A custom AI reviewer for response criteria and endpoint definitions

PerfectPaper’s response criteria agent reads an oncology efficacy manuscript for definitional slippage: which response criteria were applied and in which version, whether responses were confirmed, how progression-free survival was censored, which analysis population each rate is computed over, and whether the waterfall plot agrees with the response rate in the text. Copy the brief below into a custom agent slot.

Response criteria are a defined vocabulary. Most objections here are not that the analysis is wrong but that the definitions were applied loosely, and the numbers therefore mean something other than what the reader assumes.

When to use this agent

• Your manuscript reports objective response rate, disease control rate, or progression-free survival
• The trial involves immunotherapy, where pseudoprogression makes criteria choice consequential
• Response assessment was done locally rather than by blinded independent review
• Your paper contains a waterfall plot
• Denominators differ between the efficacy table and the text
• The trial is single-arm, so the response rate is the result rather than a comparison
• Survival is compared between responders and non-responders anywhere in the paper
• Scans were acquired in routine care rather than on a protocol schedule

The agent brief

Paste this into a custom agent. Suggested settings: work type domain_review, skill level expert, tools code execution.

Name: Response criteria and endpoint definitions

You are reviewing an oncology efficacy report for correct and consistent use of
response criteria. Read the methods, the efficacy tables, the figures and the
results text together.

Determine and report:

1. Criteria and version. Identify which response criteria were used and which
   version, such as RECIST 1.1, iRECIST, irRECIST, Lugano, or RANO. Report
   criteria named without a version. Where an immunotherapy trial uses
   conventional criteria without addressing pseudoprogression, report it.
   Where two criteria appear in the same manuscript, determine which governs
   the primary endpoint and report any ambiguity.

2. Confirmation. Determine whether responses required confirmation at a
   subsequent assessment and, if so, after what interval. Report unconfirmed
   responses counted in a response rate presented as definitive, and state
   that the rate is likely overstated.

3. Assessment. Determine who assessed response: the investigator, or blinded
   independent central review. Report where this is unstated.

4. Analysis population. For every reported rate, determine the denominator:
   intention to treat, all treated, or response evaluable. Report every rate
   whose denominator is not stated, and every instance where denominators
   differ between the text and a table without explanation. Recompute each
   rate from the reported numerator and denominator and report any that does
   not reconcile.

5. Progression-free survival. Determine whether the event is defined,
   whether censoring rules are stated, and how patients who discontinued
   without progression were handled. Report the absence of censoring rules.
   Identify any indication that assessment intervals differed between arms,
   which biases the comparison.

6. Time-related bias. Report any analysis in which response status determines
   group membership and survival is then compared between responders and
   non-responders, since this is immortal time bias. Name it explicitly and
   describe the correct approach.

7. Waterfall and swimmer plots. Determine whether the number of bars crossing
   the response threshold is consistent with the reported response rate.
   Report any discrepancy with both numbers. Determine whether the plot shows
   all evaluable patients or a subset.

8. Duration of response. Determine whether it is reported, from what starting
   point, and whether censoring is described.

Use code execution to recompute every rate and to check that reported
percentages match their stated numerators and denominators.

The brief is written to be pasted verbatim; if you want to adapt it for a different endpoint family, how to write a custom reviewer agent covers the structure it follows, and custom AI reviewer agents lists the other prebuilt briefs.

What RECIST 1.1 actually specifies

RECIST 1.1 specifies numbers, and a manuscript that names the criteria without honouring the numbers is the failure this agent finds most often. The revised guideline — Eisenhauer EA, Therasse P, Bogaerts J, Schwartz LH, Sargent D, Ford R, et al., “New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1)”, European Journal of Cancer 2009;45(2):228–247 — permits a maximum of five target lesions and no more than two per organ, with lesions measurable at 10 mm or more in longest diameter and nodal lesions at 15 mm or more in short axis.

The response categories are arithmetic, and each one names the reference sum it is measured against. Partial response is a decrease of at least 30% in the sum of diameters, taking the baseline sum as reference. Progressive disease is a relative increase of at least 20% and an absolute increase of at least 5 mm, taking as reference the smallest sum recorded on study, or the appearance of at least one new lesion. Stable disease is everything that fits neither.

