Research integrity screening that finds a duplicated figure before submission, explains what it can, and leaves every decision to a person
Duplicated figures fixed before submission, not retracted after publication.
See one case, screen by screen ↓Research integrity screening is a confidential check of a manuscript before it is submitted: its figure panels for duplication, rotation and splices, its p-values against their test statistics, and its references against retraction notices. Research Integrity Screening puts seven agents on that work for a university's Office of Research Integrity, and an integrity specialist decides every flag that reaches a person.
AI in education: a buyer's guide to agentic AI for colleges, universities and school districts →
Updated
Research integrity screening is the confidential check a university's Office of Research Integrity offers authors before they submit: are any figure panels duplicated, rotated, spliced or reused from the lab's earlier papers, do the p-values match their test statistics, and does the manuscript cite retracted work? Today that check is a specialist eyeballing figures in an image editor, perhaps one image tool licensed for a few labs, and nothing at all for statistics or references. A blot reused and turned 180° is hard to see by eye across 41 panels and nine earlier papers, so most manuscripts are never screened.
Image tools flag similarities without explaining them, so the specialist reads every legend and method by hand to learn that a loading control was disclosed as reused, and every false flag that reaches an author costs the office trust. The findings live in e-mails and screenshots, next to anything that might one day become a matter under the revised 42 CFR Part 93, which applies to allegations received from January 1, 2026. Research Integrity Screening puts seven agents in front of the comparison work, and a person decides every flag.
Estimated split of a specialist's hours on one manuscript, by hand and with the agents — our arithmetic, kept modest. By hand, most manuscripts are never screened at all.
Agents compare every panel, statistic and reference; a Second-Look agent explains what it can; a specialist decides, a second one checks.
Tuesday, October 20, 2026, at Halden State University's Office of Research Integrity. Dr. Lena Moravec's lab is about to submit MC-2026-0884, "DRP1 phosphorylation at Ser616 controls mitochondrial fission in migrating fibroblasts". Here is what happens to it, screen by screen, in the working solution.
The request comes in from the manuscript submission system: six authors, linked to award HSU-SP-2023-0417. The Intake agent splits six figures and the supplement into 59 lettered panels and matches 52 of them to raw files in the folder the lab shared. Rota AR-2 routes it to Daniel Achterberg, with the report due Thursday, October 22.
Daniel starts the screen and watches the agents' plan step by step: 71,567 panel comparisons against each other and the lab's 9 earlier papers (1,184 panels), 38 statistics recomputed and 64 references checked against 58,214 retraction notices. Eight candidates come out; four are explained; four go to Daniel. On the lightbox, an amber line joins panel 3C to the lab's 2024 paper.
0.97 · rotated 180°
Panels 3F and 4D are the same blot at 0.99. The Second-Look agent finds the Figure 4 legend disclosing it and explains the match with that passage, and 2B and S3A are explained as a labeled crop at 0.88. Nobody is told about either. Daniel can still flag one anyway, with a reason, if he disagrees.
"GAPDH loading control reproduced from Figure 3F"
Figure 3C (p-DRP1, S616A mutant, 24 h) and Figure 5B of Moravec et al. 2024 (p-DRP1, control siRNA, 48 h) sit side by side, a 212 × 148 px region outlined on both: similarity 0.97, 180°, 31,376 px², 1,284 matched points. The Second-Look agent found no disclosure, no labeled reprint and a different stated experiment. Daniel includes it in the private report with reason code INC-01.
This is a screening finding, not an allegation.
The Provenance agent finds 2026-03-14_DRP1-S616A_24h_blot2.tif in the lab's shared folder: acquired March 14, 2026 at 10:41, lanes labeled 0, 6, 12, 24, 36 and 48 h to match panel 3C, and used in none of the lab's earlier papers. It suggests the file; the author chooses.
