A custom AI reviewer for trial registration and endpoint pre-specification

PerfectPaper’s registration agent reads a trial manuscript for the pre-specification failures editors check first: a missing or malformed registration identifier, a reported primary endpoint that differs from the registered one, a secondary endpoint promoted to the abstract, and analyses presented without any indication of whether they were planned. Copy the brief below into a custom agent slot.

Outcome switching is the substantive item here. It is common, it is usually inadvertent, and it is the finding most likely to change what a paper is allowed to claim. Mathieu, Boutron, Moher, Altman and Ravaud compared 147 adequately registered trials against their registry records in JAMA (2009;302:977–984) and found discrepancies between the registered and the published primary outcome in 31% of them; where the direction of the change could be assessed, it favoured a statistically significant result in 82.6% of cases, 19 of 23. The pattern is not that trialists lie. The pattern is that the interesting number wins the abstract.

When to use this agent

Run this agent whenever a trial manuscript’s registration, endpoint, or analysis plan could plausibly diverge from what was pre-specified. The six situations below are the most common triggers:

• Your manuscript reports a prospective clinical trial
• The registered protocol was written years before the paper
• The paper’s headline result is not the endpoint the trial was designed around
• Subgroup analyses appear in the abstract
• The trial changed during conduct and the registry record may not reflect it
• The trial has more than one primary endpoint, or a composite whose components are reported separately
• You inherited the manuscript from a departed first author and did not personally file the registration

The agent brief

The agent brief below is the complete instruction set for a custom PerfectPaper agent that checks registration and endpoint pre-specification. Paste it into a custom agent slot with work type verification, skill level expert, and tools web search.

Name: Trial registration and endpoint pre-specification

You are reviewing a clinical trial manuscript for registration and endpoint
pre-specification. Read the abstract, methods and results together.

Determine and report:

1. Registration. Determine whether a trial registration identifier is given,
   in which registry, and whether the identifier matches that registry's
   format. Report its absence. Report an identifier given only in a footnote
   or acknowledgement rather than in the methods or abstract, since journal
   policy generally requires it in a findable place.

2. Prospective registration. Determine whether the manuscript states that
   registration preceded enrolment, and report where enrolment dates and
   registration date cannot be compared from what is given.

3. Primary endpoint. Identify the primary endpoint as stated in the methods.
   Identify the result presented as the paper's main finding in the abstract
   and conclusions. Where these differ, report it explicitly, quote both, and
   state that this is outcome switching unless the manuscript explains the
   change. Treat this as the highest priority item in this brief.

4. Endpoint definition. Determine whether the primary endpoint is defined
   precisely enough to be measured the same way twice: the measure, the
   timepoint, and the population. Report vagueness in any of the three.

5. Secondary endpoints. Determine whether secondary endpoints are listed and
   whether the manuscript distinguishes them from the primary. Report any
   secondary endpoint presented in the abstract with the emphasis of a primary
   result, and any secondary analysis reported without adjustment for
   multiplicity where several are tested.

6. Exploratory and subgroup analyses. Determine whether analyses are labelled
   as pre-specified or exploratory. Report any subgroup result presented as a
   finding without that label. Report subgroup claims made without a test of
   interaction.

7. Protocol changes. Determine whether the manuscript describes any change to
   the protocol during conduct, including changes to eligibility, endpoints or
   sample size, and whether the reason and timing are given.

8. Reporting completeness. Determine whether every endpoint named in the
   methods is reported in the results. Report endpoints introduced in methods
   and then absent, which is selective reporting.

If web search is enabled, you may consult the registry record for the given
identifier and compare its registered primary endpoint against the manuscript.
Report differences factually, quoting both. If the record cannot be retrieved,
say so rather than inferring anything about it.

Note the instruction in the last paragraph. It is there deliberately. An agent asked to adjudicate a registry record it could not retrieve will produce a confident sentence about what the record says, and a fabricated registered endpoint is worse than no comparison at all — you would revise a correct manuscript to match an invented record. Registry records are also genuinely unreachable in ordinary circumstances: a record under curation, a European trial visible in the Clinical Trials Information System but not indexed by a general-purpose search engine, and a registry undergoing maintenance all look identical to a search failure. The brief therefore separates the two verifiable states, retrieved and not retrieved, and forbids the third.

