Prelaunch preview. Explore AROS products and skills. Checkout is in sandbox; no real payment.
Blog

Why we turn the rules for auditing research into AI agent workflows

Science's checks catch errors, but years late. Agent workflows that audit research run them while the work is done, helping people and AI do research together.

Science's checks catch errors years after publication; agent workflows that audit research run them while the work is done.

Wang Ziyi20 min readEnglish

A language model can draft a plausible grant application, literature review, manuscript or referee report in minutes. What makes any of those documents scientific is a set of rules older than the model. Since the Second World War, public research has relied on a small set of institutions to decide which questions to back and which claims to trust: competitive grant review, structured synthesis and analysis of the literature, the standard form of an academic manuscript, and peer review. Each is a set of rules about what a scientific argument must contain and how it may be checked.

Our skills are built around those exact forms: the biomedical-literature-evidence skill, the Mendelian randomization skill, the meta-analysis skill, the scientific-figure-master skill, the grant-write-master skill, the scientific-writing skill and the academic-peer-review skill. Rather than invent a separate, uninspected workflow for AI, we encode the checks of institutional science as executable software, workflows and AI agent skills, so an agent is held to the same standard as a human researcher. And because software can run a check at any moment, the check moves to where it does the most good: while the work is being done, not years after it is published.

The rules work, but they work late

Over eighty years, the combination of competitive funding, open publication, peer review and independent replication has overturned many entrenched clinical dogmas and exposed sustained academic fraud. It works after the fact, by forcing claims into a public record where better evidence can defeat them, and that takes time.

Prefrontal leucotomy is one example. It was introduced by Egas Moniz in the mid-1930s, popularised in the United States in its transorbital form by Walter Freeman, and performed on tens of thousands of patients; Moniz shared the 1949 Nobel Prize in Physiology or Medicine for the procedure [1]. What ended it was its mounting harm, set beside a workable alternative. Later reviews report blunted personality and apathy, late seizures in 23% of patients, chronic headaches in 15% and mortality of up to 5% [2]; once chlorpromazine arrived in the early 1950s, drug therapy replaced surgery as the safer and more effective option [1, 2]. The system corrected itself, but it took about a quarter of a century.

A generation later the same machinery dismantled a stem-cell programme whose data did not survive investigation, on a similar timescale. From 2001, Piero Anversa's laboratory reported in leading journals that bone-marrow cells [3], and later resident c-kit+ cardiac cells [6], regenerated infarcted heart muscle. Independent laboratories tested the claim where it could be tested. Tracking genetically marked haematopoietic stem cells across 145 transplants into normal and injured mouse hearts, Murry and colleagues found no detectable transdifferentiation into cardiomyocytes [4]. In 2014, thirteen years after Anversa's first regeneration report, lineage tracing by van Berlo and colleagues put the contribution of c-kit+ cells to new cardiomyocytes at roughly 0.03% or less, a rate likely to be functionally insignificant [5]. Harvard Medical School investigated the laboratory, the phase I SCIPIO trial paper was retracted, and many papers from the laboratory followed [6]. Between the first claim and the retractions, a clinical trial had been run on patients.

What slips through while the checks wait

The rules were written for fallible people under strong pressure to report clean, positive stories, and the large replication projects run since 2015 have measured how much gets through when the rules are applied only at the end.

When the Open Science Collaboration repeated 100 studies from three psychology journals, statistical significance held in only 36% of the replications compared with 97% of the originals, and the replicated effects were half their original size [7]. The Reproducibility Project: Cancer Biology repeated 50 preclinical experiments from 23 high-impact papers, yielding 112 effects that could be scored on all five replication criteria [8]. Fewer than half of those scored effects met at least three of five replication criteria (51 of 112, or 46%), and among positive effects the median replication effect was 85% smaller than in the original reports [8].

Behind those gaps sit both deliberate misconduct and everyday shortcuts. In Fanelli's pooled analysis of survey data, about 2% of scientists admit to having fabricated, falsified or altered their own data at least once, and in individual surveys up to a third admit to other questionable practices [9]. Asked about colleagues, 14% of scientists report having seen falsification, and in individual surveys up to 72% report other questionable practices [9]. Across disciplines, text-mined p-values show a small excess just below 0.05, the signature of analyses adjusted until a result crosses the line, though the authors judge the effect on conclusions to be weak [10]. When Bik, Casadevall and Fang screened 20,621 papers containing Western blots, 782 of them (3.8%) held inappropriate image duplications, at least half with features suggesting deliberate manipulation [11].