Table 2, row 3 reports t(14) = 2.38, p = .012. From t, p is .032 two-tailed; the lab's quantification file gives t(14) = 2.88, p = .012 — a typing error, not a wrong result. The retraction check finds reference 17, retracted March 4, 2025 and cited as support, and reference 41, retracted November 16, 2023 and cited for a method. A stated Bonferroni adjustment and a stated one-tailed test are explained and never reach the author.
Four items — Figure 3C, Table 2 row 3, references 17 and 41 — each cited to an image region, a table cell or a notice, with a suggested fix. Daniel confirms the three items the agents proposed in one step and sends the report for a check. The Report agent never sends anything.
Confidential · not an allegation
Noor confirms that every item cites its evidence, the wording is neutral, no other author's unpublished data is shown and the fixes are the author's choice. She checks it with reason code CHK-01 and it goes privately to Dr. Moravec. Daniel, who prepared it, could not have checked it himself.
Dr. Moravec sees only her own report. She uses the original blot2.tif for 3C and the Image agent re-screens it at once: highest similarity 0.21. At 14:34 she accepts the three corrections — t(14) = 2.88, reference 17 replaced by reference 29, reference 41 cites its notice — and sends her changes back.
This is a service to you before you submit, not an allegation.
Daniel closes MC-2026-0884 as RES-01, resolved before submission. The original and the replacement 3C are both kept in the confidential screening record for 7 years. No referral was made, and nothing reached a journal.
The same manuscript, seen by the five people who carry it — what their day looked like, and what it looks like now.
Agents compare every panel, statistic and reference; a Second-Look agent explains what it can; a specialist decides, a second one checks.
Splits every figure into lettered panels, reads the legends, and matches each panel to its raw file in the lab's shared folder.
Compares every panel with every other panel and with the lab's earlier papers for duplication under rotation, flip, scale and crop, splices and generated textures, and outlines the shared region on both panels.
Recomputes every p-value from its test statistic and degrees of freedom, checks degrees of freedom against the sample sizes, and recomputes from the lab's raw quantification file where it is shared.
Checks every reference against retraction notices, expressions of concern and corrections, by DOI or by title and authors, and shows where each is cited.
Before any flag reaches a person, searches the methods, legends, supplement and raw files for an innocent explanation: a disclosed reuse, a labeled crop, a stated correction or a one-tailed test.
Finds the original acquisition for a flagged panel in the lab's shared folder, so the author is offered a fix, not only a problem.
Drafts the private pre-submission report from the specialist's decisions, in neutral words, every item cited with a suggested fix.
Decide, check, and only the RIO escalates. The agents propose; a named person decides.
A specialist, or an author about her own report, can ask in plain words and get an answer that cites the passage, cell or file it came from.
Why is 3C flagged but 3F is not?
Both match another panel. 3F matches 4D at 0.99, but the Figure 4 legend discloses it: "GAPDH loading control reproduced from Figure 3F". 3C matches Figure 5B of the lab's 2024 paper at 0.97 after a 180° rotation, the conditions differ, and nothing in the methods mentions reuse. So the Second-Look agent explained 3F and handed 3C to Daniel Achterberg.
Recheck the statistic in Table 2
Table 2, row 3 reports t(14) = 2.38, p = .012. From t, p is .032 two-tailed. The lab's quantification file gives t(14) = 2.88, p = .012, so the p-value is right and the t value was mistyped. Suggested correction for the author: t(14) = 2.88.
What still needs me today?
9 manuscripts need a person, 2 of them past due and escalated to Noor Haddad. 14 all-clear notes are ready to send in bulk after a check, 4 reports wait for a check, and referral RF-2026-007 waits for the Research Integrity Officer.
Who can see a screening flag?
The specialists who screen it, the author of that manuscript once a checked report is sent, and the Research Integrity Officer if it is referred. Internal Audit sees the record with names withheld unless a need to know is recorded. A flag is never an allegation.
13 screens from the working solution, on its sample data. Pick one to see it large.