Frequently asked questions about this agent

What is an AI reviewer for trial registration and endpoints?

An AI reviewer for trial registration and endpoints is a custom agent briefed to check pre-specification rather than science: whether a registration identifier is present and correctly formatted, whether registration preceded enrolment, whether the reported primary endpoint matches the registered one, whether secondaries and subgroups are labelled, and whether every endpoint named in the methods appears in the results. The brief above is the complete instruction set.

Does the agent read my registry record?

Only if you enable web search, and it is instructed to say so plainly rather than guess when the record cannot be retrieved. Without web search it checks internal consistency between your methods, abstract and results — comparing what the paper says about itself, not what the registry says about the paper.

Should I run this before submission or during revision?

Before submission. A missing identifier causes a return without review, and an unexplained endpoint discrepancy is far cheaper to address as a methods sentence you wrote yourself than as a reviewer comment you must answer after the discrepancy has already shaped their reading of the paper.

What this agent catches

The table below lists nine failures this agent flags, from a missing registration identifier to a protocol change reported with no stated reason, each mapped to the consequence it produces for the paper.

Why outcome switching is the finding most likely to change your claim

Outcome switching changes what a paper may claim because pre-specification is the only thing separating a confirmatory result from a hypothesis. The registered primary endpoint was chosen while the investigators were blind to the answer; a substituted endpoint may have been chosen after seeing several candidates, and the reported p-value carries no record of how many were looked at. This is p-hacking executed at the level of the endpoint rather than the model, and it is invisible in the manuscript because the discarded endpoints leave no trace on the page.

Not every switch is p-hacking, and the agent is briefed to keep the two apart. A change made for a stated reason, at a stated time, before anyone looked at unblinded results, is an amendment; the label the agent applies depends entirely on whether the manuscript discloses the reason and the timing, not on whether the endpoint changed. The distinction, and how to write it up when the change was genuinely legitimate, is covered later in this brief.

The prevalence is measured, repeatedly, and it is not a fringe problem. Chan, Hróbjartsson, Haahr, Gøtzsche and Altman compared 102 trial protocols against their publications in JAMA (2004;291:2457–2465), covering 3,736 outcomes, and found that 50% of efficacy and 65% of harm outcomes per trial were incompletely reported, that statistically significant results were substantially more likely to be fully reported than non-significant ones, and that 62% of trials had at least one primary outcome that was changed, introduced or omitted.

The COMPare project measured the same thing prospectively in the journals least likely to have the problem. Goldacre, Drysdale, Dale and colleagues assessed every trial published over a six-week period in the New England Journal of Medicine, The Lancet, JAMA, the BMJ and Annals of Internal Medicine, all five of which endorse CONSORT (Trials 2019;20:118). Across 67 trials, a mean of 76% of pre-specified primary outcomes were correctly reported, with a between-journal range of 25% to 96%; secondary outcomes ran at a mean of 55%, range 31% to 72%; and trials carried a mean of 5.4 undeclared additional outcomes each, range 2.9 to 8.3. Fifty-eight of the 67 trials, 87%, had discrepancies warranting a correction letter. Twenty-three of those letters were published, 40%, at a median delay of 99 days.

Two numbers in that paragraph are worth holding onto. A mean of 5.4 undeclared additional outcomes per trial is the reason a reviewer reads your results section against your registry entry rather than against your methods section — the methods section has already been rewritten to match the results. And only 43% of those 67 trials had a publicly available pre-trial protocol, which is why the registry record, thin as it is, does the work the protocol should have done.

Common questions about outcome switching

What is outcome switching?

Outcome switching is reporting a different primary endpoint from the one registered before the trial began, without explaining the change. It matters because the registered endpoint was chosen without knowledge of the results, and the reported one may not have been. Mathieu and colleagues found the change favoured a statistically significant result in 82.6% of the assessable cases they examined.

Reviewer 2 says my primary outcome doesn’t match my registration — what do I say?

