Editorial decision-making and desk rejection

Last updated: 2026-10-20

Synthesis

The peer-review process most authors imagine — careful reading by two or three experts, weighed by an editor against a clear standard — is not what most submissions experience. At high-impact journals, the modal outcome is desk rejection: an editorial-stage decision made within days or weeks, without the paper being sent for full peer review. The journals that publish their own figures make this concrete. In 2025 Science sent 15.0% of its research submissions for external review and accepted 5.3%, with a median of 8 days to first decision [9]. Nature states that about 8% of submissions are accepted and that “most submissions are declined without being sent out for peer review” [4]. NEJM reports publishing about 5% of original research submissions [10]. Comparable primary figures for mid-tier and mega-journals are scarce. The decision is made by an editor or associate editor reading the cover letter and abstract, sometimes glancing at the figures, and applying a small number of triage heuristics: scope fit, methodological adequacy detectable from the abstract, novelty, and language quality. “Not peer reviewed” does not always mean editor-only triage, though. At Science, editors summarily rejected 28.1% of submissions in a five-year dataset and sent 69.5% to the Board of Reviewing Editors (BoRE), outside experts who give rapid advice; editors then declined 75% of those after BoRE advice [12]. So most Science submissions that never reach full peer review are still declined with outside scientific input. A large 2026 analysis of editorial triage (110,303 submissions to Science and Science Advances) also finds selection associated with factors the manuscript cannot change — institutional prestige, team size, topic and country — and these associations are stronger at the editorial stage than at peer review [12].

For scriptorium, this matters because the value proposition includes catching what would trigger desk rejection before submission. That is a concrete, testable claim — much sharper than “improving the manuscript” — and it is the most asymmetric intervention an agentic editor can offer: a single weekend of pre-submission audit can save months of round-trip latency. Operationalising it means building (or budgeting in reviewer-simulation) a desk-rejection risk audit that probes a small number of high-leverage triage signals.

Evidence and frameworks

How editorial decisions actually get made

Smith’s 2006 JRSM essay [2] is the most-cited polemic on peer review’s empirical weaknesses, and the most useful concise account of the editorial side. Smith — then editor of BMJ for 13 years — argues that peer review is “a flawed process at the heart of science and journals”, that the evidence base for peer review’s effectiveness is thin, and that editors retain substantial weighting authority even when reviews are formally the basis of decision.

Bornmann’s broader review of peer-review research [1] surveys studies of inter-reviewer agreement (consistently low), editor weighting of reviewer recommendations (variable), and decision predictors. His meta-analysis with Mutz and Daniel pooled 70 reliability coefficients from 48 studies of journal peer review and found low inter-rater reliability (mean ICC/r² = .34, mean Cohen’s kappa = .17) [8]. The bottom line: reviewer recommendations are noisy inputs to an editorial decision that is itself partly judgement.

Desk-rejection rates and triage criteria

Quantitative data is uneven. Most journals do not publish how many submissions they decline before review; the figures below are the ones journals disclose themselves:

• Science (2025, research articles, reports and research resources): 15,088 submissions; 15.0% externally reviewed; 5.3% accepted; median 8 days from submission to first decision [9]. In other words, about 85% of submissions were not sent to full peer review. That is not the same as editor-only triage. In Zhang et al.’s five-year Science data, editors summarily rejected 28.1% of 68,047 submissions and sent 69.5% to the Board of Reviewing Editors, then declined 75% of those after BoRE ratings and advice [12]. (The 2025 figures could only be read through trade-press and search-index copies of the Science page; see [9].)
• Nature: “only about 8% of submitted manuscripts will be accepted for publication. Most submissions are declined without being sent out for peer review.” Authors are “usually informed within a week” if the paper is not being considered [4].
• NEJM: about 5% of original research submissions are published [10]. NEJM does not disaggregate desk rejection from later rejection in this statement.
• Mid-tier subject journals and open-access mega-journals (PLOS ONE, Scientific Reports) are widely described as desk-rejecting far less, with most filtering at peer review. We did not find primary, journal-disclosed figures for them, so this note no longer gives percentages for these tiers.

Nature also publishes the criteria its editors apply before review: results that “seem novel, arresting (illuminating, unexpected or surprising)”, with “both immediate and far-reaching implications”. It states that “the judgement about which papers will interest a broad readership is made by Nature’s editors, not its referees”, and that the initial decision “is not a reflection on the technical validity of the work” [4]. This is direct evidence that significance calibration is an editor-stage gate, separate from technical soundness.

