Five citation classics: retracted, corrected, or clean?

Anyone who cites a famous paper is quietly asserting a second fact on top of the first: not just "this paper says X" but "no editor has since pulled X back." The first fact gets checked constantly. The second almost never — and yet it now has a machine record. Crossref's update-to and updated-by relations are, in effect, the DOI system's ledger of retractions, corrections, errata, and expressions of concern, fed by publisher deposits and by Retraction Watch's data. This notebook audits five heavily cited papers against that ledger and against each paper's version-of-record page, and its main product is deliberately unglamorous: four honest nulls and one loud exception.

Why a null is a finding

"No retraction on record" is not the same claim as "this paper is fine," and it is not even one claim — it is as many claims as there are places a retraction could have been recorded. So each null here is scoped to the surfaces actually checked, and dates each surface. For Watson & Crick's 1953 DNA structure paper, the claim [C1] rests on the live Crossref record, which carries no update relations [A1], and on Nature's own article page, rendered live on the claim date and carrying no editorial banner [F1]. The same two-surface pattern covers LeCun, Bengio & Hinton's 2015 "Deep learning" [C4], whose Crossref record shows only a has-review relation — post- publication reviews, not an editorial event [A4, F2].

Lowry et al. 1951, the most-cited paper in the scientific literature, gets a third surface: its PubMed record, whose CommentsCorrections fields — the place PubMed flags retractions and errata — are simply absent [A19]. With the Crossref record [A3] and the ScienceDirect page [F6], that null [C3] has the broadest coverage in the set.

Shannon 1948 has the narrowest, and the claim says so [C2]. The Crossref record exists but is a thin backfile deposit [A2]; the publisher page could only be captured as a May 2025 archive snapshot, because live Wiley sits behind a challenge wall and IEEE Xplore answered our fetcher with HTTP 418 [F7]. A probe of PubMed came back empty for the whole of the Bell System Technical Journal — a genuine corpus gap, recorded as such on the question page [Q2]: PubMed cannot speak to this paper either way. A null's weight is its coverage, and Shannon's is honestly the lightest of the four.

Every null is also armed to die. Each question page carries reopen conditions — "a Crossref update relation appears for this DOI," "the publisher attaches a banner" — so the answers expire the moment the world changes [Q1–Q4].

The one that wasn't clean

Séralini et al. 2012, the Roundup-and-GM-maize rat study, is the contrast case, and the same surfaces that stay silent for the other four light up completely. Its Crossref record's title begins "RETRACTED:", and its updated-by field names the retraction notice twice — once deposited by the publisher, once from the Retraction Watch feed, record 3604 [A5]. The notice itself has its own DOI and Crossref record, placing it in Food and Chemical Toxicology volume 63, January 2014, page 244 [A6, F5, C5]. The archived version-of-record page carries the full retraction statement, which is more careful than the episode's reputation: the Editor-in-Chief "found no evidence of fraud or intentional misrepresentation of the data," and retracted because "no definitive conclusions can be reached with this small sample size" [F4]. That reading is asserted as a claim about the captured document and survived a verbatim attribution test against the raw bytes [C8].

The trail continues: Environmental Sciences Europe republished the study in 2014, framing republication as transparency rather than endorsement [A7, F3, C6]. And here the machinery shows its edge — the republication's own Crossref record carries no relation linking it back to the retracted original [C7]. Check only the 2014 DOI and the paper looks pristine. That asymmetry is the whole argument for scoped absence claims in one example.

What was ruled out, and what this can't say

The Retraction Watch Database was not queried directly: its search is an interactive form with no public mechanical per-DOI route from this deployment, and we don't screen-scrape and call it capture. We relied instead on Crossref's ingestion of the same data — visible in custody in A5's updated-by rows — plus publisher pages and PubMed (recorded as a pass, with the substitute surfaces named).

The strongest attack on this notebook is that Crossref's update metadata is incomplete — publishers deposit unevenly, and pre-DOI-era events may never have been backfilled. That is why no null here claims the world, only named surfaces; why the publisher's own page backs every Crossref silence; and why the oldest, thinnest record in the set (Shannon) carries the caveat in the claim text itself. Whether the ESE republication was freshly peer reviewed remains open [Q6], with the surfaces that could settle it named on the question page.

Four classics stand unmarked, on the record of the surfaces named, as of 2026-08-13. One classic carries the full machinery of modern correction: retracted, documented twice over, republished, and — from the republication's side — invisibly so. Both halves of that sentence are findings.