EvoCounter: Replication Package for Counterfactual Historical Replay for Evolutionary Compatibility of AI-Generated Patches

EvoCounter is a replication package for the empirical study “EvoCounter: Counterfactual Historical Replay for Evolutionary Compatibility of AI-Generated Patches.” The study evaluates evolutionary compatibility: whether a protocol-admitted alternative patch can support a repository’s observed subsequent development when substituted for the archival repair under a frozen historical-replay protocol. The package contains the final study inputs, execution source code, frozen generation and admission records, candidate patches, historical reconstruction records, replay outcomes, endpoint evidence, reviewer materials, adjudication results, and analysis scripts used to reproduce the reported quantitative results. The study includes 44 repair tasks from 19 public Python repositories and 132 task-system units across three generative systems. Eleven alternative patches met the protocol-defined admission gate. The longitudinal evaluation contains 143 replay conditions across horizons h0, h2, h5, h10, h20, and h30. The package also contains the evidence and completed records for 21 independently reviewed and adjudicated human-context cases. The reported analyses can be reproduced from the frozen records using Python 3.12 or later by running: python Scripts/Analysis/reproduce_all.py The frozen records are the authoritative evidence for the reported results. A fresh end-to-end provider execution additionally requires Git, Docker, Internet access, compatible provider SDKs, and provider credentials. Because generative-system outputs may vary across fresh executions, reproduction of the reported statistical results does not require new model calls or regeneration of the scientific outcomes.