Has anything you cite been quietly corrected?

Paste your bibliography. This tells you which of your references carry a published correction, erratum, expression of concern, withdrawal or retraction — not just the retractions everyone checks for.

Paste a reference list, a BibTeX file, or one DOI per line. Anything works — the DOIs get picked out of the mess. Vancouver-style entries that cite by PMID work too: those are looked up in PubMed first. Nothing else in your text is read.

Or open the file you already have — a .nbib straight from PubMed’s “Send to → Citation manager”, a .bib, a .ris, Zotero’s CSL JSON, a .txt, or the manuscript itself as a Word .docx. You can also drag it onto the box below. The file is read inside your browser and is never uploaded; only the identifiers found in it are sent, to Crossref and PubMed (and, for the ones neither can find, to doi.org).

Watch this bibliography over time

The notice that matters most is usually the one published a year after you filed the reference away. This page can tell you what changed since the last time you ran it — but it will not keep your reading list in this browser to do it. On a shared computer that list should not outlive your session. So you get a small file instead: keep it wherever you keep your own notes, and hand it back here next time. It is the same file the command-line version writes, so you can start here and carry on in a terminal, or the other way round.

What this is, and what it is not

Retractions are the famous case and free tools already cover them. But a retraction is the rarest kind of change to the scientific record. Corrections and errata are several times more common, and nobody emails you when the figure you built an argument on was corrected two years after you read it — the paper is still perfectly valid, which is exactly why nobody looks.

A change notice is not a verdict on the science. Most corrections are a typo, a wrong unit, a mislabelled axis. This tells you something changed; reading what changed is your job.

It asks two registers, because one is not enough. Crossref only holds what a publisher deposited, so a correction nobody registered is invisible there. PubMed keeps the same information separately, compiled by NLM indexers rather than by the publisher, and the gap between them is not small. Sampling 400 records per category on 11 September 2026: of the papers PubMed says carry a notice, Crossref’s data would let you find 78.8% of the corrections, 91.8% of the expressions of concern and 93.7% of the retractions. One corrected paper in five would have come back clean here if this page asked Crossref alone. Anything found only in PubMed is labelled per PubMed so you can see which register spoke.

That number measures Crossref against PubMed, not against the truth — PubMed will have its own holes, and this page does not know their size. It also only covers biomedicine, because PubMed only covers biomedicine. Two registers beat one; two registers are still not all of them. Anything without a DOI stays invisible to both, so preprints, books and chapters often come back “not found” — and this page says so instead of pretending they came back clean.

A reversed retraction still reads as a retraction, and that is the worst thing here. Sometimes a paper is retracted and then reinstated — the journal got it wrong, or the authors were cleared. Neither register represents that: Crossref has no such notice type, and the note restoring the paper is usually filed as an ordinary article linked to nothing. Retraction Watch lists 155 reinstated papers with a DOI, and on 16 September 2026 this page still called 31 of them retracted. There is no signal in the data to read, so nothing is done about it here beyond saying so. If you have reason to doubt a RETRACTED verdict, check it at retractiondatabase.org, which is the one place that records reversals.

PMIDs have a hole in them, and it is measured. A PubMed id is translated to a DOI before anything can be checked, and not every record has one. Sampling 1,558 random PubMed records on 9 September 2026: 72.5% listed a DOI — 95.7% of the recent ones, but only 48.6% of those from before 1990. Since 18 October 2026 a PMID with no DOI is still checked, against PubMed's own record, which lists retractions and errata without needing a DOI: on the Retraction Watch list that found 636 notices this page used to call not checked. The report says checked against PubMed only for those, because a silence from one register is weaker than a silence from two.

Your bibliography is not uploaded anywhere

There is no server behind this page. Your text stays in your browser; only the identifiers found in it leave it — and since 11 September 2026 they leave it twice, because both registers get asked: every DOI and PMID goes to Crossref’s public API and to NCBI’s public E-utilities. It used to be that only your PMIDs reached NCBI; that is no longer true, and saying so plainly matters more than the sentence looking tidy. Since 18 September 2026 there is a third destination, and a much narrower one: any DOI Crossref did not return is sent to doi.org to ask which register owns it — that is how this page can tell a mistyped DOI from an arXiv preprint. Only the DOIs that already failed go there, never your whole bibliography. And since 19 September 2026, when a DOI turns out not to exist, the digits on its tail are sent to NCBI to ask whether they are a PubMed id — that is how this page tells a broken citation from a DOI deposited with a PMID welded to the end of it. Again, only for DOIs that already failed everywhere else. All three are public services with their own privacy policies; nothing goes through me. No account, no cookies, no analytics, nothing logged by me. You can save this page and run it offline from your own disk — it is one HTML file with no dependencies.

Since 14 September 2026 this page does remember, inside the one tab you have open, which identifiers it has already checked, so that running a growing export again does not ask two free public registers the same thousand questions every time. That memory is sessionStorage: it belongs to that tab and is gone when you close it, which is the point — on a shared computer the list of what you have been reading should not outlive your session. Nothing is remembered about a lookup that failed, or one that only Crossref answered. The result says how many answers came from that memory and how old the oldest one was, and there is a button to ask both registers again.

Watching the same bibliography over time — for the notice published a year after you filed the reference away — has been here since 21 September 2026, under Watch this bibliography over time above, and it works without this page remembering anything about you. To compare today against last month something has to survive the tab, and on a shared library computer that something must not be the browser: it is a file you keep and hand back. It is byte-for-byte the file the command-line version writes with --watch, so a bibliography can be started here and carried on in a terminal, or the other way round. Treat it as what it is — the list of what you have been reading — and keep it where you would keep your own notes.

Two things it will not do. It will never report a notice as lifted because a lookup failed: if the register that knew about a retraction did not answer today, that is filed under could not check today, with the retraction still standing. And a watch run does not reuse this tab’s memory — “nothing new” has to come from the registers, not from something this page already had on hand.

References that have left your bibliography are kept in the watch file rather than dropped, so that loading the wrong file once cannot destroy the only baseline you have; tick the box on the save bar when you do want them gone. If the file you loaded turns out to have nothing in common with the text you just checked, the report says so at the top — measured on 22 September 2026, that mix-up otherwise looks exactly like a review that was rewritten, and saving merges the two into one file with no way to separate them again.