The administrator proposes a change and the Associate Vice President for Research approves it as a new settings version. The agents, models and prompts in force are written into every screen's record.
| Setting | Default | Choose from |
|---|---|---|
| Image similarity that becomes a flag | 0.90 | Below it, a match is recorded but nobody is told |
| Splice score that becomes a flag | 0.60 | Lane boundaries and pasted regions |
| Generated-texture score that becomes a flag | 0.50 | Lowered from 0.55 in settings version 14 |
| A statistic recomputes when | It agrees to the reported decimals | Or within 0.001 |
| The Second-Look agent may explain a flag only with | A cited passage | Or a cited passage and a person's confirmation |
| Expressions of concern | Reach a person | Or inform only; retractions always reach a person |
| Report to the author | 3 business days from arrival | Escalated to the senior specialist 1 day past due; author reminder after 5 days |
| Screening record retention | 7 years after closure | Both images kept with hashes; separate from misconduct records |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
Sources: Bik, Casadevall and Fang, mBio (2016): a problematic duplicated figure in 3.8% of 20,621 biomedical papers · Nuijten et al., Behavior Research Methods (2016): statistical reporting errors in psychology, 1985–2013 · Michalek et al., PLoS Medicine (2010): the direct costs of one research misconduct investigation · 42 CFR Part 93 as revised, applying to allegations received from January 1, 2026; §93.106 (confidentiality) · "target" = design goal, measured in the live solution · "demo" = seen in the working solution, on its sample data · "typical" = published figures · "estimated" = our arithmetic, kept modest. Halden State University and the people named on this page are fictional — characters and sample data in the working solution.
Research integrity screening is a confidential check of a manuscript before it is submitted: its figure panels for duplication, rotation and splices, its p-values against their test statistics, and its references against retraction notices. Research Integrity Screening puts seven agents on that work for a university's Office of Research Integrity, and an integrity specialist decides every flag that reaches a person.
The Image agent compares every panel with every other panel in the manuscript and with the lab's earlier papers, under rotation, flip, scale and crop, and outlines the shared region on both panels. In the working solution it matched panel 3C to Figure 5B of the lab's 2024 paper at 0.97 after a 180° rotation. It also scores possible splices and generated textures against their own thresholds, and only scores; a person judges.
Yes. The Stats agent recomputes every p-value from its test statistic and degrees of freedom, and from the lab's raw quantification file where it is shared, which tells a typing error from a wrong result. The Citation agent runs a retraction check on every reference and shows where each retracted paper is cited.
Before any flag reaches a person, the Second-Look agent searches the methods, legends, supplement and raw files for an innocent explanation, such as a disclosed reused loading control, a labeled crop, a stated Bonferroni adjustment or a stated one-tailed test. It may explain a flag only with a cited passage, and a specialist can still flag it anyway, with a reason.
No. The report goes privately to the author as a service before submission. Only the Research Integrity Officer can open an allegation assessment, which creates a separate record, and the agents may not contact authors, journals or co-authors or call anything misconduct.
The specialists who screen it, the author of that manuscript once a second specialist has checked it, and the Research Integrity Officer if it is referred. Names are limited to those who need to know, and Internal Audit and the Insights page see them withheld.
Yes. A specialist decides every flag with a reason code, a second specialist checks every report and never their own, the author chooses every fix, the Research Integrity Officer decides any referral, and the Associate Vice President for Research approves each settings change.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your screening policy, reason codes and a sample of past manuscripts — and it goes live once every quality gate has passed. We will walk you through it on your own manuscripts first.
Two blind votes on every record, a person on each conflict, every number tied to its page.
1-min demoArchival Transcription & Finding AidsBacklogs described in weeks, not years, with every line approved by an archivist.
1-min demoResearch Conflict of Interest & Research Security ReviewUndisclosed ties found before just-in-time, and explained by the researcher first.
1-min demoWe'll run Research Integrity Screening on a sample of your own manuscripts and figures.