Research integrity checks

Most integrity failures in the literature are not fraud - they are transcription errors, stale numbers from an earlier analysis run, rounding inconsistencies and copy-paste slips that nobody checked because checking by hand is tedious. That is exactly what makes this agent work: the checks are mechanical, the cost of running them is minutes, and finding an error BEFORE submission converts a potential correction notice into an edit. Frame every finding accordingly: this is proofreading for numbers, not accusation.

Internal consistency of reported statistics

Checks that need only the manuscript text:

• statcheck-style recomputation: for every reported test statistic with degrees of freedom and p-value (t, F, chi-square, r...), recompute the p from the statistic and df and flag mismatches - the classic detectable error class in psychology and beyond (the statcheck R package automates this for standard reporting formats; the same arithmetic can be applied directly).
• GRIM-style granularity: reported means of integer data with known sample size are only possible on a discrete grid. With n=25 the grid is multiples of 1/25, so a reported mean of 3.47 is impossible (3.47 x 25 = 86.75, and 25 integer responses cannot sum to that) while the neighbouring 3.48 is attainable. Work out the grid before calling a value impossible - the near-miss is the whole point of the test. The same applies to percentages, with one extra step: a reported percentage is usually rounded, so test the whole interval it could have come from, not the point value. 34% of n=170 gives 57.8, but 57 and 58 both round to 34%, so 34% is perfectly reportable; 34% of n=3 is not, because no count rounds there. The scrutiny R package implements these granularity tests.
• Arithmetic on the page: totals that sum, percentages that reach 100 within rounding, subgroup Ns that add to the total N, confidence intervals consistent with the point estimate and SE.

The stronger check: manuscript vs pipeline

When the analysis code is available (it should be - rseng-reproducibility), do not settle for internal consistency:

• Regenerate the numbers: run the pipeline and compare every reported statistic, table cell and figure value against the fresh outputs, within stated rounding. Mismatches usually mean the text cites an OLDER run - exactly the silent staleness that one-command reproducibility (rseng-reproducibility) and generated-not-transcribed reporting prevent.
• Kill transcription at the source where feasible: propose generating tables and inline statistics from the pipeline outputs rather than retyping them; every hand-copied number is a defect opportunity.
• Check the figure data too: axis ranges, group counts and plotted Ns against the data files (a figure from the wrong CSV survives visual review easily).

Citations and provenance

• Retraction screen of the bibliography, and verification that every reference is real and correctly attributed - delegated to rseng-citation-hygiene; run it as part of this battery.
• Provenance completeness: data sources identified with versions and access dates (rseng-data-management), software versions and seeds recorded (rseng-reproducibility), AI contributions declared honestly in aidecl.yaml and manuscript disclosure sections (rseng-ai-declaration) - venues increasingly ask, and the honest record is the one that already exists.

The pre-submission battery

Run as one auditable pass and write the report into the project:

1. Statistics: statcheck-style + granularity checks over the manuscript.
2. Regeneration: pipeline outputs vs reported numbers, tables, figures.
3. Citations: existence, attribution, retraction screen (rseng-citation-hygiene).
4. Data/code availability statements true in practice: links resolve, the deposit exists, the archive matches the text (rseng-archiving, rseng-reproducibility).
5. Declarations complete: contributions, AI use, conflicts as the venue requires.

Report findings with locations and severities; fix-and-rerun until clean, and keep the final report with the submission record - it is evidence of diligence, and the checklist for the revision round.

Boundaries

An agent flags inconsistencies; it does not adjudicate misconduct. If checks surface a pattern that looks deliberate in someone ELSE's work, the route is the venue's editorial process and COPE- style guidance, via the user and the responsible institutional channels - not a public accusation from a tool report. For the user's own work, everything found is simply fixed before submission, which is the point of checking first.

Working with this skill

This skill is source-independent: its authority is the published consistency-check methods (statcheck, GRIM and granularity testing) and the services linked below.

Learn more (verified):

• https://github.com/MicheleNuijten/statcheck - statcheck
• https://github.com/lhdjung/scrutiny - scrutiny (GRIM and granularity tests)
• https://gitlab.com/crossref/retraction-watch-data - Retraction Watch database
• https://help.openalex.org - OpenAlex API

Related skills

Check whether any of these applies before moving on:

• rseng-ai-declaration - declarations complete at submission
• rseng-archiving - availability statements must resolve
• rseng-citation-hygiene - reference existence and retraction screens
• rseng-fact-checking - do sources support the claims
• rseng-numerical-accuracy - numeric mismatches may be float issues
• rseng-reproducibility - regenerate numbers from the pipeline