The dominant triage criteria reported by editors:

1. Scope mismatch. The paper is competent but wrong for the venue (basic-science result at a clinical journal; method paper at an applications journal).
2. Insufficient novelty or significance. Detected from the abstract — “we extend X to a new dataset” rarely survives triage at a flagship journal.
3. Methodological inadequacy detectable from the abstract. Small N, no controls, missing reporting-guideline elements (reporting-guidelines) flagged at a glance.
4. Language quality. Severe ESL issues (see esl-writers-swales-hyland) are read by editors as a competence signal even when the science is sound — an uncomfortable reality that drives the EAL editing market.
5. Cover letter and significance positioning. A cover letter that fails to articulate why the work matters here, now, for this venue is itself a desk-reject trigger; see significance-positioning and nih-significance-patterns.

Who gets past the desk: selection beyond the manuscript

Zhang et al. (2026) released deidentified data on 110,303 submissions over five years at Science and Science Advances and analysed the editorial-review and peer-review stages separately [12]. They report “strong selective effects associated with higher institutional prestige, larger team size, and certain topics and countries”; a small but significant advantage for corresponding authors who are men at Science; and a significant disadvantage for authors based in China. These associations “are generally stronger in editorial review than in peer review, even as final editorial decisions correlate strongly with reviewer advice” [12].

Two consequences follow (researcher inference). First, desk-stage outcomes are not a pure function of manuscript quality, so a pre-submission audit can lower risk but cannot remove it. Second, the editorial stage is where most of the non-manuscript variance enters, which strengthens the case for treating desk triage as its own problem rather than as early peer review.

Editors and associate editors

The “editor’s introduction” papers in journals like eLife and Annual Reviews describe the assignment-to-AE step: a handling editor reads the abstract, identifies an associate editor (AE) with topical fit, and either triages or hands off. AE selection is itself a high-leverage decision — AEs choose reviewers, weight their reports, and write the decision letter.

NIH study sections are the grant-review analogue. A small number of assigned reviewers (usually three) review each application in detail and report to the panel; panel members’ individual scores are anchored to the scores those assigned reviewers declare publicly [7]. In constructed NIH-style panels, discussion made reviewers within a panel agree more but made different panels agree less, and “score calibration talk” played a pivotal role in that variability [11]. The parallel matters: in both peer review and grant review, the person who introduces the work to the deciding body has disproportionate influence on outcome (researcher inference from [7, 11]; the journal-AE side of the analogy is not directly measured). This is why reviewer-archetypes-grants applies as much to journal AEs as to grant reviewers — the AE is functionally the first reviewer.

Open / transparent decision-making

Several journals have experimented with transparency to test whether visibility changes review quality or outcome:

• eLife: from 2023 eLife stopped making accept/reject decisions after peer review. Papers it sends for review are published as Reviewed Preprints with public reviews and an eLife Assessment [13]. An earlier opt-in trial (submissions in 2018, final outcomes reported in 2019) let 313 authors choose a process in which eLife committed to publish any paper it sent for in-depth review; 20.8% of trial papers were accepted against 14.9% in the regular process, and median time to acceptance was 175 days against 138 [6]. So the trial did raise the final acceptance rate, even though more trial papers were rejected before review. The model has had an indexing cost: Web of Science put eLife on hold in October 2024, and eLife now provides only a partial feed (papers assessed as “solid” or better), which makes it ineligible for a Journal Impact Factor [15].
• BMJ: publishes signed reviews and the full prepublication history. In the 1999 BMJ randomised trial [3], asking reviewers to have their identity revealed to authors had no significant effect on review quality, the publication recommendation, or the time taken to review, but it made reviewers more likely to decline to review (35% v 23%).
• Nature: for manuscripts submitted from 16 June 2025, every paper published as a research article carries its peer-review file (reviewer reports and author responses). Reviewers stay anonymous unless they choose to be named. This replaced the opt-in transparent review Nature had offered since 2020, so the share of Nature papers with a public review file grows over time rather than switching over at once [14].
• PLOS: opt-in publication of peer-review history; modest uptake.

The pattern: making review visible has not, on the trial evidence, changed review quality much. The constraint on review quality is time and expertise, not opacity. Transparency now reaches the author side too: under Nature’s 2025 policy, reviewer reports on newly submitted papers that are accepted become public. Changing the decision rule itself, as eLife did, changes outcomes and has consequences for indexing.