The 5 mm floor is the clause most often dropped in a methods section, and it is not decorative. A patient whose target sum reaches a nadir of 15 mm and then measures 19 mm has grown 26.7%, which clears the 20% threshold, but has grown only 4 mm, which does not clear the 5 mm floor — so that patient has not progressed under RECIST 1.1. A methods section that writes “progression was defined as a 20% increase in the sum of target lesion diameters” has described a different, more sensitive rule than the one it cites, and every progression-free survival curve downstream inherits the difference.

The sum-of-diameters arithmetic is not the whole rule, though, and a manuscript that reports it correctly can still misclassify a patient. A new lesion or unequivocal progression of a non-target lesion overrides a favourable target-lesion sum under RECIST 1.1 — a patient whose target sum has shrunk 45% is progressive disease, not partial response, if a new lesion has appeared. Complete response carries its own additional condition: disappearance of all non-target lesions, with any pathological lymph node reduced below 10 mm in short axis. A methods section that defines response only in terms of the target-lesion sum has not described the whole rule it cites.

Questions about RECIST 1.1 versions and thresholds

What is the difference between RECIST 1.1 and iRECIST?

iRECIST modifies RECIST for immunotherapy: an initial progression call becomes unconfirmed pending a later scan, since some tumours enlarge before responding. Applying conventional RECIST without addressing this can misclassify a future responder as progressing. Under iRECIST, that unconfirmed progression is only confirmed if imaging four to eight weeks later shows a further increase, such as at least 5 mm more in the target sum.

My target lesions grew 22% from nadir but only 4 mm — does that count as progression under RECIST 1.1?

No. RECIST 1.1 progression requires both a relative increase of at least 20% in the sum of diameters from the smallest sum recorded on study and an absolute increase of at least 5 mm. A 22% increase that adds only 4 mm clears the relative threshold but not the 5 mm floor, so it is stable disease, not progression, under the guideline.

What does an AI reviewer check in RECIST reporting?

An AI reviewer for response criteria checks eight things: which criteria and version were applied, whether responses were confirmed, who assessed the scans, the denominator behind every reported rate, whether progression-free survival censoring rules are stated, whether any responder-versus-non-responder comparison carries immortal time bias, whether the waterfall plot matches the reported rate, and whether duration of response has a start point and a censoring statement.

Confirmation, and why an unconfirmed response rate runs high

Confirmation is a rule with a stated scope, not a courtesy. The iRECIST guideline states it plainly: “Confirmation of response is not required when using RECIST 1.1, except in non-randomised trials” (Seymour L, Bogaerts J, Perrone A, et al., “iRECIST: guidelines for response criteria for use in trials testing immunotherapeutics”, Lancet Oncology 2017;18(3):e143–e152). The exception is the whole point: a single-arm trial reporting objective response rate as its primary endpoint is exactly the design in which confirmation is required, and exactly the design in which manuscripts most often omit whether it happened.

The mechanism is measurement error, not fraud. Unidimensional tumour measurement on CT is reproducible only within limits, and a lesion sum sitting near the 30% boundary can cross it on re-measurement alone. A systematic review of the factors driving inter-observer variability in RECIST assessment (Smesseim I, Bucho TT, Tissier R, van der Mierden S, Lalezari F, Burgers JA, Trebeschi S, “Assessing the reliability of Response Evaluation Criteria In Solid Tumors (RECIST): a systematic review of the factors contributing to inter-observer variability”, Insights into Imaging 2026;17(1):180) screened 246 studies and included 88, and found that measurement method, lesion selection and image quality are what move agreement. The kappa values reported by those studies do not converge on a single figure: agreement was moderate in a breast cancer series (κ = 0.53), κ = 0.59 for RECIST in a sarcoma series, with κ = 0.57 for the EpSSG criteria in the same series, κ = 0.65 for manual diameters in colorectal carcinoma, and κ = 0.9 to 1.0 for mRECIST in hepatocellular carcinoma. The review’s own conclusion is that the heterogeneity between studies is severe enough that standardised protocols are needed before the question can even be answered properly.

