The science behind
the service.

Submedit’s editorial protocols are grounded in a formal evidentiary framework developed from the published literature on citation error, verification methodology, and the structural limitations of peer review. This page describes that framework and the research that informs it.

What makes a citation valid?

The question of citation validity is more demanding than it first appears. A citation is not valid merely because the referenced paper is topically related to the claim, or because it contains findings in the same domain. A citation is valid when, and only when, the referenced source directly and explicitly supports the exact claim at the point of citation.

This distinction is the foundation of Citation Integrity™ verification. Conceptual proximity, causal association, or thematic overlap — however reasonable they may seem to the author — does not constitute citation validity in the sense that a careful peer reviewer applies the term.

A citation is valid when the referenced source directly and explicitly supports the exact claim at the point of citation. Conceptual proximity is not sufficient.

The Equivocation Error

The most frequently identified category of citation invalidity in our review workflow is the equivocation error: the citation of evidence established for one construct to support a claim about a related but distinct construct. This error is systematic and ubiquitous. It arises not from negligence but from a structural feature of how scientific literature is read and cited.

The mechanism is straightforward. An author reads a paper establishing finding X. Later, when writing a paper about finding Y — a related but not identical phenomenon — the author cites the X paper as support for a claim about Y, reasoning that the findings are close enough to be equivalent. A peer reviewer with deep familiarity with both constructs will recognise this as an unsupported inference. The citation does not do what it claims to do.

The published literature documents this category of error at scale. Independent meta-analyses across 32,074 references establish that approximately 1 in 6 citations contains a significant error,^1,2,3 with 38% of those errors representing the systematic misrepresentation of what the source paper actually demonstrates — not simple misquotation.⁵

Primary vs. Secondary Citation Integrity

A second category of citation invalidity concerns the distinction between primary and secondary support. A citation is primary when the referenced paper presents the finding as its own original result. A citation is secondary when the author cites a paper that merely relays a conclusion from a third paper — reproducing it without being its original source.

Secondary citations are not inherently invalid, but they are systematically more fragile. The intermediate paper may have mischaracterised the original finding. The original finding may have been superseded. The cited secondary source may itself contain a citation error in its relay of the primary result. Each hop in the chain introduces a new point of potential distortion.

Simkin and Roychowdhury (2003) estimated, through statistical analysis of copying errors in bibliographies, that approximately 80% of scientific authors do not read the papers they cite — they copy citations from the reference lists of other papers.⁴ This finding has not been substantially challenged in the two decades since. It is the mechanistic explanation for why citation error rates have not improved since they were first systematically documented in 1984 — despite decades of awareness.

In our verification workflow, every citation is assessed against both criteria: (1) does the source directly support the specific claim, and (2) is the support primary — the original finding — rather than a relay of a third paper’s conclusion?

Forward and reverse analysis.

Our citation verification protocol employs independent forward and reverse analysis for each citation–claim pair. This dual-direction approach addresses a structural limitation in conventional single-direction reading.

The combination of both directions provides an independent check on each direction’s conclusion. A citation that passes forward analysis but fails reverse analysis — where the source paper reaches a different conclusion from what the manuscript claims — is a more severe validity failure than a simple mischaracterisation: it represents the systematic misrepresentation category documented by Pavlovic et al. (2021) as constituting 38% of all citation errors.⁵

Every discrepancy identified is returned to the author with a precise annotation specifying: (1) the nature of the validity failure (equivocation, contradiction, secondary relay, or mischaracterisation); (2) the specific passage in the source paper that is relevant; and (3) the correction required — whether revision of the claim, replacement of the reference, or explicit acknowledgment of the inferential step.

Cross-validation across every data surface.

Citation errors are the most documented category of pre-submission manuscript failure, but they are not the only category that triggers reviewer doubt. Numerical inconsistency — discrepancies between values stated in the body text and those appearing in figures, tables, or supplementary materials — is among the most common causes of reviewer concern at high-scrutiny journals.

The mechanism is straightforward: during the revision process, values are updated in one location but not propagated consistently across all locations where they appear. A p-value revised in the body text may remain unchanged in the corresponding table. A percentage stated in the Discussion may not match the figure caption from which it was derived. A sample size mentioned in the Methods may conflict with the n reported in a supplementary table.

A numerical inconsistency identified in the Methods triggers a reviewer to look more carefully. If they find a second in the Discussion and a third between the text and a figure, the manuscript’s evidentiary reliability has been placed in doubt before the scientific evaluation has begun.

Our numerical consistency protocol systematically cross-validates all quantitative claims in the body text against every corresponding data surface in the manuscript: figures, figure captions, tables, table notes, and supplementary files. The scope of validation covers reported statistics (p-values, effect sizes, confidence intervals), sample sizes and subject counts, percentage calculations, and any numerical values that appear in multiple locations.

Each discrepancy is annotated with the specific locations of the conflicting values and the correction required. Where the discrepancy reflects an update not propagated across all instances, the correction is flagged for authorial resolution — we do not correct data; we ensure the author is aware of every inconsistency before submission.

Solving the architecture problem in human review.

The structural limitation of conventional manuscript review is not a quality problem — it is an architecture problem. A single editor reading a complex paper sequentially encounters an irreducible cognitive constraint: as the manuscript grows longer, the attention available for later sections is not equal to the attention applied to earlier ones. A language editor scrutinising sentences on page 3 has finite cognitive resources remaining for verifying citations on page 28.

This limitation cannot be resolved by hiring more experienced editors. It is inherent to the sequential, unaugmented reading of a long document. The AI component of Submedit’s workflow is designed specifically to address this constraint — not to replace human expert judgment, but to restructure the problem the human expert is solving.

This is the architecture of orchestrated AI-human review: the exhaustive, consistent scanning of AI agents combined with the irreplaceable judgment of domain experts who understand what the findings mean, why the citations matter, and what the reviewers at the target journal will evaluate. Neither achieves this standard alone.

Importantly: no client manuscript is used in the training, fine-tuning, or evaluation of any AI model. The AI component of Submedit’s workflow is a proprietary tool applied exclusively within the editorial process. This commitment is binding under our Terms of Service and Privacy Policy.