Why this matters for scriptorium

The value proposition implicit in “reviewer-simulation as a sparring partner before submission” is two distinct services:

1. Reviewer-style critique — simulated reviewer reports of the kind the work would receive if it survives triage.
2. Desk-rejection risk audit — a pre-triage check that flags the small number of high-leverage signals editors actually use to triage.

These are different products. The second is potentially much more asymmetric: catching a scope mismatch before submitting to Nature saves a desk-rejection cycle (Nature says authors are “usually informed within a week” [4]) plus the reformatting and fresh queue at the next venue; catching a reviewer-style critique saves perhaps 2 weeks on a second round. (Corrected in the 2026-10 audit: the earlier “6+ weeks of round-trip” figure was unsourced and conflicted with Nature’s own turnaround statement.) The desk-rejection audit should be cheap, fast, and focused on a small triage checklist — not a comprehensive review.

How this informs scriptorium

1. Build a desk-rejection-risk audit, either as its own v0.2 skill or as a flagged output mode of reviewer-simulation. The checklist is short:

   • Does the abstract (and cover letter, if present) articulate a finding that fits the target venue’s scope?
   • Are reporting-guideline elements (reporting-guidelines) present in the abstract for the relevant study type?
   • Are language-quality issues likely to trigger an editor’s “send back for English editing” reflex?
   • Does the significance positioning (significance-positioning) match the venue’s calibration (clinical relevance for NEJM; mechanistic insight for Nature; method novelty for ML venues)?
   • Are any policy-compliance issues likely to be caught at triage (ICMJE authorship; AI-disclosure; data-sharing; ethics statements)? The ICMJE Recommendations ask authors to describe AI use in both the cover letter and the manuscript (since the May 2023 update), and, when AI was used to conduct the study, to give the tool, version and prompts in the Methods (added in January 2024 [16]). The January 2026 update consolidated the AI guidance into a top-level Section V covering authors, reviewers and editors [5]. These are established requirements, not new 2026 ones.

2. Use MANUSCRIPT_STATE.target_venue aggressively. A desk-rejection-risk audit is incoherent without a target. Skills should refuse to run (or run with a warning) if target_venue is empty.

3. Honest reporting. The skill should distinguish desk-rejection signals (editorial triage heuristics) from reviewer concerns (peer-review-stage critique). These map to different stages of the decision pipeline and should be reported separately. It should also say that desk outcomes carry selection the manuscript cannot change (institution, team size, country, topic) [12], so a clean audit lowers risk rather than guaranteeing review.

4. Conservative-edit posture applies. The audit reports risk signals; it does not rewrite the abstract or cover letter unsolicited. If the user then asks for a rewrite, that’s a separate transformation skill invocation.

Implementation priority for scriptorium

Verdict: Yes — v0.2 skill desk-rejection-risk (or explicit output mode of reviewer-simulation).

Status (2026-10-20 refresh): the skill now exists as skills/desk-rejection-risk/. The verdict holds, and the refreshed evidence strengthens it: journal-disclosed figures confirm that most submissions to the top general journals are not sent to full peer review [4, 9], and the editorial stage is where selection is strongest [12]. Three things in the skill’s grounding prose now need to change. First, the rate figures it quotes from the previous version of this note (Cell/Lancet figures; “30–60%” for mid-tier journals) are no longer supported here. Science’s 85% figure [9] should not be relied on until someone reads science.org directly, because the only copies available to this refresh were trade-press and search-index copies. Second, at Science that 85% includes a Board of Reviewing Editors consultation for most papers [12], so the skill should not describe the gate it simulates as an editor skimming the abstract alone. Third, its policy-compliance category should check the ICMJE AI-disclosure requirements: cover letter plus manuscript (May 2023), and tool, version and prompts in the Methods (January 2024 [16]), now consolidated in Section V (January 2026) [5]. The skill should present these as established ICMJE requirements, not new 2026 rules. Tracked for the skills pass, not edited here.

