A citation is a claim about a source. SaneCite reads the source.
Of 2,156 random citations we could judge in two tests, blind AI reads found 7.7% (95% CI 6.6–8.9%) that badly misreport the cited paper and 16% to 20% with some error, depending on how reader disagreements are settled (AI only, not yet human-checked). The checker reads the source behind each citation and quotes the passage its verdict rests on; when it can’t tell, it says so. Every result is published for everyone.
Check a paper’s references. See which were retracted.
Enter a DOI, PubMed ID or PMCID. We check every reference with a DOI or PubMed ID against the Retraction Watch database, say whether each retraction came before or after the paper came out, and, where our census read the paper, how it uses the retracted work.
Or paste a reference list
We pick out the DOIs and PubMed IDs and check those. Your text stays in your browser; we keep only a count of how many were checked.
Retraction data: the Retraction Watch database, published openly by Crossref, refreshed every working day. A flag shows what the database records; read the notice before you rely on it. References without a DOI or PubMed ID are not checked. Readings are AI readings from our census (method and accuracy). Something wrong? Write to SaneCite at letters@sanecite.com.
Every citation in the census. Search any paper.
Each citing passage the Phase 2 census read, with the label its three AI models gave it. The checker described further down, which tests whether a citation matches its source, is a separate tool.
Loading the census…
Download everything: citations.csv (one row per citing paper; its retraction year and reason group come from Retraction Watch, not the notices: column notes) · census.json (the results) · citing sentences: contexts.csv · codebook · citing_papers.csv (each citing paper’s PubMed ID and DOI). Bulk files, the live API and how to reuse them: Use our data.
Three readers, one verdict.
- Cited as supportThe citing sentence rests on a finding, number, method or conclusion of the retracted paper, or a review counts it as data. If the retracted data were made up, that sentence would lose its support. This covers using a method or saying results agree, not only building on a result.
- Passing mentionThe paper is one of several references for a point, or backs general background, or is only described.
- FlaggedThe text around the citation warns about the paper: retracted, withdrawn, fraudulent, unreliable, or not replicated. A paper may warn elsewhere; the census reads only the passage around each citation.
- Models splitThree AI models read every passage. Where they did not all agree, the call is marked so you can judge it yourself. On a blind random sample of 150 citing papers, the panel matched a separate AI reader on 142 (94.7%), and on 123 of the 126 where all three models agreed. That check was done by AI, not people; the census is pending human review.
- Found a wrong call?Write to letters@sanecite.com with the citing paper's PMC number. Corrections go into the next release and are listed.
Four steps, every citation.
- 01
Find
Pull every citation out of the paper and give it a stable ID. The whole reference list, every time.
- 02
Locate
Find the passage in the cited source that the claim points at, by meaning, not keyword luck.
- 03
Judge
Compare the claim with the passage and give a verdict, quoting the passage either way.
- 04
Review
Disputed verdicts are marked for people to check. The finished record keeps its audit trail, mistakes included.
Three answers, and one of them is “I don’t know.”
Do citations say what their sources say? About 1 in 13 do not.
Before trusting the checker with named papers, we tested it on ordinary science. We drew 1,000 random citations from open-access papers in PubMed Central (2015 to 2025), none of which the checker had seen, and fixed the reading rules and the pass marks before it ran. A second test drew 1,320 more the same way. Each citation was read next to the paper it cites.
of the 2,156 citations that could be judged in the two tests badly misreport the paper they cite (166 of 2,156; 95% CI 6.6–8.9%). Some error: 20.3% in the first test; 19.2% or 15.7% in the second, depending on how reader disagreements are settled.
of the first test’s serious errors were caught by the checker (56 of 66; 95% CI 74–92%). 56 of its 81 serious-error calls were right (69%).
per 1,000 citations to run it, at list price. A low-cost screen reads every citation and passes about 3 in 10 to a vote of three models.
These figures are measured against AI readers, not people. A check by two human labellers on a random sample is designed, and we will publish its result here whatever it shows.
- Citations that credit a specific finding are the riskiest.In the first test, 9.7% of them badly misreport the cited paper (40 of 413) and 28.3% have some error, against 5.6% for background citations and 4.8% for methods and tools. A separate, earlier set of 500 citations shows the same order (7.8%, 5.4%, 3.6%). This comparison was not in the test’s plan.
- Many serious errors are not close calls.In some, the cited paper is from another field altogether; in others, it reports the opposite of what it is cited for. This comes from the readers’ one-line reasons, not a coded count.
- The checker matches the answer key closely.Its agreement with the answer key (kappa 0.74) is higher than the agreement between the two AI readers who built the key (0.66).
- It over-calls a little.On the first test it marks 8.9% as serious against 7.2% in the answer key, and about 3 in 10 of its serious calls are wrong. A serious call is a lead to check, not a verdict.
How the test worked
- Sample1,000 random pairs of a citing passage and the paper it cites, both open access, with citing papers not used in any earlier test. 996 could be read and 915 judged; the rest gave too little to decide. The second test read 1,320 and judged 1,241. Its serious-error rate (8.1%) passed the plan’s test for agreeing with the first within 3 points, so the headline pools the two.