State the fact rather than argue with it: name the registered endpoint, name the reported one, give the date of the protocol amendment that changed it, and say whether the change happened before investigators saw unblinded outcome data. If there was no amendment, say so and revise the conclusion to the registered endpoint — defending an unexplained switch usually means a lost revision cycle; disclosing it is usually two sentences.

Can I switch my primary endpoint if the original one didn’t accrue enough events?

Yes, but only if the change is dated, blinded, and disclosed. A lower-than-projected event rate is one of the reasons the FDA’s own guidance and CONSORT both recognise as legitimate, provided the amendment happened before anyone examined unblinded results by treatment arm and the manuscript states the original endpoint, the new one, and the date of the change.

What counts as a registration identifier, and where it has to appear

A registration identifier is an accession issued by a public registry, and the ICMJE names the acceptable set precisely: registration must be “in any registry that is a primary register of the WHO International Clinical Trials Registry Platform (ICTRP)… or in ClinicalTrials.gov”. The WHO network currently lists 18 primary registries, so a number issued by an institutional or sponsor-run database is not a registration however official it looks. The agent checks the prefix against the registry the manuscript names, because a mismatched prefix is the single most reliable textual signal that a number was transcribed from the wrong document.

A trial can legitimately carry two European identifiers at once, and the agent should not flag that as an error by itself. The Clinical Trials Regulation entered into application on 31 January 2022; from 31 January 2023 sponsors had to apply to start a trial through the Clinical Trials Information System; and from 31 January 2025 any trial approved under the old Directive that was still running had to comply with the Regulation. A trial that opened before those dates holds a legacy EudraCT number, and the same trial, if it was still running when it transitioned, also carries an EU CT number. Two European prefixes on one trial is therefore expected for anything that spanned the changeover, which is most European trials publishing now; the check that matters is whether each number’s format matches the registry named for it, not whether there is only one.

Where the number appears matters as much as whether it exists, because the person who first looks for it is not reading the paper. The ICMJE recommendation is explicit: “The ICMJE recommends that journals publish the trial registration number at the end of the abstract.” An identifier that appears only in an acknowledgement, a funding statement or a supplementary file is functionally missing during technical check, which is the screening pass that produces most cases of being desk rejected without review on a trial paper with no scientific problem. The rest of what that pass looks at is covered in desk screening at scale.

The registry record itself is not a bare number. An acceptable registry, per the ICMJE, must capture “the minimum 24-item trial registration data set… at the time of registration and before enrollment of the first participant” — the WHO Trial Registration Data Set, currently version 1.3.1. Those 24 items include the primary and key secondary outcomes, which is precisely why a registry entry can adjudicate an endpoint dispute even when no protocol was ever published.

Prospective registration, and the 21-day gap that catches good trials

Prospective registration means registration before the first participant is enrolled, and two authorities set that line in different places — which is how a trial can be fully compliant with United States law and still ineligible at an ICMJE journal. The ICMJE recommends “registration of clinical trials in a public trials registry at or before the time of first patient consent for enrollment”, and states flatly that “retrospective registration, for example at the time of manuscript submission, meets none of these purposes.” United States regulation, at 42 CFR § 11.24(a), requires registration of an applicable clinical trial no later than “21 calendar days after the first human subject is enrolled”. A sponsor who files on day 20 has met the regulation and missed the journal policy.

Retrospective registration is the majority case, not the exception. Trinquart, Dunn and Bourgeois pooled 40 studies covering 8,773 randomised trials in BMC Medicine (2018;16:173) and found that 53% of published trials were registered at all (95% CI 44% to 58%), and that across 5,529 trials the pooled proportion registered prospectively was 20% (95% CI 15% to 25%). Registration rose from 25% to 52% between 2005 and 2015, so the direction is right and the level is still low.

The practical instruction, when your own trial was registered late, is to state the fact rather than obscure it. Give the registration date and the enrolment start date in the methods, in that order, and say in one sentence why the gap exists. An editor who reads “registered on 4 March 2023, four months after enrolment opened, because the sponsor treated the study as observational until protocol amendment 2 introduced randomisation” is being handed a decision. An editor who reads a registration number with no dates and finds the gap themselves is being handed a reason to distrust the rest. The agent reports the second case as an item the reviewer will raise, because it is the one the manuscript can still fix.

