Asking Retraction Watch Twice: Retraction Data Revise Their Own Past

Retraction data are increasingly treated as infrastructure: metascience studies count retractions by year and reason, reference managers warn authors who cite retracted work, and retrieval systems filter retracted papers out. Since Crossref opened the Retraction Watch database in 2023, it has been republished daily, and studies cite it as if it were one fixed dataset. We replayed 39 snapshots spanning its entire public history (2024-09-02 to 2026-10-16) and asked the same four questions of each, all about retractions dated 2023 or earlier, so that any change is the database revising the past. Over two years the number of such retractions grew 9.1%, the number citing “Paper Mill” grew 80.6%, the median time from publication to retraction rose from 379 to 420 days, and the count citing “Fake Peer Review”, 4,654 in the first snapshot, fell to zero when the label was renamed. Almost all of the growth is late cataloguing: 4,482 records added after the first snapshot, 3,921 of them Hindawi retractions. We also join Retraction Watch with OpenAlex, PubMed, and PubMed Central. Of 89,681 papers that at least one registry flags, all eligible registries agree on 51.5%, and 49.3% of the works OpenAlex marks as retracted are retraction notices rather than retracted papers. We then checked by hand, blind to the source and with the rules committed publicly before any label, 200 of the 36,401 papers that only one registry flags. An estimated 95.0% (91.5% to 98.6%) were in fact retracted, so most disagreement comes from a registry missing a retraction, and few single flags point at papers that were never retracted. We recommend that studies cite and archive the snapshot they used, and we release the pipeline that rebuilds any snapshot on demand.