Skill spec:

• Name: desk-rejection-risk (preferred) or reviewer-simulation with a --mode=desk-triage flag.
• Phase: v0.2 (next, alongside the first orchestrator).
• Scope: read MANUSCRIPT_STATE.yaml, the abstract, the cover letter (if present), and the figure captions. Do not read the full body. Emit a structured triage report with a small number of high-signal risk flags.
• Required data:
  • project.target_venue (mandatory; refuse to run without).
  • project.target_type (informs which checklist applies).
  • core_claims (for scope-fit assessment).
  • known_weaknesses (so triage doesn’t flag what authors already acknowledge).
  • contributors: (if implemented per credit-taxonomy-authorship) for policy-compliance checks.
  • Venue-specific scope/significance rules from knowledge/grants/ and knowledge/scientific-writing/.
• Output: a structured markdown report with sections scope_fit, methodological_signals, significance_calibration, policy_compliance, language_quality, each with severity flags (high, medium, low, none) and one-paragraph rationale.

Why v0.2, not v0.1: the v0.1 release prioritises proving the skill composition pattern works. desk-rejection-risk is most useful once reviewer-simulation is mature enough to share infrastructure (venue calibration data, archetype prompts), which v0.2 is the natural moment for.

Operational note: this is the skill that turns scriptorium’s value proposition from “a useful editing tool” into “a weekend-of-time-for-months-of-round-trip-latency” trade. Sean should treat it as load-bearing for the project’s external narrative.

Open questions / weak evidence

• Desk-rejection rates are reported variably; many journals do not disaggregate triage from peer-review rejection in their public statistics. Science is the exception, publishing its external-review rate [9]. The earlier “70–80%” figures for Nature/Cell and “~90%” for NEJM/Lancet came from trade-press compilations and have been dropped. Science’s 85% not-reviewed figure and Nature’s “most submissions” statement are the defensible replacements; we found no primary figure for Cell or The Lancet. The Science figure is stated by the journal but was not read on the journal’s own page in this refresh (see [9]), and it counts Board of Reviewing Editors declines as well as summary rejections [12].
• Field-to-field heterogeneity is large: ML conference venues don’t really have “desk rejection” — every submission goes to review under tight deadlines. The audit needs discipline awareness (discipline-conventions).
• Editor weighting of reviewer recommendations is poorly characterised in the literature. The Bornmann review [1] summarises what little is known. The Science/Science Advances data [12] are the first large release that separates editorial and peer-review stages, but they come from two journals in one publisher family, and one author is a deputy editor at Science Advances (declared in the paper).
• The study-section analogy now rests on peer-reviewed work [7, 11], but that work uses constructed NIH-style panels and a small number of reviewers. No equivalent study measures how much a journal AE’s first framing drives the outcome.
• The PLOS “modest uptake” statement above is unsourced.

References

[1] Bornmann, L. (2011). Scientific peer review. Annual Review of Information Science and Technology, 45(1), 197–245. DOI: 10.1002/aris.2011.1440450112. (The canonical Bornmann review of peer-review research. The previous entry pointed at DOI 10.3102/0034654309338847, which does not resolve.)

[2] Smith, R. (2006). Peer review: a flawed process at the heart of science and journals. Journal of the Royal Society of Medicine, 99(4), 178–182. DOI: 10.1177/014107680609900414. PMID: 16574968.

[3] van Rooyen, S., Godlee, F., Evans, S., Black, N., & Smith, R. (1999). Effect of open peer review on quality of reviews and on reviewers’ recommendations: a randomised trial. BMJ, 318(7175), 23–27. DOI: 10.1136/bmj.318.7175.23. PMID: 9872878. (Foundational BMJ open-review randomised trial.)

[4] Nature. (Ongoing.) Editorial criteria and processes. https://www.nature.com/nature/for-authors/editorial-criteria-and-processes. (Primary source for Nature’s ~8% acceptance figure, the statement that most submissions are declined without review, and the criteria for sending a paper out. Replaces the earlier trade-press compilation.)

[5] International Committee of Medical Journal Editors. ICMJE Recommendations (updated January 2026), including the new Section V, “Use of Artificial Intelligence in Publishing”. https://www.icmje.org/recommendations/. (Section V consolidates AI guidance that ICMJE first added in May 2023 (AI not an author; disclosure in the cover letter and the submitted work) and extended in January 2024 (tool, version and prompts; see [16]). icmje.org blocked automated fetching during the 2026-10 refresh and audit, so the May 2023 dating rests on search-index copies of ICMJE’s May 2023 update and annotated PDF.)

[6] eLife. (2019). Peer review: final results from a trial at eLife. https://elifesciences.org/inside-elife/e8fa8f7a/peer-review-final-results-from-a-trial-at-elife. (Final outcomes of eLife’s 2018 opt-in trial in which eLife committed to publish any paper sent for in-depth review, using a cut-off of one year after the trial closed. Not an evaluation of published decision letters, as the earlier description said.)

