A custom AI reviewer for early-phase safety and dose reporting

PerfectPaper’s early-phase safety agent reads a dose-finding manuscript for the definitional and reconciliation failures that dominate review of these papers: a dose-limiting toxicity defined loosely or not at all, an unstated grading version, safety denominators that change between tables, and a recommended phase 2 dose asserted rather than justified by the escalation data shown. Each finding names the dose level and the table.

In a phase 1 report the definitions are the science. A dose-limiting toxicity that is not defined precisely makes the entire escalation uninterpretable, however carefully it was conducted. Copy the brief below into a custom agent slot.

When to use this agent

• Your manuscript reports a first-in-human or dose-escalation study
• The design is 3+3, BOIN, CRM, or another model-based scheme
• You are recommending a phase 2 dose
• Adverse events are reported across several tables with different populations
• Some patients were replaced during escalation
• A dose-expansion cohort sits in the same paper as the escalation, so two populations share one safety section
• Accrual ran long enough that adverse events may have been graded under two versions of the same grading system

What CONSORT-DEFINE asks a dose-finding paper to report

CONSORT-DEFINE is the reporting standard written for exactly these manuscripts, and it is recent enough that most submitted papers were designed before it existed. Yap and colleagues published the CONSORT Dose-finding Extension in BMJ 2023;383:e076387, adding 21 new checklist items and modifying 19 existing CONSORT items for early phase dose-finding trials; the explanation and elaboration document followed in EClinicalMedicine 2025;79:102987.

Two modified items carry most of what this agent automates. Item 13a asks for, verbatim, “For each group, the number of participants who were assigned to each dose level at each interim analysis (eg, for dosing decisions), received intended treatment, and were analysed for the primary outcome and, if applicable, any other outcomes used to inform prespecified adaptations.” Item 19 asks for “All important harms (eg, adverse events or effects, toxicities) reported by dose level in each group (for specific guidance, see CONSORT for harms).” Assigned, treated and analysed are three different denominators; item 13a exists because dose-finding papers habitually report one number where three are needed.

The extension’s own glossary defines a dose-limiting toxicity in general terms only — “Side effects of an intervention that are serious enough to prevent an increase in the dose of that intervention” — which is the right gloss for a reader and the wrong level of precision for a methods section. A manuscript has to supply the operational version: which grades, in which organ systems, over which window, with which named exceptions. The parent statement this extends is covered in the CONSORT checklist.

The agent brief

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

Name: Early-phase safety and dose reporting

You are reviewing an early-phase dose-finding study for completeness and
internal consistency of its safety and dose reporting. Read the methods,
every safety table, and the results text together.

Determine and report:

1. Dose-limiting toxicity definition. Determine whether the manuscript defines
   what constitutes a dose-limiting toxicity, including which grades and which
   organ systems, and any exceptions such as manageable haematological events.
   Report a study that reports dose-limiting toxicities without defining them
   as having an uninterpretable escalation.

2. The observation window. Determine the period during which a toxicity counts
   as dose limiting, and whether it is stated. Report its absence. Report any
   toxicity discussed as dose limiting that occurred outside the stated window.

3. Evaluability. Determine the criteria for a patient to be evaluable for
   dose-limiting toxicity, how many patients were replaced, and why. Report
   replacements that are mentioned without criteria, since replacement rules
   materially change escalation behaviour.

4. Grading. Determine which adverse event grading system and version was used.
   Report a version that is not stated.

5. Attribution. Determine whether adverse events are reported as all-cause or
   as treatment related, and whether the attribution method is described.
   Report tables that mix the two, and report any place where the text
   describes a rate that a table does not support.

6. Denominators. For every safety table and every rate in the text, determine
   the population it is computed over. Report every denominator that changes
   between tables without explanation, and recompute each percentage against
   its stated numerator and denominator, reporting any that does not
   reconcile. Treat this as high priority; it is the most common finding.

7. Grade reporting. Determine whether all-grade and grade 3 or higher events
   are distinguished consistently. Report any summary statement that does not
   make clear which it refers to.

