The 2026 manuscript findings report

The sample is every manuscript submitted to Manusights pre-submission review between 2026-03-23 and 2026-05-14, with a parsed top-risks payload. That is 5,495 manuscripts across 52 days of operation, covering authors who specified a target journal across more than 50 peer-reviewed venues plus a long tail of unspecified or niche targets.

Each manuscript is scored by the Manusights review engine, which classifies the issues it identifies into five categories: claim_support, science_core, journal_fit, structural_hygiene, and confidence. Severity is critical, major, or minor. Earlier engine versions used a different schema (severity high or medium, no issue class). To avoid mixing schemas, all sectioned-by-class analysis in this report uses only the 835 manuscripts evaluated under the v4 classification engine. Site-wide counts use the full 5,495.

The engine as instrument. The v4 classification engine is a Manusights-internal system that combines language-model-based issue detection with a rule-layer for severity assignment. The five issue classes were defined a priori from our editorial taxonomy, refined through iterative review of engine outputs against human-reviewer judgments. We have not yet published a separate validation study of the engine's false-positive rate or inter-rater reliability against human peer reviewers. That work is planned for a follow-up methodology paper. Treat the findings in this report as engine outputs, not human-confirmed determinations.

Target journal is self-reported by the author at upload. It is present on roughly 35% of jobs across the full 5,495 sample (and roughly 30% of the v4 classified subset of 835). Per-journal subsections use the classified subset that also has a stated target. We do not infer target journal from manuscript content.

One thing this sample is not: a random sample of all manuscripts being written in 2026. The authors here chose to run a pre-submission review, which selects for some baseline awareness that the manuscript might benefit from one. Read every percentage in this report as “of manuscripts whose authors chose to stress-test them before submitting,” not “of all manuscripts in the field.” We have not yet estimated the magnitude of this selection bias against a comparison group of un-reviewed manuscripts. A validation against a random-sample holdout is in scope for the August 2026 refresh.

Public context this report does not replace. Several of the journals discussed publish their own editorial statistics, which provide useful triangulation for any reading of this report:

• Nature reports a desk-rejection rate around 60% and overall acceptance around 8% (Nature editorial statements and Nature Portfolio transparency disclosures).
• Science (AAAS) publishes an overall acceptance rate under 7%.
• Physical Review Letters publicly describes a desk-rejection rate near 60% within its four-page-limit constraint (APS editorial statements).
• PLOS ONE reports a roughly 50% acceptance rate consistent with its lower-bar scope policy.

None of those public rates can be tied directly to our engine's flag rates without outcome data on individual manuscripts. The public rates are the editorial-outcome context; this report is about what an external pre-submission review surfaces before any editorial decision. We discuss the relationship hypothetically in each per-journal subsection and never as a causal claim.

Across the 835 classified manuscripts, the engine flagged 1,850 issues. Five categories account for 99.8% of them. claim_support dominates: it is more common than the next three combined.

claim_support (60.6% of issues, 82.9% of manuscripts). The single most common pre-submission risk. claim_support is the engine's category for "the data shown does not fully back the claim made." It captures overstatement (the abstract promises more than the methods deliver), undersupported conclusions (a result is asserted without the supporting cell, table, or figure), and missing-control patterns. It is not a writing problem. It is a logic-of-evidence problem.

science_core (20.8% of issues, 38.3% of manuscripts). Issues with the methodology, validation set, or core scientific argument. This is where reviewers will get to in week 2 of peer review. Things like inadequate controls, sampling issues, or a benchmark that doesn't actually test the claim. When science_core trips, it usually means a longer revision cycle, not a quick fix.

journal_fit (10.3% of issues, 22.6% of manuscripts). The work is fine; it is in the wrong venue. journal_fit is the engine's flag for scope mismatch: a methods-paper at a broad-impact journal that wants a finding, a regional case study at a journal that wants global generalizability, a narrow chemistry result at a journal that wants cross-field implication. Fit issues are disproportionately critical (79% of journal_fit risks are critical severity), because once fit is wrong, no amount of revision fixes it. You change the target.

structural_hygiene (5.5% of issues, 11.4% of manuscripts). Formatting, section ordering, missing required statements, statement-of-significance length, citation format, table layout. Notably small. Modern desk rejection is not driven by formatting. The widespread belief that "good formatting prevents desk rejection" is wrong in this sample.

confidence (2.6% of issues). The opposite of overclaim: places where the manuscript hedges past the point of asserting anything. A surprising number of strong results are obscured by the authors' own hedging. Less common than the others, but each instance is high-leverage because the fix is usually one or two sentences.

