remain visible. Action validity checks whether the resulting transition is allowed.

Some failures should prevent an overall pass regardless of other strengths. Closing a review while knowingly leaving a mandatory requirement unsupported is different from writing an awkward but accurate summary. The rubric should define those critical failures before comparing systems.

MirrorFirm could generate evaluator notes that explain each dimension and its decisive evidence. Buyers can inspect a sample of scored runs to see whether the rubric rewards the behaviour they need, including cases where a plausible narrative conceals an invalid operational outcome.