How should a lab score the quality of an assembled evidence bundle?

A useful rubric separates several questions that a single overall judgement can hide. Coverage asks whether the bundle addresses each required claim. Relevance asks whether included material bears on those claims. Sufficiency asks whether the combined evidence supports the stated conclusion. Explanation quality asks whether the model makes that connection accurately.

A model can score well on one dimension and poorly on another. It may retrieve every required record but describe an approval as proof that the approved action actually occurred. It may reach a supported conclusion while attaching large amounts of unrelated material that make review unnecessarily difficult.

Reference answers should allow alternative sufficient bundles where the procedure permits them. Exact attachment matching can penalise a defensible selection or reward a memorised list whose reasoning is wrong.

A buyer can ask for claim-level scoring notes, accepted alternatives and explicit treatment of unsupported assertions. Reviewing disagreements against those notes helps distinguish ambiguous task design from model error. For training, the same dimensions can support targeted feedback: retrieve the missing execution record, remove irrelevant correspondence or narrow a conclusion that exceeds its evidence.