Questions about registering a trial before enrollment

Do I have to register a trial before enrolling?

Most journals following ICMJE policy require prospective registration in a public registry as a condition of publication, regardless of the trial’s quality. The ICMJE line is registration “at or before the time of first patient consent for enrollment”; United States regulation at 42 CFR § 11.24(a) allows 21 days after the first participant enrols, so meeting the regulation does not guarantee meeting the journal policy.

I registered three weeks after my first patient enrolled — am I disqualified from publishing?

Not automatically, but you are outside ICMJE’s prospective-registration line even if you met the US regulatory deadline of 21 days. Some journals enforce the ICMJE policy strictly and will treat the trial as retrospectively registered regardless of the gap’s size; others weigh it case by case. State both dates plainly in the methods and the reason for the gap rather than letting an editor find it unexplained.

My registration number doesn’t start with NCT — is that a problem?

Not if the prefix matches a registry on the WHO’s list of 18 primary registries — ISRCTN, ACTRN, DRKS, ChiCTR and the others in the table above are all acceptable under ICMJE policy. It is a problem only when the prefix doesn’t match the registry named in the text, which usually means the number was transcribed from the wrong document.

What a completely defined primary endpoint contains

A completely defined primary endpoint names four things: the measure, the metric applied to that measure, the timepoint, and the population in which it is assessed. The WHO Trial Registration Data Set states the requirement at item 19, Primary Outcome(s) — “The Primary Outcome should be the outcome used in sample size calculations, or the main outcome(s) used to determine the effects of the intervention(s). Most trials should have only one primary outcome” — and asks registrants for the outcome name without abbreviations, the measurement metric or method, and the timepoints of primary interest.

SPIRIT, the reporting guideline that stands to trial protocols as CONSORT stands to trial reports, asks for the same specificity a step earlier — at the protocol stage, before a single participant is enrolled, so there is no results-driven pressure on the wording yet. A protocol written to SPIRIT’s standard is what a registry entry should already summarise; when the two disagree, the protocol is dated first and generally governs.

The United States regulatory definition is tighter still and worth quoting to a co-author who thinks the choice is negotiable after the fact. 42 CFR § 11.10 defines a primary outcome measure as “the outcome measure(s) of greatest importance specified in the protocol, usually the one(s) used in the power calculation”, and a secondary outcome measure as one “of lesser importance than a primary outcome measure, but is part of a pre-specified analysis plan… and is not specified as an exploratory or other measure.” Two consequences follow directly. If the endpoint in your abstract was not the one in your sample size calculation, it is not your primary endpoint. And “exploratory” is a category the protocol assigns in advance, not a label a discussion section can apply retroactively.

Vagueness in any of the four components is what the agent reports at item 4 of the brief, and each has a characteristic failure. “Change in pain” names a measure but no metric — mean change from baseline, proportion achieving a threshold and time to a threshold are three different endpoints from one instrument, and only the second and third require a stated responder definition. “Improved function at follow-up” names no timepoint, and a trial with assessments at 6, 12 and 24 weeks then has three chances to report a positive one. “Overall survival” without a censoring rule and an analysis population is not measurable twice the same way; that is the boundary where this agent hands off to competing risks and censoring and to intention to treat. A composite endpoint is defined only when every component carries its own diagnostic criteria and ascertainment window, and a surrogate endpoint needs its validation named as well as its measurement. Where the endpoint is a patient-reported scale, the threshold that counts as benefit belongs in the methods as a minimal clinically important difference with a citation, not in the discussion as an assertion.

Worked example: one abstract sentence, six findings

Consider an abstract sentence of a kind that reaches submission every week: “In this randomised trial of 240 patients, the intervention significantly improved six-minute walk distance at 12 weeks (p = 0.03), and reduced hospitalisation in patients with baseline NT-proBNP above the median.”