- Answer keyTwo AI readers from different model families read every citation blind. Where they disagreed (169 citations), a third careful read decided. In the second test the readers disagreed on 213 citations; settling those by a two-of-three vote instead changes 113 of them, which moves the any-error rate (19.2% of 1,241 judged, to 15.7% of the 1,191 the vote can judge) far more than the serious-error rate (8.1% to 8.5%, on the same bases). Where one reader’s tool stopped partway, a reader from a third family took over (66 citations), as the plan set out in advance.
- CheckerCloudflare’s Clef screens every citation; DeepSeek, Qwen and GLM vote on the ones it flags. The models, settings and threshold were fixed and recorded before the test.
- Pass marks, set in advanceAgreement (kappa) of at least 0.60, at least 70% of serious errors caught, and at least 65% of serious calls right. All three passed. A second version with TypeSafe’s Jev as the screen also passed (83.3% caught, 66.3% right, kappa 0.715).
- Changes to the planEach is dated in the registered plan. 22 Oct 2026, before any result: smaller reading batches with a backup reader; Clef as the screen unless Jev was clearly better; the backup reader for the stopped citations (changed twice); how disagreements were settled. 23 Oct 2026, after the result: publish now, pending human review, instead of after the human check.
The results as data: batch6-result.json and batch7-result.json. Earlier, a first set of 500 citations gave 6.2% (29 of 471).
9.7% of citations that credit a specific finding badly misreport the paper they cite.
About 1 in 14 citations overall (7.2%). Background citations: 5.6%. Methods and tools: 4.8%.
Badly misreport the cited paper, by what the citation is used for
Why it matters. Citations that credit a specific finding are the ones readers act on, and they are the ones most often wrong about what the cited paper says.
Not in the test's plan; the same order held in an earlier, separate set of 500 citations (7.8%, 5.4%, 3.6%).
Not in the test's plan. The same order held in an earlier, separate set of 500 citations.
Source: SaneCite citation-accuracy test (batch 6): 915 random citations from open-access PubMed Central papers, 2015–2025. Measured against AI readers, not people; pending human review. Data as of 22 Oct 2026.
Source: SaneCite citation-accuracy test (batch 6), 915 citations, 22 Oct 2026, pending human review · sanecite.com/charts
Show the numbers
| What the citation is used for | Judged | Badly misreport | Share (95% CI, %) | Any error |
|---|---|---|---|---|
| Credits a specific finding | 413 | 40 | 9.7% (7.2–12.9) | 117 (28.3%) |
| Background | 357 | 20 | 5.6% (3.7–8.5) | 50 (14.0%) |
| Method or tool | 124 | 6 | 4.8% (2.2–10.2) | 17 (13.7%) |
| All judged (incl. 21 data-source citations) | 915 | 66 | 7.2% (5.7–9.1%) | 186 (20.3%) |
1,000 random citation pairs from Open-access PubMed Central research and review articles, 2015-2025, citing an open-access paper with a PubMed ID; 915 could be judged. Answer key: two blind AI readers from different model families, disagreements settled by a third read. 95% intervals are Wilson intervals.
Models picked by a public test.
The census reading ran on models served by Cloudflare. No model is trusted on its word: each candidate is scored against an answer key, and the winner is picked for accuracy first, then speed, then cost.
For the Phase 2 pilot, 12 models were tested on 73 citations labelled by a separate AI model (Claude), not by people, and the best 8 again on 118. The best single model scored 93% on the 118. A panel of three (Cloudflare’s Clef, TypeSafe’s Jev and Kimi K2.6) scored higher, so a panel reads every citation and the majority wins. On 73 more citations that Claude labelled blind, that pilot panel agreed 96% of the time. The full census used Jev, Clef and Google’s Gemma 4; where Jev had no vote, Gemma and Clef decided and Clef broke ties. On a fresh blind sample, the census panel agreed with a separate AI reader 94.7% of the time. A check of a fresh sample by human labellers is designed and pending. The full scorecard (the 8 models on 118 citations) is public, and so is every model test we have run since, with every score, cost and setting.
Corrected 31 Oct 2026: this paragraph said 12 models were tested on 118 citations (12 were tested on 73, then the best 8 on 118), and gave the pilot panel’s 96% after the census panel as if it were the census panel’s.
That keeps the cost of reading close to nothing. The expensive part, and the part your support pays for, is people checking the machine.
A checker you can’t check is just another authority.
So there is no black box. Every census dataset is free. All the code behind the Campaign One census is public under the MIT license, with the data, on GitHub: the Phase 1 count, the Phase 2 reading panel and the code that calls the models. Don’t trust our census verdicts. Re-run them.
The checker used in the accuracy test above is public in the same repository, in its checker folder: the scripts, the locked settings, the reference reads and every call, with the code under the MIT license and the data under CC BY 4.0. The quoted passages from other authors’ papers are left out; each row gives the IDs to find them.