Same repo, same task, both diffs side by side is the only comparison that holds up. I would add one column: how many turns each needed before the tests passed, not just the final diff. Did Pi need more nudging on multi-file edits?