, by Timon Vonk
93% on RepoMod-Bench: porting 200K-line codebases with a proof in the loop
Bosun's Parity workflow on all 21 cases of RepoMod-Bench, three runs each. Taking the best run per case, 93% of hidden tests passed, against 74% for the paper's Codex setup on the same model and 48% for the best published agent; a 211K-line C++ port passes every hidden test. Half of the differences the proofs found are behaviour the tests never check. Method, limits and the full dataset.