The agent reports six distinct items from that sentence, before reading the methods. First, whether six-minute walk distance is the registered primary endpoint, or whether the registry names a different one — a common real case is a registered composite of death or hospitalisation, with a walk-distance component promoted because the composite did not separate. Second, the timepoint: if the registry names 24 weeks and the abstract reports 12, the endpoint has been switched even though the measure is unchanged, and this is the version authors most often do not recognise as switching. Third, “significantly improved” with a p-value and no effect estimate fails the requirement that every result carry its precision, which is a separate objection about effect size rather than about pre-specification. Fourth, the hospitalisation result is a second endpoint reported in the same sentence with no indication of whether the trial’s alpha was split. Fifth, “in patients with baseline NT-proBNP above the median” is a subgroup, presented as a finding, with no interaction test and no statement that it was pre-specified — and a median split defined from the trial’s own data is a post-hoc subgroup by construction. Sixth, the sentence reports two endpoints and the methods will name more; the agent will check that every endpoint named in the methods reappears in the results, because the ones that quietly vanish are selective reporting.

The repaired sentence looks unremarkable and answers all six: the registered primary endpoint named as such with its registered timepoint, reported with an effect estimate and a confidence interval; the hospitalisation result labelled as a secondary endpoint with the multiplicity strategy stated; and the NT-proBNP result labelled as pre-specified or exploratory, with the interaction p-value given. That is roughly thirty extra words in the abstract, and it converts three findings a reviewer would challenge into three a reviewer can evaluate. The general habit behind it — matching claim strength to the design that supports it — is the subject of the overclaim check.

Secondary endpoints, multiplicity, and the promoted secondary

A secondary endpoint may be reported and may be interesting; what it may not do is carry the paper’s conclusion without its label. The failure the agent reports is the promoted secondary — a secondary result written into the abstract, the conclusion and the title with the grammar of a confirmatory finding, so that a reader who does not open the methods will take it as the trial’s answer. The label is what tells the reader which result the trial was powered and pre-specified to deliver.

Multiplicity is the second half of the same problem. The FDA’s guidance Multiple Endpoints in Clinical Trials: Guidance for Industry, finalised in October 2022, states the concern plainly: “When more than one endpoint is analyzed in a single trial, the likelihood of making false conclusions about a drug’s effects with respect to one or more of those endpoints could increase if there is no appropriate adjustment for multiplicity.” A trial with one primary and eleven secondary endpoints, each tested at 0.05, has run twelve chances at a positive result, and reporting the two that came back below 0.05 without saying so misrepresents the evidence. Name the strategy — hierarchical testing with a pre-specified order, a gatekeeping procedure, a Hochberg or Holm adjustment, or a stated decision that secondaries are descriptive and no inferential claim is made from them. Any of those is acceptable. Silence is the finding. The underlying machinery is covered in multiple testing correction, and what a reviewer writes when it is missing is in reviewer says multiple comparisons.

Naming a strategy is not the same as naming the right one, and a reviewer who knows the family will check the fit. A Holm step-down procedure controls the family-wise error rate under any dependence structure among the endpoints and is never less powerful than a plain Bonferroni correction, which is why it is the more common default; a Hochberg step-up procedure is uniformly more powerful again, but its validity assumption — non-negative dependence among the test statistics, which usually holds for correlated clinical endpoints but is not guaranteed — is a sentence the statistical methods section should state rather than assume.

There is a quieter form of switching that changes no endpoint’s name at all: substituting a per-protocol population for the trial’s registered intention-to-treat analysis without saying so. A per-protocol result is not automatically wrong, but if the registry or protocol specifies intention-to-treat as the primary analysis, reporting the per-protocol result as the paper’s headline finding is a form of outcome switching that operates on the analysis population rather than the endpoint’s name, and the agent’s methods-versus-results comparison is built to catch a mismatch there too.

Two primary endpoints deserve their own sentence, because they are not a doubling of the trial’s chances of success unless the alpha says so. A trial with co-primary endpoints requiring both to succeed is more conservative than a single-endpoint trial on the false-positive side — but the same requirement lowers statistical power, since both hypotheses must clear their bar, and a co-primary design underpowered for the harder of its two endpoints can fail to detect a real effect while producing no multiplicity red flag at all. A trial with two alternative primary endpoints where either suffices needs the alpha split or an ordered testing procedure. The agent reports two primaries with no stated allocation of alpha, because that combination is ambiguous rather than wrong, and ambiguity at the primary endpoint is the thing a statistical reviewer will not let through.