Most telling are the errors that pass authors and referees when a standard database already holds the answer. In an October 2026 preprint, Sholto David took 20 commonly used cell lines with confirmed deletions or biallelic inactivating mutations that abolish 10 specific proteins: MCF-7 breast cancer cells lack caspase-3, Jurkat leukaemia cells lack BAX and p16-INK4a, and PC-3 prostate cancer cells lack p53, PTEN and the whole STAT3 gene [12]. A literature search that one person could carry out for Western blots claiming to detect those proteins without the gene having been re-expressed found 230 such reports across 221 published papers, led by 49 papers showing caspase-3 in MCF-7 cells, 47 showing BAX in Jurkat cells and 35 showing STAT3 in PC-3 cells [12]. Some of the errors trace to antibodies: at least one BAX blot used an antibody reported to bind a different protein of the same size, and several papers used an antibody (ab51243) raised against p16-ARC rather than the similarly named tumour suppressor p16-INK4a [12]. Each of these deficiencies is documented in public resources such as Cellosaurus and the Human Protein Atlas. The rule existed, yet nobody ran it at the moment it mattered. And, as Derek Lowe observed in Science, the AI systems now being asked to read the whole literature for us will be fed these blots along with everything else [13].

A model has no career stake, and no commitment to truth either

An agent now produces the very documents those rules were written for, so the first question is whether the rules can judge it. They can. When Yamada and colleagues submitted three manuscripts written autonomously by The AI Scientist-v2 to a double-blind ICLR 2025 workshop, one scored 6.33, in the top 45% of submissions and above the acceptance line, before being withdrawn by prior agreement with the organisers [16]. The paper's hypothesis, that a compositional regularisation term would help networks generalise, did not hold; the experiments showed no benefit and sometimes harm. Where a researcher with a grant to renew might have hunted for a subgroup that worked, the agent reported the negative result plainly, at a workshop devoted to negative results, and the reviewers praised it for doing so [16].

The same paper shows what peer review misses. When Yamada and colleagues audited that manuscript and its code afterwards and re-ran the synthetic data generator, which could produce only a few dozen distinct expressions per operator, they found that about 57% of test items also appeared in training, that the caption of Figure 3 misstated the validation-loss trend, and that the prose called two architectures equal while Figure 5 showed one clearly ahead [16]. The agent had no ego to protect, but it also had no built-in loyalty to the truth. Left to write a literature review from its weights, a model invents references. When Walters and Wilder checked the citations in 84 short model-written reviews, 55% of the 222 citations from GPT-3.5 and 18% of the 414 from GPT-4 were fabricated [17]. Even among the citations to real papers, substantive errors appeared in 43% of those from GPT-3.5 and 24% of those from GPT-4 [17]. Assistants trained on human feedback also tend to bend towards the user's stated belief rather than the evidence [18].

A second model is a weak referee for the first. In a blinded study of NLP research ideas, every language-model evaluator tested agreed with expert rankings less often (43% to 53%) than the experts agreed with one another (56%) [15]. Model-written review is useful as one voice at the table: the points GPT-4 raised overlapped with a human reviewer's about as often as two human reviewers overlapped with each other, across 3,096 Nature-portfolio papers and 1,709 ICLR submissions [14]. But the standard it is held to has to come from the institution, and it has to be enforced by something more reliable than another draw from a model.

Rules as nodes

In our previous essay we argued that an agent should never rewrite a standard check from scratch: it should compose deterministic, tested workflow nodes, each packaging its code, with narrow AI nodes that make bounded semantic judgements. Applied to scientific writing and review, that architecture gives an agent a procedural honesty that depends on no human preference. The rule stops being a norm the agent is asked to respect and becomes a step it cannot skip.

Take the missing-protein check. Turning it into a workflow pairs the two kinds of node:

  1. Narrow AI nodes, each with a customised prompt, the skills and context it loads, and a gated hook for quality control, read a manuscript or protocol one question at a time. One extracts each (cell line, target protein, antibody catalogue number, assay) tuple from the prose and figure labels. A second checks the Methods for a vector or genetic tool that re-expresses an otherwise absent gene.
  2. Deterministic code nodes look the extracted cell line and protein up in curated databases such as Cellosaurus, the Human Protein Atlas, Cell Model Passports and deposited sequencing records, and check the antibody catalogue number against known cross-reactivity and naming mix-ups such as p16-ARC versus p16-INK4a [12].

