Evidence Credibility Engineering

An AI system summarizes a flawed study as smoothly as a sound one, and a reader who sees only the summary cannot tell which is which. I believe this because of how these systems write, not because I have measured it here.

The intake names twelve separate dimensions: study design, risk of bias, statistical integrity, reporting, registration and protocol consistency, data availability, conflicts and funding, journal and editorial process, replication, applicability to a decision, certainty across a body of evidence, and post-publication history. A system that records them separately can show why a study is weak. A system that merges them cannot. I recommend that no output in this program merges them until a validated reason says it may.

A credibility record should be built in layers: source acquisition, extraction of facts with their location in the source, checks that are deterministic (the same input always gets the same answer), framework-based appraisal, applicability to a declared context, external evidence, human review, and a versioned record. Where a rule can answer, such as whether a retraction record exists or when a registration was first submitted, it should answer, and a model should handle only what rules cannot. A rule that depends on data the sources only partly hold has to say what it could not check. The farther a system moves from extracting what a paper says toward judging how much to trust it, the more validation it needs.

Sources: the ClinicalTrials.gov API specification, ICMJE’s registration policy, 42 CFR 11.24, eCFR version of 2026-10-01, Crossref’s Retraction Watch page, Crossmark’s update rules and live API calls, read 2026-10-05, with the record histories read 2026-10-06 (source-check.md, Parts 3 to 8). The record history of one example study shows the page as a reader sees it.

The public ClinicalTrials.gov API describes itself as returning "the most recent version of the clinical trials" and lists no history path, so a check of a paper’s primary outcome against the registered one would compare against the current record, which may have changed after publication. The website does keep that history: each study has a Record History tab that lists every version with its submission date and the sections changed, and the page’s data include the primary outcomes as first registered. In the one record where I compared two versions, the primary outcome changed nine days after first submission. The endpoint behind the tab is internal, is not in the published API specification, and refused my requests outside the page, so a system cannot depend on it without NLM’s agreement. Other registries differ: I read ISRCTN’s dated editorial notes, DRKS’s version list, jRCT’s change reports and CTIS’s application chronology, found no history feature in the EU Clinical Trials Register, and could not read ANZCTR or ChiCTR, which block automated page loads. None gave me a documented machine-readable history. A check has to record whether the registry’s history was available, so "no change found" and "history not available" stay different results. Registry terms also differ on automated access: jRCT asks for permission before comprehensive collection and refrains from automatic crawlers, WHO’s search portal disallows all crawlers, ISRCTN’s terms limit copying into a database while licensing its outcome fields under CC BY, and DRKS licenses only records from 2025 under CC BY 4.0 (source-check.md, Part 6). The first-submitted date is a registry field. The ICMJE counts registration "at or before the time of first patient consent", and the US rule allows submission up to 21 calendar days after the first participant is enrolled, so one registration can pass the law and fail the policy. Crossref lists retractions with their source, and its own pages say update records exist only where publishers deposit them and only for editorially significant changes. PubMed flags a retracted article with the publication type Retracted Publication and links its notice, from the journal’s notice and with no reason, and it treats a partial retraction as an erratum, since October 2016, and an expression of concern was an explicit link only from March 2017 (source-check.md, Parts 7 and 8). A search that finds no retraction record shows that none was found on that date. It does not show the paper was not retracted.

The assessor has to earn trust before it assigns any. I recommend a public benchmark that tests any assessor, human or machine, against an expert reference standard built from two independent reviews and adjudication, and that reports a profile and never one number. The profile covers fact extraction, provenance, methodological agreement, errors that would change a reader’s conclusion, reproducibility, calibration, correct abstention, and expert time saved. A system that says "unknown" when the source does not say should score better than one that invents certainty. The working name is the Research Credibility Benchmark.

A number is easy to rank and easy to build into software. It can also hide two studies that fail in entirely different ways, and a reader may treat 7.9 as meaningfully different from 7.6 when no such precision exists. The intake lists five ways to present an assessment: no composite, a category, a profile of separate domains, a composite shown with every part, and a score tied to one decision. I do not know which serves readers best. A test that shows the same studies to clinicians, researchers and others in each of the five forms, and measures whether each reader finds the main flaw and how much they trust the study, would resolve it.

