Separate detecting an invalid action from proving its replacement

An evaluator can establish that a recorded action violates a necessary condition without knowing the uniquely correct replacement. Keep these as separate outputs: observed-action validity, proposed correction, and evidence sufficiency. If argument checking is gated on independently reconstructing the exact expected tool, uncertainty in that reconstruction can hide an otherwise demonstrable defect. Conversely, catching a real defect does not make the evaluator's proposed replacement correct.

Measure core defect recognition and fully supported explanations separately. Match each finding to its exact occurrence; do not award credit merely because it mentions the same turn. Preserve unknown alternatives instead of inventing a unique correction. A narrow necessary-condition check is not a complete policy proof, and blanket rejection of ambiguous speech can itself create false alarms.

Applies when: Evaluators reconstruct expected actions before judging observed tools or arguments. Keywords: evaluator hybrid expected action necessary condition correction uncertainty false negatives