An agent working under that workflow cannot report endogenous STAT3 in PC-3 cells or a caspase-3 band in MCF-7 cells, two of the missing-protein errors in David's 230 reports across 221 papers [12]. A researcher reviews what the AI nodes extracted on the canvas as the work is done, rather than in a referee report two years later.

The seven skills apply the same rule to each form that institutional science takes:

Run the check before the vial is ordered

Catching an impossible protein or a wrong subtotal in peer review protects the literature. For the laboratory that ran the experiment, it comes years too late.

Take the most charitable view of those 221 papers [12]: that many began as an honest oversight, a researcher choosing a cell line or ordering an antibody without knowing that PC-3 lacks STAT3 or that ab51243 binds p16-ARC. When the mistake surfaces after eighteen months of grant money, cell culture and thesis work, the researcher faces a sunk-cost trap. Admitting that the foundational assay never measured its target means discarding the project. Rationalising a non-specific band, cropping a blot or adding a control that was never run lets the paper go forward to publication. Carelessness at the planning stage is the breeding ground for misconduct at the publication stage, and the slow check that eventually catches the fraud is the same slow check that created the pressure for it.

So the check belongs where the plan is written, and that is where we are putting it: in the AI-native writing tool we are building around the grant-write-master skill and the scientific-writing skill, which joins our other tools as a node in the Nexitia desktop app. There, the same deterministic cell-line and antibody nodes run when a proposal or protocol names a cell line and a target, before a vial is ordered. A deleted gene or a cross-reactive antibody found at the protocol stage costs seconds of compute instead of years of benchwork, and the sunk-cost pressure that turns an honest error into fraud never builds. That is what it means to move the rules of science upstream and have non-human programs help enforce them: the institution's checks, run not as a verdict at the end but as a step at the start, for the agent and for the person it works with.

A rule that runs as software has a second property that a rule in a reviewer's head does not: it can be read, kept and handed to someone else. How a scientist and an agent should share memory, workflows and methods, beyond the chat window where most of that collaboration happens today, is the subject of the next essay.

What comes next

The institutions built since the Second World War work because they require every scientific argument, whether a proposal, a systematic review, a causal estimate, a figure or a manuscript, to be in a form that peers can audit and falsify. We are testing these seven skills together with the cell-line and reagent check against unassisted agent baselines, measuring how often their deterministic and narrow semantic gates catch missing-protein claims, fabricated citations, prose-to-table mismatches and caption errors before a document reaches human peer review. We will publish the benchmark here in future posts.

