Corrections
The mistakes we have found in our own published numbers, dates and claims, newest first: what was wrong, what we changed and who found it.
17 Nov 2026 · Lists: a journal’s reply
What was wrong. The entry for the journal Cells on the Lists said no reply was received by 8 November. Its publisher, MDPI, had sent a statement on 6 November from an address we had not written to, and we missed it.
What we changed. The statement is beside the entry, with a dated note. Our check for replies now covers mail from other addresses at an organisation we wrote to.
Found by. The publisher.
16 Nov 2026 · Reference checker
What was wrong. The checker looks up a paper’s reference list in two outside sources, Crossref and Europe PMC. When one of them did not answer, the checker showed the part of the list it had as a full check.
What we changed. The checker now says “Partial check” and names the source that did not answer.
Found by. Us.
9 Nov 2026 · Which citations count as made after a retraction
What was wrong. We counted a citation as made after a retraction when the citing paper’s year was later than the retraction year. A paper’s year is the year of its printed issue, and many papers go online months earlier. So some papers on the Lists were online before the work they cite was retracted.
What we changed. We now compare the date a paper first went online with the date of the retraction. We found 302 such citations on the Lists and checked every date again. 296 were confirmed and are now left out of the Lists. The rule in full
Found by. Us.
9 Nov 2026 · Lists: readings and retraction dates
What was wrong. 25 citations on the Lists were labelled cited as support. Read in full, each citing paper flags the cited work as retracted or unreliable. Nine retraction dates were a year late: the notice was online a year before the date we held.
What we changed. The 25 readings now say flagged as retracted or unreliable, and the nine dates are corrected. Every change is dated in the corrections file.
Found by. Us.
9 Nov 2026 · Check of the AI audit of Saez and Zucman
What was wrong. Where the audit disputed a choice of method, we graded the choice against the paper’s own text and sources only. A method the paper explains could pass even when others measure the same thing another way.
What we changed. A third pass compared each disputed method with what others do. It moved 4 problems to a new verdict: changes a main result if you use the standard method. The check
Found by. A reader.
4 Nov 2026 · Blinded review table
What was wrong. Our table of scores for BARS, a tool that grades papers with the title, authors and journal removed, first drew on a pilot table in the tool’s repository instead of its final results. That gave one retired label and ten retired scores.
What we changed. The table uses the final results. Every score in it changed; BARS’s own row went from 0.927 to 0.946.
Found by. Us.
31 Oct 2026 · How the AI models were tested
What was wrong. The checker page said 12 models were tested on 118 citations. In fact 12 were tested on 73 citations, then the best 8 on 118. It also gave one test panel’s 96% agreement as if it belonged to the census panel, whose figure is 94.7%.
What we changed. The page gives the right figures, with a dated note.
Found by. Us.
30 Oct 2026 · Lists and Stories: check after publication
What was wrong. A check of every entry the day after the Lists were published found errors in single entries. A retraction notice dated 11 Dec 2022 was online on 2 Dec 2022. One journal’s count of citations as support was one too high (68, now 67). A card said two of three registered trials use a survey from a retracted paper, when all three do. A story quoted one retraction notice as the reason for ten studies, when one of the ten was retracted for a different reason.
What we changed. Each entry is corrected. Changes to the Lists are dated in the corrections file.
Found by. Us.
30 Oct 2026 · Download files on the Lists
What was wrong. Two download files published with the Lists on 29 Oct included our readings for 9 authors and 8 journals who were not named on the page and whom we had not yet written to.
What we changed. We removed those rows from the site the next morning, and from the copy on GitHub on 1 Nov. The site now refuses to build if a file names anyone the page does not.
Found by. Us.
20 Oct 2026 · Census
What was wrong. 12 census citations were matched to the wrong retraction record.
What we changed. Without them the census read 55,676 citations, not 55,688, and judged 51,614, not 51,626. The headline rates did not change. The same check on all 22,553 retracted papers in the census found 70 wrong records with 86 later citations. The note on the charts page
Found by. Us.
16 Oct 2026 · First count
What was wrong. Our first count of citations made after retraction, 286,050, included paperwork retractions, where the journal found no fault in the science.
What we changed. We left those out, which gave 262,998.
Found by. Us.
15 Oct 2026 · Headline figure
What was wrong. We published 40.8% as the share of citations that still cite a retracted paper as support. The figure had been rounded twice. From the counts (21,037 of 51,626) it is 40.7%. The figure for the first two years after a retraction was given as 43.5% and is 43.6%.
What we changed. Every page, chart and file carries the corrected figures, each worked out once from the counts.
Found by. Us.
Smaller fixes
9 Nov 2026: the letters report said “the other nine” letters where the number is eight. 22 Oct 2026: a home page chart said 1,833 citations “since 2010”; its bars ran from 2012 and added up to 1,834. 20 Oct 2026: one chart group was labelled “Neither” and is now “Includes another subject”, and the chart notes said the likely ranges came from “thousands of” re-runs where it is 400. All found by us.
Found a mistake?
Write to letters@sanecite.com. We reply within two working days.