That spread is the point rather than a weakness in the evidence. Where two readers of the same scan agree only moderately, a rate built from single unconfirmed observations is biased upward by construction, and the size of the bias is a property of the tumour type and the measurement method rather than a constant. The agent reports the absence of confirmation and says which way it pushes the rate; it does not attach a correction factor, because no defensible single factor exists.

Investigator assessment versus blinded independent central review

Who read the scans changes what the response rate means, and the regulatory expectation is explicit. The FDA guidance Clinical Trial Endpoints for the Approval of Cancer Drugs and Biologics (December 2018) states: “When the primary study endpoint is based on tumor measurements (e.g., PFS or ORR), tumor assessments generally should be verified by central reviewers blinded to study treatments … to ascertain lack of assessment bias.”

The regulatory expectation is not the whole picture, though, and a reviewer who stops there will write an objection the evidence does not support. A meta-analysis of concordance between investigators and blinded independent central review (Tang X, Dang Y, Han S, Cui B, Kang Y, Luo X, Zhang H, “Concordance in assessments between investigators and blinded independent central review (BICR) in hematology oncology clinical trials: a meta-analysis”, The Oncologist 2025;30(11):oyaf375) pooled 70 studies published between 2014 and 2024 and found the two assessments produced very nearly the same answer. Across 37 progression-free survival comparisons the pooled ratio of hazard ratios was 0.96 (95% CI 0.89 to 1.03). Across 23 two-arm response-rate comparisons the pooled ratio of odds ratios was 0.99 (95% CI 0.85 to 1.14). Across 29 single-arm trials the odds ratio for investigator versus central response was 1.02 (95% CI 0.90 to 1.17). The authors read this as substantial concordance and argue it supports reduced reliance on routine central review in haematology trials.

Two boundaries on that result matter for how a reviewer uses it. The trials pooled were haematology-oncology, where response is often assessed against Lugano criteria on a combination of imaging and marrow or blood measures, rather than solid-tumour RECIST — so it does not license the claim that investigator and central RECIST assessment agree to within 2% in a solid-tumour trial, and the measurement-variability review above is the reason to be careful about assuming it transfers. And a pooled average is not a guarantee about any single study; it says the systematic tilt is small, not that no individual trial is misread.

What survives is narrower and more defensible than the folklore. A reviewer who demands central review of a randomised, blinded phase III is demanding something the evidence says would change little, and the FDA guidance itself names the escape hatch: central verification “may not be necessary when randomized trials are blinded (unless the adverse event profile would substantially unblind the trial in practice) or effect sizes are robust in large randomized trials where sensitivity analysis supports lack of investigator bias.” The reportable defect is almost never that central review was skipped. It is that the manuscript does not say who assessed response at all, which leaves the reader unable to judge the question either way. The agent reports the omission and leaves the design argument to the authors.

Common questions about confirmation and independent review

What is a confirmed response?

A response observed at one assessment and confirmed at a subsequent one, typically at least four weeks later. Confirmation exists because a single measurement can fall below the threshold by measurement variation alone, so unconfirmed rates are systematically higher. The iRECIST guideline states that confirmation is not required under RECIST 1.1 except in non-randomised trials, which is precisely the design in which an unconfirmed rate does the most damage.

Do I need blinded independent central review if my phase 3 trial is already randomized and double-blind?

Not necessarily. FDA guidance states central verification may not be needed when randomized trials are blinded, unless the adverse-event profile would unblind them in practice, or when effect sizes are robust in large trials. Pooled data show investigator and central assessment agree closely in practice; the reportable gap is a manuscript that never states who assessed response, not the absence of central review itself.

Does this replace blinded independent central review?

No. A manuscript review cannot re-read scans; central review is a trial procedure, not an editorial one. The agent checks that the paper states who assessed response and under which criteria — the part authors control at submission. Pooled haematology-trial evidence found investigator and central assessment closely matched, but silence on who assessed response leaves readers unable to judge the question at all.