References

  1. Staudt MD, Herring EZ, Gao K, Miller JP, Sweet JA (2019). Evolution in the treatment of psychiatric disorders: from psychosurgery to psychopharmacology to neuromodulation. Frontiers in Neuroscience 13: 108. doi:10.3389/fnins.2019.00108. Reviews the rise of prefrontal and transorbital lobotomy across tens of thousands of patients, Moniz's 1949 Nobel Prize, and the displacement of psychosurgery in the 1950s by chlorpromazine and lithium.
  2. Mehta SS, Vadali S, Singh J, Sadana SK, Singh A (2024). The legacy of Egas Moniz: triumphs and controversies in medical innovation. Cureus 16(10): e72056. doi:10.7759/cureus.72056. Documents prefrontal leucotomy's late sequelae, including seizures in 23% of patients, chronic headaches in 15% and mortality reaching 5%, alongside its sharp post-1960 decline after chlorpromazine was introduced in 1952.
  3. Orlic D, Kajstura J, Chimenti S, Jakoniuk I, Anderson SM, Li B, Pickel J, McKay R, Nadal-Ginard B, Bodine DM, Leri A, Anversa P (2001). Bone marrow cells regenerate infarcted myocardium. Nature 410: 701–705. doi:10.1038/35070587. Claimed that locally delivered bone-marrow cells generated new myocardium occupying 68% of the infarcted portion of the mouse ventricle nine days after transplantation.
  4. Murry CE, Soonpaa MH, Reinecke H, Nakajima H, Nakajima HO, Rubart M, Pasumarthi KBS, Virag JI, Bartelmez SH, Poppa V, Bradford G, Dowell JD, Williams DA, Field LJ (2004). Haematopoietic stem cells do not transdifferentiate into cardiac myocytes in myocardial infarcts. Nature 428: 664–668. doi:10.1038/nature02446. Tracked haematopoietic stem cells with cardiomyocyte-restricted and ubiquitous reporter transgenes across 145 transplants into normal and injured adult mouse hearts and detected no transdifferentiation into cardiomyocytes.
  5. van Berlo JH, Kanisicak O, Maillet M, Vagnozzi RJ, Karch J, Lin SCJ, Middleton RC, Marbán E, Molkentin JD (2014). c-kit+ cells minimally contribute cardiomyocytes to the heart. Nature 509: 337–341. doi:10.1038/nature13309. Genetic lineage tracing at the Kit locus in mice showed that endogenous c-kit+ cells generated new cardiomyocytes at roughly 0.03% or less (below 0.008% after accounting for cell fusion), a likely functionally insignificant rate.
  6. Bolli R, Tang XL, Guo Y, Li Q (2021). After the storm: an objective appraisal of the efficacy of c-kit+ cardiac progenitor cells in preclinical models of heart disease. Canadian Journal of Physiology and Pharmacology 99(2): 129–139. doi:10.1139/cjpp-2020-0406. Reviews the laboratory's claims that resident c-kit+ cardiac progenitor cells regenerate infarcted myocardium, and documents the Harvard misconduct investigations into Anversa's laboratory, the retraction of the phase I SCIPIO trial paper by The Lancet over unreliable laboratory data generated at Harvard, and the retraction of many papers from that laboratory.
  7. Open Science Collaboration (2015). Estimating the reproducibility of psychological science. Science 349(6251): aac4716. doi:10.1126/science.aac4716. High-powered replications of 100 studies from three psychology journals: 97% of original studies reported P < 0.05, 36% of replications did, and replication effects were half the original magnitude.
  8. Errington TM, Mathur M, Soderberg CK, Denis A, Perfito N, Iorns E, Nosek BA (2021). Investigating the replicability of preclinical cancer biology. eLife 10: e71601. doi:10.7554/eLife.71601. Repeated 50 preclinical experiments from 23 high-impact papers, yielding 158 effects (117 numerical, of which 112 could be scored on all five criteria): 46% of those 112 (51) met at least three of five binary replication criteria, and the median positive replication effect size was 85% smaller than the original.
  9. Fanelli D (2009). How many scientists fabricate and falsify research? A Systematic Review and Meta-Analysis of Survey Data. PLoS ONE 4(5): e5738. doi:10.1371/journal.pone.0005738. Pooled synthesis of 18 surveys: a weighted average of 1.97% of scientists (7 studies) admitted to fabricating, falsifying or modifying data at least once and up to 33.7% admitted to other questionable practices; 14.12% (12 studies) reported observing falsification by colleagues and up to 72% observed questionable practices.
  10. Head ML, Holman L, Lanfear R, Kahn AT, Jennions MD (2015). The extent and consequences of p-hacking in science. PLoS Biology 13(3): e1002106. doi:10.1371/journal.pbio.1002106. Text-mined p-values from open-access PubMed Central articles; across disciplines p-values bunched immediately below 0.05 compared with the adjacent lower bin, which the authors read as p-hacking that is probably common but whose effect seems weak.
  11. Bik EM, Casadevall A, Fang FC (2016). The prevalence of inappropriate image duplication in biomedical research publications. mBio 7(3): e00809-16. doi:10.1128/mBio.00809-16. Visually screened photographic images in 20,621 biomedical papers matching the search term "Western blot" across 40 journals (1995–2014): 782 papers (3.8%) contained problematic image duplications, at least half with features suggestive of deliberate manipulation.
  12. David S (2026). Erroneous reports of protein expression in cell lines with genetic deletions and mutations. Zenodo (21 October 2026). doi:10.5281/zenodo.23498482. A preprint that took 20 commonly used cell lines with confirmed homozygous deletions or biallelic inactivating mutations across 10 proteins (checked against Cellosaurus, the Human Protein Atlas, Cell Model Passports and GEO sequencing) and, in a one-person literature search, found 230 erroneous Western blot reports across 221 published papers, led by caspase-3 in MCF-7 (49 papers), BAX in Jurkat (47 papers) and STAT3 in PC-3 (35 papers).
  13. Lowe D (2026). The case of the should-be-missing proteins. Science, In the Pipeline, 21 October 2026. https://www.science.org/content/blog-post/case-should-be-missing-proteins. A commentary on David's preprint: 230 examples in 221 papers from a one-person search; the causes named are misidentified or contaminated cell lines and non-specific or wrong antibodies; and a warning that AI systems meant to make sense of the literature will be fed these errors.
  14. Liang W, Zhang Y, Cao H, Wang B, Ding D, Yang X, Vodrahalli K, He S, Smith DS, Yin Y, McFarland DA, Zou J (2024). Can large language models provide useful feedback on research papers? A large-scale empirical analysis. NEJM AI 1(8): AIoa2400196. doi:10.1056/AIoa2400196; arXiv:2310.01783. Across 3,096 Nature portfolio papers and 1,709 ICLR submissions, point overlap between GPT-4 and a human reviewer (30.85% and 39.23%) matched overlap between two human reviewers (28.58% and 35.25%).
  15. Si C, Yang D, Hashimoto T (2024). Can LLMs generate novel research ideas? A large-scale human study with 100+ NLP researchers. arXiv:2409.04109; ICLR 2025. Blinded study with 49 expert idea writers and 79 expert reviewers (298 reviews): every tested LLM evaluator had lower ranking agreement with the experts (43.3% to 53.3%) than the experts had with one another (56.1%).
  16. Yamada Y, Lange RT, Lu C, Hu S, Lu C, Foerster J, Clune J, Ha D (2025). The AI Scientist-v2: workshop-level automated scientific discovery via agentic tree search. arXiv:2504.08066. Submitted three autonomous manuscripts to a double-blind ICLR 2025 workshop: one paper reporting negative results on compositional regularisation scored 6.33 (top 45%, above the acceptance threshold), while the authors' own audit of its code and text found a 57% train–test overlap, an inaccurate Figure 3 caption and prose contradicting Figure 5.
  17. Walters WH, Wilder EI (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports 13: 14045. doi:10.1038/s41598-023-41032-5. Across 84 short literature reviews on 42 topics containing 636 citations (222 from GPT-3.5, 414 from GPT-4), 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated, and 43% of real GPT-3.5 citations and 24% of real GPT-4 citations contained substantive errors.
  18. Sharma M, Tong M, Korbak T, Duvenaud D, Askell A, Bowman SR, Cheng N, Durmus E, Hatfield-Dodds Z, Johnston SR, Kravec S, Maxwell T, McCandlish S, Ndousse K, Rausch O, Schiefer N, Yan D, Zhang M, Perez E (2023). Towards understanding sycophancy in language models. arXiv:2310.13548; ICLR 2024. Five state-of-the-art assistants trained with reinforcement learning from human feedback consistently exhibited sycophancy across four free-form text-generation tasks, matching user views over truthful answers.