[7] Pier, E. L., Raclaw, J., Carnes, M., Ford, C. E., & Kaatz, A. (2019). Laughter and the chair: social pressures influencing scoring during grant peer review meetings. Journal of General Internal Medicine, 34(4), 513–514. DOI: 10.1007/s11606-018-4751-9. PMID: 30604119. (Replaces the earlier pointer to NIH OER blog posts. States that panelists’ scores are anchored to those declared by the assigned reviewers.)

[8] Bornmann, L., Mutz, R., & Daniel, H.-D. (2010). A reliability-generalization study of journal peer reviews: a multilevel meta-analysis of inter-rater reliability and its determinants. PLoS ONE, 5(12), e14331. DOI: 10.1371/journal.pone.0014331. PMID: 21179459.

[9] Science / AAAS. (Updated semi-annually; 2025 data.) Journal metrics. https://www.science.org/content/page/journal-metrics. (Submissions, external-review rate, acceptance rate and median turnaround for the Science family of journals. The page blocked automated fetching during the 2026-10 refresh, its independent re-check and the evidence audit. The only copies of the 2025 figures found were trade-press copies (Manusights blog posts) and search-index summaries, the same class of source this refresh dropped for the older desk-rejection rates. Treat the 85% not-reviewed figure as stated by the journal but unverified here; it needs a manual read of science.org before the skills pass relies on it.)

[10] Massachusetts Medical Society. (Ongoing.) About NEJM. https://www.nejm.org/about-nejm/about-nejm. (“About 5% of original research submissions achieve publication by NEJM.” The page blocked automated fetching during the 2026-10 refresh, so the quotation was checked through a search index’s copy of the page.)

[11] Pier, E. L., Raclaw, J., Kaatz, A., Brauer, M., Carnes, M., Nathan, M. J., & Ford, C. E. (2017). ‘Your comments are meaner than your score’: score calibration talk influences intra- and inter-panel variability during scientific grant peer review. Research Evaluation, 26(1), 1–14. DOI: 10.1093/reseval/rvw025. PMID: 28458466.

[12] Zhang, S., LaBerge, N., McElhiney, Q., Larremore, D. B., & Clauset, A. (2026). Editorial and peer review dynamics at elite general science journals. Science Advances, 12(29), eaec0494. DOI: 10.1126/sciadv.aec0494. PMID: 42455887. (Competing interest declared in the paper: A. Clauset is a Deputy Editor at Science Advances. Source for the Science triage breakdown: “Editors at Science summarily rejected 28.1% of the 68,047 submissions … and sent 69.5% to external experts on the BoRE. Given BoRE ratings and advice, editors declined 75% of those manuscripts”. The BoRE is “a set of outside experts who would provide rapid advice to the editors on whether and to whom to send a manuscript for formal peer review”. Checked against the open-access full text, PMC13371883.)

[13] Eisen, M. B., Akhmanova, A., Behrens, T. E., Diedrichsen, J., Harper, D. M., Iordanova, M. D., Weigel, D., & Zaidi, M. (2022). Peer review without gatekeeping. eLife, 11, e83889. DOI: 10.7554/eLife.83889. PMID: 36263932.

[14] Nature. (2025). Transparent peer review to be extended to all of Nature’s research papers. Nature, 642(8068), 542. DOI: 10.1038/d41586-025-01880-9. See also the Springer Nature press release of 19 June 2025: https://www.springernature.com/gp/group/media/press-releases/transparent-peer-review-now-standard-for-nature/27788498.

[15] eLife. (2024, addendum 2025). Changes to eLife’s indexing status in Web of Science and Scopus. https://elifesciences.org/inside-elife/ae620829/changes-to-elife-s-indexing-status-in-web-of-science-and-scopus.

[16] Ochsner Journal. (2024, 25 January). Additional AI recommendations from the International Committee of Medical Journal Editors. Ochsner Journal blog. https://ochsnerjournal.wordpress.com/2024/01/25/additional-ai-recommendations-from-the-international-committee-of-medical-journal-editors/. (Dated secondary record of the January 2024 ICMJE amendment, quoting it verbatim: “Authors who used AI technology to conduct the study should describe its use in the methods section in sufficient detail to enable replication of the approach, including the tool used, version, and prompts where applicable.” Cited because icmje.org blocked automated fetching.)