How journal editors raise REFerence [IN]tegrity before publication

Peer review is thorough about the science and silent about the references. Reviewers assess method, evidence and argument, because that is what they were asked to do and what their expertise is for. Almost none of them opens a reference list and checks that every entry exists and that every citation points where it claims.

Nor should they have to. It is not a scholarly judgement — it is a lookup task, and asking a subject expert to spend an evening on it is a poor use of the scarcest resource a journal has.

The result is a gap that sits between two roles rather than inside either of them.

Between the reviewer and the copy-editor

Author            assembles the list, often merging co-authors' libraries

Desk screen       scope, fit, formatting compliance, plagiarism screen

Reviewer          method, evidence, argument
                  — reads the references cited in the text, not the list

Copy-editor       house style, punctuation, capitalisation, order
                  — checks that entries are FORMATTED, not that they EXIST

Production        deposits the reference list as structured metadata

Nobody            checks that each citation resolves to the entry it
                  names, and that each entry is a real record

Every role in that chain is doing its job. The copy-editor is not being careless by formatting a reference to a paper that does not exist — verifying existence was never in the brief, and no house style guide asks for it. The gap is structural, not individual, which is also why exhortation has never closed it.

The reference list is the only part of a submission that no single role is responsible for verifying — and the only part that is machine-readable, permanent and cited onward.

Why this matters more for a journal than for a thesis

A broken reference in a thesis is a correction the candidate makes. A broken reference in a published paper is three other things as well.

It becomes the version of record. Once published, the fix is a corrigendum — a permanent, visible, editorially expensive artefact for something that would have cost a minute to catch at submission.

It is deposited as metadata. Production sends the reference list out as structured data. An identifier that points at the wrong article does not merely sit there looking wrong; it enters the citation graph as an assertion that this paper cites that one.

In the paper     "…consistent with earlier work on ward staffing¹²"

Reference 12      Okonkwo & Ferreira (2022), Workflow variance in
                  regional centres
DOI deposited     10.xxxx/jhs.2019.0447   ← belongs to a different paper

Downstream        the citation is credited to the WRONG article
                  indexes, metrics and reference managers all inherit it
                  a reader following the link lands somewhere else
                  the actual source receives no credit at all

And it is inherited. Readers cite what you published. A reference copied out of your paper into three others carries the same defect, and by then the error has a provenance that looks like corroboration.

What changed in the last two years

Two things, pulling against each other. Generative tools now produce reference lists that are fluent, correctly formatted and sometimes partly fictional — a problem we have written about from the author's side, and which arrives at your desk as a submission that looks impeccable. And the obvious countermeasure, asking a model to verify the list, answers a different question than the one you need: whether a reference is plausible, not whether it exists.

The practical consequence for an editor is that formatting quality has stopped being a proxy for care. A beautifully formatted list used to suggest an author who had been diligent. It now suggests a tool, and tells you nothing about whether the entries are real.

Where a check belongs: the desk screen

Not at proof stage, where fixing anything is expensive and slow. At the desk screen, alongside the scope check and the plagiarism screen, before a reviewer's time is committed.

Three reasons it belongs there. It is cheap — the same automated pass you already run for plagiarism, applied to a different property. It protects reviewers, because a submission whose reference list does not resolve is one you may reasonably return before spending two experts on it. And it is fixable by the author at no cost to anyone, which is not true at any later stage.

What an editor gets back is a short, specific list: citations with no matching entry, entries never cited, duplicates from a co-author merge, and identifiers that resolve to something other than the entry they sit in. It is the same arithmetic we describe for a single document — and one or two findings is the normal result, not a red flag.

And again at acceptance

A second, cheaper pass at acceptance catches the class of defect that appears without anyone touching the manuscript. Papers sit in review for months; during that time preprints get published, hosts move, and occasionally a cited paper is retracted. The list that was correct at submission is not necessarily the list you are about to deposit.

The retraction case is the one worth the trouble on its own. A published paper citing a retracted result without acknowledging the retraction is a correction waiting to be requested, and the notice usually appears while the manuscript is with you rather than before it arrived. There is a fuller account of what ageing does to a reference list, which applies exactly to a review cycle.

What this is not

Most of what a reference screen surfaces is ordinary assembly damage: a duplicate created when two co-authors merged libraries, a citation added in response to a reviewer and never carried into the list, a DOI copied from a manager record that was stale before the paper was drafted. Reporting those to an author as suspicion would produce defensive authors and nothing else.

One finding is not in that category. A reference in a submission that resolves to no record anywhere is not assembly damage, and an editor knows what it may indicate — fabricated citation is academic misconduct, and it is the failure most likely to become a retraction rather than a correction. A desk screen does not adjudicate that. It surfaces it early, while it is still a question you can put to an author, rather than after publication when it is a question someone else puts to you.

It is also not a judgement about the quality of the scholarship. A paper can have an immaculate reference list and be wrong, and it can have three merge duplicates and be excellent. The check answers one narrow question — do the references resolve — and should be reported to authors in exactly those terms.

The value to a journal is not that it catches bad actors. It is that the part of the published record which is permanent, machine-readable and inherited by everyone who cites you gets verified once, by something built for the task, before it becomes the version of record. There is an annual measurement of how often that verification does not happen, and the shape of it is not reassuring.