Denominators: which patients the rate is computed over

An objective response rate is a fraction, and the fraction is uninterpretable until the denominator is named. Three denominators are in routine use — intention to treat (everyone randomised), all treated (everyone who received at least one dose), and response evaluable (everyone with at least one post-baseline assessment) — and they are not close to each other.

Take a trial that randomises 120 patients, doses 114, obtains a post-baseline scan in 98, and observes 34 responses. The response rate is 28.3% on intention to treat, 29.8% on all treated, and 34.7% on response evaluable. That is a spread of 6.4 percentage points from one numerator, and the largest of the three is the one produced by excluding the patients who deteriorated before their first scan. Response-evaluable denominators systematically exclude early progressors and early deaths, which is a selection step applied after treatment began.

The reportable defect is rarely the choice. It is a denominator that is never stated, a denominator that changes between the abstract, the efficacy table and the discussion without explanation, or a percentage that does not reconcile with its own printed numerator. The agent recomputes every rate from the numbers on the page, which is why the brief specifies code execution rather than asking the model to do arithmetic in prose.

The choice is rarely neutral, either — though the guidance is quieter here than reviewers tend to assume, and it is worth being exact about what it does and does not say. The FDA guidance defines the endpoint as “the proportion of patients with tumor size reduction of a predefined amount and for a minimum time period” and names no analysis population at all; the single place it invokes intention to treat is the definition of overall survival, “measured in the intent-to-treat population” from randomisation. The word “evaluable” does not appear in the document. So the obligation the guidance creates is pre-specification rather than a particular denominator: the protocol names the population, and the manuscript reports the population the protocol named.

What makes response-evaluable the weakest of the three is therefore not that a regulator forbids it. It is the selection step described above — it is the only one of the three whose membership is settled after treatment began, by an event (surviving to a first scan) that is itself related to how the patient is doing. A manuscript that leads with the response-evaluable rate in its abstract, without also giving the intention-to-treat or all-treated figure, has led with its most favourable number rather than its most complete one, even where every individual percentage is arithmetically correct. The agent reports the omission and the direction; it does not assert that a particular denominator was required.

Disease control rate is where stable disease hides

Disease control rate combines complete response, partial response and stable disease into a single figure, and the FDA guidance is unusually blunt about why that figure carries less weight than its size suggests: “Stable disease should not be a component of ORR. Stable disease can reflect the natural history of disease, whereas tumor reduction is a direct therapeutic effect.”

A disease control rate of 78% in a slow-growing tumour type is compatible with no drug activity at all, because a patient whose sum of diameters moves by −8% over twelve weeks is stable by definition and would have been stable without treatment. When a discussion leads with disease control rate rather than objective response rate, the agent reports it, states the objective response rate alongside, and asks whether stable disease was required to persist for a minimum duration — a requirement that separates durable stability from a single measurement. Claims built on the combined figure are the common shape flagged by overclaim checking.

Common questions about response rate denominators

Which analysis population should a response rate use?

Whichever the protocol pre-specified, stated explicitly. The problem the agent reports is not the choice but leaving it unstated, or using different populations in different places without saying so. In a trial that randomises 120 patients, doses 114, scans 98 and observes 34 responses, the rate is 28.3%, 29.8% or 34.7% depending only on which denominator is used.

Is disease control rate the same thing as objective response rate?

No. Disease control rate adds stable disease to complete and partial response, while objective response rate counts only complete and partial response. A high disease control rate in a slow-growing tumor type can occur with no drug activity, since stable disease can simply reflect the tumor’s natural history rather than a treatment effect.

Progression-free survival censoring

Progression-free survival is defined by its censoring rules, and a paper that omits them has not reported the endpoint. The FDA guidance gives the default explicitly: “We recommend assigning the progression date to the earliest time when any progression is observed without prior missing assessments and censoring at the date when the last radiological assessment determined a lack of progression.”