Of the 835 v4-classified manuscripts, nine peer-reviewed journals each had at least 10 manuscripts in the classified subset and a stated target. Together they account for 219 of the 835. We exclude long-tail and non-journal targets (e.g., conferences, theses, memoir-style submissions) from this section because their findings patterns are not directly comparable to peer-reviewed journal submissions.

Read this table with caution. Rows with N<20 (Science, Nature, NeurIPS, IEEE Access, Environmental Health Perspectives, PLOS ONE) should be treated as directional only. A 90% rate on N=10 has a 95% binomial confidence interval of roughly 55-99%. We mark these rows with an asterisk and discuss them as hypotheses, not estimates.

A few patterns are worth pulling out before the journal-specific deep dives. Energy Research & Social Science is the surprise at the top of the table: 65 classified manuscripts targeting it, more than Nature Communications. The dominant issue is claim_support, with science_core close behind. The interpretation: ER&SS authors push policy-relevant claims that the evidence base is often not yet large enough to anchor, and the engine catches the gap.

Nature, Science, and Nature Communications cluster together at the top-tier flagship venues but their failure modes are different. Nature has a structural_hygiene component (6 occurrences, 4 critical) that none of the others share at that level, because Nature has tightly-specified format and length requirements. Science skews hard toward claim_support criticality. Nature Communications skews toward critical journal_fit issues, which we discuss separately below.

Physical Review Letters shows a cleaner pattern in the table: 18 of 43 claim_support issues are critical (42%). PRL has a four-page limit and a high evidence-density requirement; in our sample, claim-support issues at PRL co-occurred with critical severity more often than at the all-sample baseline (35%). The next-highest claim-support critical rate in the table is Science at 52% (N=19, small cell).

This section interprets a small-N cell: Science N=19, claim_support occurrences = 21. Treat the patterns described here as hypotheses, not estimates. Public Science acceptance rate is reportedly under 7% per AAAS editorial statements, with which our descriptive flags are not in conflict but also not yet causally connected.

In our sample, 52% of claim-support issues on Science-targeted manuscripts are critical severity (11 of 21 occurrences). The all-sample baseline is 35%. Science-targeted manuscripts in this small-N cell co-occur with higher critical-claim-support rates than the rest of the corpus.

One reading of this pattern, consistent with Science's publicly stated editorial bar: Science triages on whether a single result is general enough, surprising enough, and complete enough to merit a one-paper publication in a venue that publishes very few. Gaps between headline claim and figure-level evidence become more load-bearing at venues that publish single-finding flagship papers than at high-volume archival venues. Our data is descriptive only, so we cannot test this mechanism directly.

A practical audit, valid at any venue but especially worth running for Science-targeted drafts: read every claim in the abstract and the first paragraph of the discussion against the corresponding figure. If the figure does not show the result the claim is making, in the exact strength the claim is making it, treat that as a high-leverage edit before submitting. We have no outcome data on whether tightening these claims improves acceptance rates.

Science is also the only journal in the top 10 where 5 of 5 journal_fit issues are critical. PLOS ONE shows the same 4-of-4 pattern at N=4. Both are small-cell observations and neither supports a stable rate estimate. The directional reading: when the engine flags a fit problem at any of these venues, the pattern across our (small) samples is that the flag is rarely a minor one.

This section interprets a small cell: Nature Communications N=37, of which journal_fit occurrences = 10. A 9-of-10 rate has a 95% binomial confidence interval of roughly 55-99%. The pattern is directional; treat it as a hypothesis rather than a stable rate estimate.

In our sample, 9 of 10 journal_fit issues on Nature Communications-targeted manuscripts are critical severity. The 90% directional rate is the highest in the table for any journal with N at least 10, on any issue class, but it sits on a small denominator.

What this might mean in practice: Nature Communications occupies a specific position in the Nature family. It is open-access, broad-scope, and explicitly designed to be more accepting than Nature itself (Nature Portfolio publicly reports Nature Communications as a higher-throughput venue than Nature), but it still wants results with significance across multiple subfields. The manuscripts in this small sample that received a journal_fit flag for Nature Communications usually fell into one of three patterns. First, the methods-paper case: the work would fit Nature Methods or a specialty methods journal better than a broad-impact venue. Second, the regional-finding case: a result that is valuable but specific to one population, dataset, or system, where the manuscript stretches to claim universality the data does not support. Third, the incremental case: a clean follow-up to a paper that arguably belongs in a more specialty journal.

