Abstract
Terminal resolve rate obscured where repository-level language-model patches lost utility between generation and execution. This study addressed that measurement gap by proposing an auditable patch-validity framework whose conceptual novelty was to represent each artifact as monotonic survival through six evidence-linked checkpoints: non-empty generation, syntactic validity, repository applicability, harness completion, behavioral resolution, and review readiness. The first five were measured from archived generations, normalization records, and Docker-harness outcomes; review readiness remained a proposed deployment check. In a 300-artifact diagnostic slice, 64 failed syntax, 77 failed applicability, 151 reached execution but remained unresolved, and 8 resolved. Across 1,600 larger-run artifacts, 776 (48.5%) were invalid or inapplicable and 22 resolved. Staged prompting improved completion in some configurations, but its strongest diagnostic resolution advantage, 4/50 versus 1/50, was exploratory (p = 0.25) and reversed at larger scale. By converting terminal outcomes into syntactic, repository-state, and semantic failure tiers, the framework enabled low-cost pre-execution filtering, targeted remediation, and auditable CI resource allocation. It was a measurement and triage instrument, not a new repair algorithm.