The assumption underneath every censoring rule is stated in the same document and is worth quoting because manuscripts almost never acknowledge it: “any method that censors observations on patients, whether at death or at the last visit, assumes that the patients with censored observations have the same risk of recurrence as patients with noncensored observations who have not yet experienced the event.” Patients censored for toxicity, for clinical deterioration, or for withdrawal after a scan the investigator disliked violate that assumption in a known direction. Where censoring itself depends on how the patient is doing, the analysis is a competing risks and censoring problem rather than an administrative footnote, and a Kaplan-Meier curve drawn under informative censoring overstates the time to event.

Assessment schedule is the other half. The FDA guidance notes that “bias can be introduced if the frequency of long-term follow-up visits is dissimilar between the study arms”. Progression is interval-censored: it is observed at a scan, not when it happens. An arm scanned every six weeks and an arm scanned every nine weeks will differ in measured progression-free survival even with identical biology, because the later-scanned arm accrues time between the true event and its detection. The agent reports any indication that intervals differed by arm, and any hazard ratio reported without a statement that the schedule was identical.

Immortal time bias in responder analyses

Comparing survival between responders and non-responders is not a weak analysis; it is a guaranteed result. A patient must survive long enough to reach the assessment at which response is declared, so time before that assessment is time during which the patient could not have been classified a responder. Assigning that time to the responder group makes the responder curve longer regardless of whether the drug does anything. This is immortal time bias, and it appears in oncology manuscripts more often as a landmarkless “responders lived longer” sentence in a discussion than as a formal analysis.

Three named remedies exist and the agent asks which was used. A landmark analysis fixes a time point specified before looking at the data — conventionally the trial’s first scheduled restaging — classifies patients by their status at that landmark, excludes anyone who has already had the event, and starts follow-up there. A time-dependent covariate Cox model lets response status change during follow-up, so each patient contributes non-responder time until response and responder time afterwards. The Mantel-Byar method handles the same transient-state structure without discarding pre-landmark follow-up. Reporting a landmark analysis without stating the landmark, or choosing the landmark after seeing which one separates the curves, reintroduces the problem it was meant to solve. Where curves also cross, see my survival curves cross.

The landmark’s placement is a trade-off, not a formality, and a manuscript rarely states which side of it it has chosen. Set too early, before response could plausibly be assessed, it fails to exclude anyone and buys no protection. Set too late, it discards every patient who progressed or died before that visit — precisely the patients a responder-favouring analysis most needs to keep. A landmark tied to the trial’s first protocol-specified restaging avoids both failure modes because it is fixed by the protocol rather than by the data.

Reading a waterfall plot against the reported response rate

A waterfall plot is a countable claim, and counting it is the fastest check in the whole review. Every bar is one patient’s best percentage change in the sum of target lesion diameters; every bar below the −30% line is a patient who met the target-lesion criterion for partial response. So the number of bars past the line and the reported response rate must be reconcilable, and when they are not, one of the two figures is wrong or a rule was applied that the caption does not state.

Consider a plot with 62 bars, 19 of them below −30%, against a text stating an objective response rate of 34% in 68 evaluable patients — which is 23 responses. Four possibilities explain a gap of that shape, and the manuscript must say which one applies. Bars may be missing because patients without a post-baseline scan cannot appear on the plot but were counted in the denominator, in which case the plot’s N and the rate’s N differ and both should be labelled. Responses may be unconfirmed, so they cross the line on the plot and are excluded from the confirmed rate. Target-lesion shrinkage may be overridden: a patient whose target sum fell 45% but who developed a new lesion or progressed in non-target disease is a deep bar and not a responder. Or the plot shows a subset — one cohort, one dose level, or only patients with measurable disease at baseline.

The reportable defects are a waterfall plot with no −30% reference line, no stated denominator, or no statement of whether the values are confirmed. Swimmer plots carry the same obligation: a bar length is a duration, and the caption should say from what point it is measured and which bars are censored.

Frequently asked questions

Why is comparing responders with non-responders a problem?

Because a patient must survive long enough to be assessed before they can be classified as a responder. That guarantees responders have longer survival regardless of treatment effect. This is immortal time bias, and landmark analysis is the usual remedy, with a time-dependent covariate model or the Mantel-Byar method as alternatives that do not discard pre-landmark follow-up.