When the engine flags a Nature Communications fit problem, the directional advice the data supports is to consider the redirect path rather than the revise path. Nature Methods for methods. A Nature Portfolio specialty title (Nature Energy, Nature Plants, Nature Microbiology) for field-specific findings. eLife or Cell Reports for clean incremental advances. We cannot estimate the relative outcome rates between “revise and resubmit at Nature Communications” vs “submit fresh at the sister journal” without outcome data; this advice reflects editorial reasoning, not measurement.

The broader implication: in our sample, fit issues at Nature Portfolio venues co-occur with critical severity more often than fit issues elsewhere. A pre-submission scope-fit check is one tool to surface those issues before submission rather than after editorial triage.

The 82.9% claim_support rate is not an indictment of the authors in this sample. It is what a thorough pre-submission review finds when it is looking. Almost every manuscript has at least one place where the claim is slightly ahead of the evidence, or the evidence is slightly off the claim. The question is whether you catch it before the editor does, or after.

Four checks the data supports:

1. Read the abstract back against the last figure. If the abstract makes a claim that is not visibly shown in the final figure, that gap is a claim_support issue waiting to be flagged. The fix is to either show it or stop claiming it.
2. If you are submitting to a Nature Portfolio journal, audit fit early. In our (small) Nature Communications sample, the directional pattern is that journal_fit flags arrive at critical severity. We have no outcome data, but the editorial logic of broad-scope Nature Portfolio venues is consistent with the pattern: revision rarely fixes a scope mismatch.
3. If you are targeting Science, Nature, or PRL, audit every claim against its supporting figure carefully. The critical-severity rate for claim_support at these venues ranged from 39% to 52% in our sample, above the 35% all-sample baseline. Sample sizes are small (N=19-27), so treat this as a directional signal: top-tier venues co-occur with stricter engine flags in our data.
4. Do not over-invest in formatting. structural_hygiene drove 5.5% of issues. Time spent making the references perfect is time not spent auditing the claims. Re-order.

The thing this dataset cannot tell you is whether your specific manuscript has any of these issues. A Manusights pre-submission scan uses the same engine that produced these numbers, on your draft, in about five minutes. The findings in this report are the population; the scan is the individual measurement.

Self-selected sample. The 5,495 manuscripts in this dataset belong to authors who tried Manusights pre-submission review. They likely skew toward authors who suspect their manuscript has an issue (otherwise they wouldn't use the service). The 82.9% figure should be read as "82.9% of manuscripts whose authors brought them in for pre-submission review had a claim-support issue we caught," not "82.9% of all 2026 manuscripts have claim-support problems."

No outcome data. This report describes what pre-submission review FINDS. We do not have systematic outcome data on which manuscripts were subsequently accepted, rejected, or revised. We make no causal claim that the issues we identify predict rejection.

Internal classification. The five-class taxonomy (claim_support, science_core, journal_fit, structural_hygiene, confidence) is Manusights' internal scheme, refined through the 5,495-manuscript review window covered in this report and prior pilot windows. It is not identical to the categories journal editors or peer reviewers use in their own decision frameworks. We have not yet published an inter-rater reliability study against external human reviewers.

Self-reported target journal. Approximately 35% of jobs included a stated target journal. Per-journal subsections use only that subset. We do not infer target journal from manuscript content.

Engine version mix. The full 5,495-manuscript sample includes records from earlier engine versions that did not classify issues by class. To avoid mixing schemas, the classified subsections of this report use only the 835 records evaluated under the v4 classification engine.

A persistent identifier (DOI via Zenodo) is being minted for the underlying dataset and will be added to this page before any formal academic citation. In the meantime, the canonical URL is stable and suitable for citation.

@misc{jia2026manuscriptfindings,
  author = {Jia, Erik},
  title  = {The 2026 Manuscript Findings Report: What Pre-Submission Review Catches},
  year   = {2026},
  month  = may,
  version = {1.0},
  publisher = {Manusights},
  howpublished = {\url{https://manusights.com/research/manuscript-findings-2026}}
}

Data availability. Aggregate, de-identified findings data backing every chart and table in this report is published at /research/data/manuscript-findings-2026.csv under a CC BY 4.0 license. Re-use with attribution. We do not publish manuscript-level data; only aggregate counts.

Press contact. Journalists, researchers, and editors interested in the underlying methodology or a specific subset of the data can reach the report's author at erik@manusights.com.

This report covers the 52-day window from 2026-03-23 to 2026-05-14. We plan a quarterly refresh: the next version will publish in August 2026 with data through July 31. If you'd like to be notified when the next version drops, we publish updates via the Manusights blog.