Author here. We build tooling for rewriting legacy systems, and wanted to test our approach on code we didn't pick, graded by someone else.

Setup: all 21 RepoMod-Bench cases, three unattended runs each, one model throughout. As a control we reran the paper's Codex setup on the same model, since ours is newer than the paper's.

Before anyone asks: we report the best of three runs (control as well); most of the jump from the paper's numbers is the model; two cases are regraded (missing a fixture); and we tuned on the initial three cases, then froze the workflow before running the rest.

Everything is on Hugging Face, about 60 GB.