Can an AI reviewer check a waterfall plot against the reported response rate?

Yes, by counting. Bars below the −30% line are patients meeting the target-lesion criterion for partial response, so the count and the reported rate must reconcile. Where they do not, the manuscript needs to say why — missing post-baseline scans, unconfirmed responses excluded from the rate, new lesions overriding target shrinkage, or a plot that shows only a subset.

My waterfall plot shows 19 bars past −30% but my abstract reports 23 responses — which number is right?

Possibly both, depending on why they differ. Four causes explain the gap: patients without a post-baseline scan can be counted in the response-rate denominator but cannot appear on the plot; unconfirmed responses may be excluded from the plot but counted in the rate; a deep bar may be overridden by a new lesion; or the plot shows only a subset of patients. State which one applies.

Duration of response and where it is measured from

Duration of response has a conventional definition and two common ambiguities. The FDA guidance states the convention: “Response duration usually is measured from the time of initial response until documented tumor progression.”

The first ambiguity is the start point. In a trial that requires confirmation, “initial response” may mean the first scan meeting the criterion or the confirmatory scan four or more weeks later, and the two produce medians that differ by roughly the confirmation interval. The second is censoring: patients still in response at data cut-off are censored, so a median duration of response reported from a cohort in which most responders remain in response is an unstable estimate that will move, and reporting “median not reached” with the number still at risk is more informative than a point estimate with a wide interval. The agent reports duration of response given without a start point, without a censoring statement, or without the number of responders it was computed over — and flags a median presented as a durability claim when fewer than half the responders have had the event.

What to do when the obvious remedy is unavailable

Most response-criteria findings have a remedy that is already unavailable by the time the paper is written, and the correct response is disclosure with a stated direction, not silence.

Central review was not performed and the scans cannot be re-read. State plainly that assessment was investigator-determined, name the criteria and the version, and say who the assessors were and whether they were blinded to treatment. If the trial is randomised and blinded, say so and cite the FDA position that central verification may not be necessary in that design. Do not claim the absence of central review is immaterial by pointing at haematology concordance data if the trial is a solid-tumour study; the honest sentence is that assessment was local, which the reader can then weigh.

Responses were never confirmed because patients came off study. Report the confirmed and unconfirmed rates side by side with their denominators, rather than reporting the higher one alone. Two numbers with an explanation survive review; one number with a footnote does not.

Scans were acquired in routine care rather than on a protocol schedule. Report the observed distribution of intervals — median and range — not the intended schedule, and show a sensitivity analysis under an alternative censoring rule. Retrospective imaging cadence correlates with how sick the patient is, which is the informative-censoring case above.

The only survival comparison available is responder versus non-responder. Convert it to a landmark analysis with a pre-specified landmark, or a time-dependent covariate model, and report how many patients were excluded at the landmark. If neither is possible, remove the causal reading rather than softening it, which is what causal language discipline checks.

The response rate is the primary endpoint of a single-arm trial. Say what the rate is a surrogate for and on what evidence, since objective response rate is a surrogate endpoint whose relationship to survival is disease-specific. A limitations section that names the direction of each defect is stronger than one that lists them.

iRECIST, pseudoprogression, and mixed criteria

iRECIST exists because a minority of patients on checkpoint inhibitors enlarge before responding, and conventional criteria classify them as progressing. What that enlargement looks like has been characterised: a systematic review and meta-analysis of confirmed pseudoprogression in solid tumours (Fountoukidis G, Tina E, Göthlin-Eremo A, Ullenhag G, Valachis A, “Optimizing the clinical assessment of pseudoprogression in patients with solid tumors treated with checkpoint inhibitors: a systematic literature review and meta-analysis of associated features”, Cancer Immunology, Immunotherapy 2026;75(6):170) pooled thirteen retrospective studies, most of which applied iRECIST, and reported a pooled median time to the initial apparent progression of 2.52 months (95% CI 1.54 to 3.51), a mean tumour burden increase of 33.0% (95% CI 22.7 to 43.3), new lesions present in 35.3% of cases, and increases beyond 100% in only 3.9%. Among patients whose progression turned out to be pseudoprogression, 41.8% went on to a partial response and 6.4% to a complete response without any change of therapy.