Subgroups: label them, and test the interaction

A subgroup claim requires a test of interaction, not a p-value inside each subgroup, and the distinction decides whether the claim survives review. Reporting that the treatment effect was significant in men (p = 0.02) and not in women (p = 0.31) says nothing about whether the effect differs by sex; the two p-values sit either side of a threshold and the interaction may be nowhere near significant. Wang, Lagakos, Ware, Hunter and Drazen set out the standard expectation in the New England Journal of Medicine (2007;357:2189–2194): report the interaction test, state whether the subgroup analysis was pre-specified, and limit the number of subgroups examined.

The interaction test itself is usually underpowered, which cuts against a common misreading in both directions. Brookes, Whitely, Egger, Davey Smith, Mulheran and Peters quantified this in the Journal of Clinical Epidemiology (2004;57:229–236): a trial with 80% power for the overall effect has only 29% power to detect an interaction of the same magnitude, and “for interactions of this size to be detected with the same power as the overall effect, sample sizes should be inflated fourfold, increasing dramatically for interactions smaller than 20% of the overall effect.” A trial sized only for its main effect will therefore miss most genuine interactions, so a non-significant interaction test supports “this trial could not resolve the question,” not “the subgroups behave alike.” The opposite misreading matters just as much when many subgroups are examined: with several interaction tests run at the conventional 0.05 threshold, an apparently significant one is exactly what chance alone predicts a fraction of the time, which is why the agent also checks whether the number of subgroups examined was limited and stated in advance.

CONSORT asks for the same labelling. CONSORT 2010 item 18 required “Results of any other analyses performed, including subgroup analyses and adjusted analyses, distinguishing pre-specified from exploratory”, and CONSORT 2025 carries the requirement at item 28. An unlabelled subgroup is read as post hoc by default, which means the labelling is worth adding even when the analysis genuinely was planned — the absence of the label removes evidence you actually have.

Three subgroup patterns draw a reviewer’s attention immediately. A subgroup defined by a median split of a continuous variable measured in the trial itself, since the cut point depends on the data. A subgroup that appears in the abstract but not in the methods. And a subgroup whose sample size makes the estimate uninterpretable, where the honest response is a limitations section stating that the analysis is hypothesis-generating rather than a discussion paragraph treating it as a finding.

Questions about secondary endpoints and subgroup findings

Can I report a secondary endpoint as my main finding?

A secondary endpoint can be reported, but it should be labelled as secondary and interpreted accordingly. The problem the agent reports is a secondary result presented with the emphasis of a primary one, which invites the reader to treat an exploratory finding as confirmatory rather than as evidence to be tested in a future trial designed around it.

Is a subgroup analysis okay if I’m upfront that I didn’t preregister it?

Yes — an honestly labelled post hoc subgroup analysis is legitimate as hypothesis-generating evidence; what draws a reviewer’s objection is the same analysis presented with confirmatory language, or a p-value reported without the interaction test that would show whether the subgroups actually differ. Label it exploratory, report the interaction test if you have it, and frame the conclusion as a hypothesis for a future trial.

When the endpoint change was legitimate

An endpoint change made for a defensible reason, at a defensible time, and reported completely is not misconduct — it is an amendment, and CONSORT expects you to report it. What converts an amendment into outcome switching is silence. Three facts make the difference: what changed, why, and when relative to any look at outcome data.

Report the change in the methods, not the discussion. State the original endpoint, state the replacement, give the date of the protocol amendment, and give the reason in concrete terms — a lower-than-projected event rate, a regulatory request, an instrument that proved unvalidated in the population, a data monitoring committee recommendation. Then state explicitly that the change was made while the investigators remained blind to treatment-arm outcome data, if that is true, because it is the fact a reviewer most wants and most rarely receives. CONSORT 2025 places this within item 10, Changes to trial protocol, which asks for important changes to the trial after it commenced including any outcomes or analyses that were not pre-specified, with reason, and item 10 was one of the three items revised in 2025 specifically to ask for outcomes and analyses that were not pre-specified.

