Repair Evidence Resolution: Separating Prediction, Structural Resolution, and Transfer in Automated Patch Correctness Assessment

Automated patch correctness assessment (APCA) is usually evaluated by predictive performance: accuracy, precision, recall, F1, ranking quality, or related measures. These metrics answer whether an assessor tends to make the right decision, but they do not determine whether the evidence visible to the assessor is structurally capable of separating repairs that require opposite correctness decisions. We introduce Repair Evidence Resolution (RER), a claim-relative framework for studying that prior question. RER models assessor-visible evidence as a partition of a declared repair population and asks whether every evidence cell is homogeneous with respect to a repair claim. It yields repair-claim resolution (RCR), concrete claim-conflict witnesses (CCWs), resolution deficit (RD), claim-resolution profiles, and claim-boundary completeness (CBC) for transferring resolution conclusions from evaluated repair subsets to larger target populations. We prove basic separation results and explicitly reduce exact-identification special cases to established teaching/certificate-complexity machinery. We then implement a deterministic audit pipeline and evaluate a publisher-verified ASE 2020 APCA artifact. Across 1,762 generator-aware records, one released evidence channel produces nine mixed correctness cells covering 43 patches (RD=2.44%), whereas a second channel and the combined evidence are exact-cell claim-resolving on the observed corpus. We also derive the exact probability that restricted evaluation subsets falsely appear resolved: for a sample of 100 records it is 91.29%. The results motivate a three-layer evaluation view: prediction, resolution, and transfer.