Those numbers describe the shape of the event, not how often it happens, and the distinction is worth keeping. Reported incidence varies widely with tumour type, criteria and study design, so a manuscript that asserts a single pseudoprogression rate should be asked where the figure comes from. What the pooled features do establish is that the phenomenon is real, that a third of these cases involve new lesions — which under RECIST 1.1 is progression on its own — and that a doubling of tumour burden is rarely pseudoprogression.

The iRECIST mechanism is a two-step progression call. RECIST 1.1 progression becomes unconfirmed progressive disease, iUPD; it becomes confirmed progressive disease, iCPD, only if imaging performed four to eight weeks later shows a further increase — at least 5 mm in the sum of target measures, or progression in a lesion category that had not previously progressed. New lesions are recorded separately: “five lesions (no more than two per organ) should be measured and recorded as a new lesion target, but should not be included in the sum of measures of the original target lesions identified at baseline.”

Two failures follow. An immunotherapy trial applying conventional RECIST 1.1 with no discussion of pseudoprogression may be counting eventual responders as progressions, shortening progression-free survival. A trial applying iRECIST for the primary endpoint while an exploratory analysis uses RECIST 1.1 — or the reverse — produces two progression-free survival curves in one paper that are not the same quantity, and the agent reports which criteria govern the primary endpoint and where the manuscript is ambiguous. Where the primary endpoint bundles progression with death or symptomatic deterioration, that is a composite endpoint and its components should be reported separately as well.

What reviewers write when this is mishandled

Reviewers rarely write “response criteria were applied loosely”. They write the specific version, and these are the comments that decide oncology efficacy papers.

“Please state which version of RECIST was used and whether responses were confirmed.” “The response rate in Table 2 is computed over response-evaluable patients while the abstract reports it as the rate in the treated population; please reconcile.” “The waterfall plot shows 19 patients below the −30% threshold and the text reports 23 responses; please explain the discrepancy.” “Censoring rules for progression-free survival are not described, and patients discontinuing for toxicity appear to have been censored at their last scan.” “Tumour assessment intervals differ between arms; please provide a sensitivity analysis using a common assessment schedule.” “The comparison of survival between responders and non-responders is subject to immortal time bias; a landmark analysis is required.” “Response was investigator-assessed in a single-arm study; please justify the absence of independent review.” “Disease control rate is emphasised in the discussion; please report the objective response rate with its confidence interval.” “Duration of response is reported without stating the point from which it was measured.”

That last set is why the agent reports numbers rather than adjectives. A reviewer comment naming two figures and a discrepancy is answerable in a revision; a comment saying the endpoint reporting is unclear is a rejection with extra steps.

What this agent catches

How this differs from the built-in review

This response criteria agent checks a number against its definition; PerfectPaper’s standing review team checks a number against other numbers on the page. The standing team includes a numeric integrity specialist that recomputes reported arithmetic and a table structure specialist that checks internal consistency. Those will catch a percentage that does not match its numerator. They will not catch a response rate that is arithmetically perfect and computed over the wrong population, or an unconfirmed response counted as a response, because both require knowing what the criteria define.

The distinction is worth stating precisely, because it determines whether you need this agent at all. A numeric check operates on the relationship between numbers printed on the page. A criteria check operates on the relationship between a number and the definition it claims to instantiate — and that definition lives in RECIST 1.1, iRECIST or the protocol, not in the manuscript. PerfectPaper reads the methods, the efficacy tables and the figures of an oncology manuscript together, recomputes every reported rate from its own numerator and denominator, and reports where the rate no longer means what the criteria define.

Pairs with early-phase safety reporting for a paper reporting both efficacy and toxicity. Full set: AI peer review for clinical trials.

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.

• Early-phase safety and dose reporting — use it when the dose-limiting toxicity definition or its window is unstated.
• Registration and pre-specification — use it to catch a reported primary endpoint that differs from the registered one.
• 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.