Where the amendment happened after an interim analysis, say so and say who saw what. A blinded sample-size re-estimation based on the pooled event rate is a different object from an endpoint change made after an unblinded interim look, and a reviewer who cannot tell them apart will assume the worse one. The same discipline applies to a switch from superiority to non-inferiority, which changes the question the trial answers and requires its margin to have been justified in advance.

Adaptive and platform trial designs complicate this without weakening it. A response-adaptive design, a group-sequential design with pre-specified boundaries, or a platform trial with a master protocol can change which endpoint or which arm is analysed as data accumulate, and that is not outcome switching provided the rule for the change was written into the statistical analysis plan before the trial started and the decision was executed by a process — typically an independent data monitoring committee acting on a charter, not the investigators — that did not have access to unblinded results by treatment arm. The agent cannot verify a statistical analysis plan it has not been given; where a manuscript describes an adaptive element, it reports whether the plan and its date are cited, and treats an undated or absent one the same as any other unexplained change.

When the registry record is wrong rather than the paper

Sometimes the manuscript is right and the registry entry is stale, and the fix is to correct the record openly rather than quietly. Registries are updated by the responsible party, and an entry written in 2019 by a coordinator who has since left may name an endpoint the protocol amended in 2021. That is a real and common situation, and it is repairable.

What it is not is invisible. 42 CFR § 11.64 requires that “the Director will retain prior clinical trial registration information… and make it publicly available… through ClinicalTrials.gov so that updates do not result in the removal of any information from the original submission or any preceding update.” Every prior version of a ClinicalTrials.gov record remains readable, with dates, and reviewers do read them. An endpoint edited into the registry the month before submission is more damaging than the discrepancy it was meant to remove.

The correct sequence is: update the registry record to reflect the amended protocol, with the amendment’s own date; state in the manuscript that the endpoint was amended, giving the amendment date and the reason; and note the registry update in the cover letter so the editor is not the one to discover a version history that does not match the paper. Where a reviewer has already raised the discrepancy, the same three facts form the whole answer, and how to write a response to reviewers covers the shape.

How reliably can outcome switching be detected automatically

Automated detection of outcome switching is now measurable rather than hypothetical, and the published accuracy is high enough to be useful and low enough to require a human read. Cummins, Drysdale, Elson, Hussey, Goldacre and DeVito benchmarked an LLM workflow named RegCheck against the COMPare project’s own hand-coded assessments across 62 trials from high-impact journals, reporting 91.2% recall on outcome extraction, 83.6% accuracy on classifying outcomes as primary, secondary or non-pre-specified, and 85.6% accuracy on misreporting detection, rising to 94.8% after adjudication (medRxiv preprint, 2026; not peer reviewed at the time of writing).

Read those figures as a specification of where the residual work sits. Extraction — finding the outcomes named in a document — is close to solved. Classification, deciding which of several stated outcomes the trial actually treats as primary, is where a fifth of the errors live, and it is exactly the judgement a manuscript makes hard when the methods hedge. Adjudication lifting detection from 85.6% to 94.8% means a meaningful share of apparent discrepancies dissolve once a human reads both texts carefully, which is the argument for running this agent before submission, where a false positive shows you an ambiguous sentence, rather than after, where a false negative has already reached a reviewer.

How to check your own manuscript in twenty minutes

Run the check as a reconciliation between three documents, not as a re-read of one, because the defect is always a mismatch.

1. Open the registry record and copy its primary outcome verbatim into a scratch file, including the metric and the timepoint. Copy the key secondary outcomes too.
2. Copy your methods section’s primary endpoint sentence beneath it, verbatim. Compare word by word. A different timepoint is a switch; a different metric on the same instrument is a switch.
3. Copy the first result sentence of your abstract beneath both. If it is not the endpoint in lines 1 and 2, you have either a switch to explain or a label to add.
4. List every endpoint your methods name, then tick each one off against the results section and the tables. Anything unticked is selective reporting and must be reported or removed from the methods.
5. List every subgroup and secondary analysis in the paper, and mark each pre-specified or exploratory. Anything you cannot mark from the protocol is exploratory.
6. Check that the registration number appears at the end of the abstract, that its prefix matches the registry named, and that the registration and enrolment dates are both stated in the methods.
7. Open the registry record’s version history and confirm nothing was edited after your database lock that you have not disclosed.