Our records:

  • Biomedical Literature Evidence 1.0.0-public.1 validation record: 19 unittest methods passed; 3 of 3 RIS and CSL-JSON records retained as unverified; synthetic completed-insufficient ledger check (completed: true, support_ready: false).
  • Mendelian randomization 1.0.0-public.2 reproduction report: Walker et al. 2019 v2, case 2; 163 archived rows, 157 retained, 19 of 19 headline comparison cells matched, and the prose-versus-Table-4 P-value discrepancy recorded.
  • Meta-analysis 1.0.0 performance report: 28 open-access Cochrane reviews, 1,060 of 1,061 pooled analyses and 10,797 of 11,351 checkable quantitative items reproduced (95%).
  • Scientific Figure Master 1.1.2 skill page: 61 of 61 scored caption fields correct across 20 real captions; 32 unit tests passing; no significance star when a recomputed p = 0.2 contradicts the caption.
  • Grant Writing Master 7.0.0-public.1 validation record: 6 unittest methods and 8 routing cases passed; the fact-preservation record shows a synthetic revision keeping p = 0.20 and a non-causal association in English and Japanese.
  • Scientific Writing 1.1.0 release notes: same-input complete manuscript revision comparison with a separate editorial decision record and preserved null and contrary findings.
  • Academic Peer Review 1.0.2 skill page and benchmark: coverage calibrated against 1,035 reviewer comments across 60 published open reviews; on the PeerJ 20995 submission, read blind, 6 of 49 row-level reviewer points caught, 4 of them substantive matches after audit, including the decisive unsupported ranking, and 5 of 9 review-driven v1-to-v2 changes demanded. The skill's review of that submission is public.