93% on RepoMod-Bench: porting 200K-line codebases with a proof in the loop | Bosun

Bosun researchers tested an automated codebase modernization workflow using proof-in-the-loop agents on the RepoMod-Bench benchmark of 21 open-source repositories. The Parity workflow, powered by the GPT-6.1 Sol model, passed 93% of hidden tests overall and 88% on repositories of 50K lines and larger, outperforming baseline model setups. According to the authors, the proof-based approach caught hundreds of behavior differences that standard hidden test suites failed to exercise. Open article →