The agent performs steps 2 through 6 against the manuscript text in one pass and reports the mismatches with locations; with web search enabled it attempts step 1 as well. Step 7 is yours, because it requires knowing your own database-lock date. If you are assembling several agents for one submission, how to write a custom reviewer agent covers the brief structure this one uses.

What editors and reviewers actually write

Reviewers rarely write the words “outcome switching” first. They write the specific version, and these comments decide papers.

“The primary outcome reported here does not match the primary outcome in the registry record; please explain the discrepancy.” “The registration number is not given in the abstract.” “Registration appears to postdate the start of enrolment; please state both dates.” “The primary endpoint is described as change in symptom score without a timepoint.” “Several secondary endpoints are tested and no adjustment for multiplicity is described.” “The subgroup finding highlighted in the abstract is not accompanied by a test of interaction.” “The authors report a protocol amendment but do not state when it occurred relative to the interim analysis.” “Outcomes listed in the Methods do not all appear in the Results.” “The conclusion is based on a secondary endpoint; please revise it to reflect the primary result.”

The last of those is the one authors most often try to negotiate, and it is the one editors most often hold. A conclusion is a claim about what the trial showed, and a trial shows what it was designed and powered to show. Adjusting the conclusion is usually a two-sentence edit; defending the original is usually a lost revision cycle. The habit of matching the strength of a claim to the design that supports it is the same discipline enforced by causal language review.

What this agent does not do

The registration agent checks pre-specification and reporting, not the trial. It does not evaluate whether the endpoint was a good choice, whether the analysis was correct, whether the survival model was appropriate, or whether the effect is clinically meaningful — those are separate questions and, in a PerfectPaper review, largely separate agents in the clinical trials set. It does not judge whether an endpoint change was justified; it reports that a change occurred, whether a reason and a date were given, and leaves the judgement to a human who knows the field. It does not confirm that a registry record exists, for the reason stated after the brief. It does not read LaTeX source; manuscripts are reviewed as .docx.

Nor does it substitute for a trial statistician. The agent finds absences and mismatches reliably, with locations, because those are textual; it cannot tell you that the endpoint your registry names was the wrong endpoint for the question you care about. Because manuscripts at this stage are unpublished and often carry unblinded results, how PerfectPaper handles unpublished research is worth reading before uploading.

How this differs from the built-in review

This custom agent differs from PerfectPaper’s built-in review in one specific comparison: it checks the registered endpoint against the one the abstract actually reports, which the standing team’s checklist-based routing does not do on its own. The standing team includes a submission-requirements specialist and a reporting-guideline router that will identify CONSORT as applicable and check the checklist items. Checking a checklist establishes that a primary endpoint is stated somewhere in the paper. It does not compare that endpoint against the one the abstract actually leads with, which is the comparison that finds outcome switching, and which requires reading the two parts of the paper against each other rather than against a list.

That difference is the whole argument for spending a custom agent slot here. A checklist is a list of presences; outcome switching is a relation between two present things. The COMPare finding that all five journals endorsed CONSORT while breaching it extensively is the empirical form of the same point — endorsement of a checklist did not produce correct outcome reporting, because nobody was performing the comparison. For what else journals check on arrival, see journal policies.

Pairs with response criteria, which examines whether the endpoint was measured correctly once chosen. Full set: AI peer review for clinical trials.

PerfectPaper reads the registry entry, the methods and the abstract against one another, and reports every place the endpoint the trial was designed around is not the endpoint the paper leads with.

Review my manuscript

Related agents

The agents below share a review with this one; you can run three at a time. See AI peer review for clinical trials for the full set of five therapeutics agents.

• Response criteria and endpoints — use it for confirmed responses, censoring rules and waterfall consistency.
• Early-phase safety and dose reporting — use it when the dose-limiting toxicity definition or its window is unstated.
• Drug synergy claims — use it when synergy is asserted from a three-bar figure with no reference model.
• Translational dose relevance — use it when the effective concentration may exceed achievable human exposure.