Sources: RobotReviewer (2016), RobotReviewer against humans on 1955 trials (2024), ChatGPT-4o against Cochrane RoB 1 (2025), Claude 2 on RoB 2 (2025), four models on RoB 1 and RoB 2 (2026) and ChatGPT-o3 on 50 trials (2026). I read each abstract on 2026-10-05, not the full papers (source-check.md, L1 to L4).

On RoB 2, one study found kappa from 0.06 to 0.13 across four models, with models "consistently over-flagging concerns". Another found kappa 0.10 on selective reporting with Claude 2, the lowest domain. On trials rated against Cochrane RoB 1, ChatGPT-4o agreed with the human consensus on overall risk 50.7% of the time. Each study concludes the model cannot yet assess risk of bias alone. The studies that used published review judgements as the standard cannot say how often the published judgement was the one in error.

Source: PLOS ONE 2026, abstract read 2026-10-05. On 50 trials, ChatGPT-o3 rated 34% high risk, a masked human panel 22%, and the original systematic-review authors 12%. Agreement was median kappa 0.33 with the panel and 0.14 with the original authors. This is one study of 50 trials, and it supports building an adjudicated reference standard that keeps the disagreement.

Sources: RegCheck, arXiv 2601.13330, DARPA SCORE and its 2020 forecasting paper, read 2026-10-05, and the search results in literature-search.md. RegCheck compares a registration with a paper and keeps a person in the loop, and its authors write that its accuracy cannot be stated in general. SCORE built confidence scores for the replicability of social and behavioural science claims, which is a different question from methodological quality. I then ran a systematic search (literature-search.md, 2026-10-05) of Europe PMC and arXiv, 6,780 records and 5,743 unique, with the protocol and the inclusion rule written before the first query. No study reported three or more of six whole-assessor measures (provenance, abstention, calibration, reproducibility, error severity, expert time) for one system. Seven studies reported two (some only in part), usually reproducibility or provenance with time saved, and none reported calibration. The search covers two sources, coded most records from abstracts, and used one AI screener, so it supports "none found by this method" and not "none exist".

I read each named standard’s home page on 2026-10-05 and confirmed what it says it is for (source-check.md, Part 1). RoB 2, ROBINS-I, QUADAS and PROBAST assess risk of bias. CONSORT, PRISMA and STROBE guide how studies are reported, and STROBE says its checklist is not a way to judge quality. GRADE rates certainty across a body of evidence. The current tools are QUADAS-3 and PROBAST+AI. I have not read what any of them requires item by item. The literature above shows the risk-of-bias part has been tried with modest results, so this program is not first, and it has to rest on what those studies did not do: test the whole assessor, with provenance, abstention and conclusion-changing errors counted. Whether that is enough to justify the program is open. If a published test of a whole assessor turns up in a source I did not search, the program should narrow or stop. The sources not searched are Embase, Web of Science, Scopus and Google Scholar, and a person with access should run the same queries there.

I recommend 120 studies across three designs (randomized trials, observational cohorts, and systematic reviews or meta-analyses) and four clinical domains, with about equal numbers of straightforward, difficult and adversarial cases. Two qualified reviewers assess each independently, disagreements are adjudicated and kept, and an AI-assisted workflow is run blind against the result. We measure agreement domain by domain, repeat runs, the time experts spend correcting, and the time saved. Success is defined before the first run. Failures are published as prominently as successes. The pilot tests the benchmark as much as it tests any system.

The intake treats the benchmark, its expert-adjudicated corpus and its record of expert disagreement as the assets that would be hard to copy, and a venture as something that follows only if the benchmark shows a reliable, differentiated capability. Who would pay, and whether the better business is a product, a data asset, an audit service or an open standard with commercial implementations, is unknown. Evidence from the pilot and from people who would integrate the work would resolve it. The research does not depend on a venture existing.