8. Recommended dose. Determine whether a recommended phase 2 dose is stated
   and what justifies it. Report a recommendation that rests only on the
   absence of dose-limiting toxicity, without reference to pharmacokinetic
   exposure, target engagement, or cumulative and late toxicity. Report where
   the recommended dose is not the highest dose tested without an explanation.

9. Escalation account. Determine whether the number of patients treated at
   each dose level, and the number of dose-limiting toxicities at each, are
   reported. Report their absence, since without them the escalation cannot be
   followed.

Use code execution to reconcile patient counts across dose levels and tables.

What this agent catches

Dose-limiting toxicity is a protocol definition, not a description

A dose-limiting toxicity is whatever the protocol says it is, so a manuscript reporting the count without the definition has reported a number with no unit. The definition is not a convention shared across trials; it is a decision each protocol makes, and reviewers of dose-finding work read the methods for it before they read anything else.

That is a measured claim rather than a rhetorical one. Le Tourneau and colleagues reviewed 155 phase 1 trials of 111 molecularly targeted agents and found that “for any given toxicity, there was substantial variability in the degree of severity required”, with duration, reversibility and dose-modification criteria specified only rarely; discrepancies between the initial and the final definition of a dose-limiting toxicity were reported in 25% of trials (European Journal of Cancer 2011;47(10):1468–75).

A complete definition has four parts: a grade threshold — usually grade 3 or higher for non-haematological events, and commonly a grade higher than that for haematological ones, since some reversible cytopenia is expected on most regimens; the organ systems in scope; the named exceptions; and the assessment window. The exceptions carry most of the between-trial variation. Common carve-outs include grade 3 nausea, vomiting or diarrhoea that resolves to grade 1 or better within 48 or 72 hours on standard supportive care; grade 4 neutropenia lasting fewer than 5 to 7 days without fever; grade 3 laboratory abnormalities with no clinical consequence and no intervention; alopecia at any grade; and isolated lymphopenia. Two trials of the same molecule with different exception lists will derive different maximum tolerated doses from identical patient data.

The arithmetic downstream is unforgiving. Under a 3+3 design, 0 of 3 patients with a dose-limiting toxicity escalates, 1 of 3 expands the cohort to 6, 2 or more stops escalation, and the maximum tolerated dose is conventionally the highest dose at which no more than 1 of 6 patients had one. Whether a single grade 3 transaminase rise that resolved in four days counts therefore decides the dose the paper recommends. Report the rule, not only the outcome of applying it.

Common questions about dose-limiting toxicity definitions

What exactly counts as a dose-limiting toxicity?

A dose-limiting toxicity is whatever the protocol defines it to be, not a fixed clinical category: a grade threshold (usually grade 3 or higher for non-haematological events), the organ systems it applies to, named exceptions such as short-lived grade 3 nausea, and an assessment window. A manuscript reporting dose-limiting toxicities without stating all four has reported a number with no unit.

Do haematological and non-haematological toxicities use the same grade threshold?

Usually not. The more common convention sets a higher bar for haematological events — grade 4 rather than grade 3 for neutropenia or thrombocytopenia — because some reversible cytopenia is expected on a cytotoxic regimen, and then qualifies that further by duration or by accompanying fever or bleeding: grade 4 neutropenia lasting fewer than 5 to 7 days without fever, or grade 3 thrombocytopenia counting only where it comes with bleeding. It is a convention rather than a rule, though, and a substantial minority of protocols do apply the same grade 3 threshold to both classes. That is precisely why the threshold has to be stated per class instead of assumed: a reader cannot recover which convention you used from the words “grade 3” alone.

What does a reviewer check first in early-phase safety reporting?

The dose-limiting toxicity definition and its assessment window, before any result. After that: the grading system and version, whether tables report all-cause or treatment-related events, whether the denominators reconcile across abstract, tables and text, and whether the number of patients and dose-limiting toxicities is given at each dose level.

The observation window decides what the escalation could have seen

The dose-limiting toxicity observation window is the period during which a toxicity counts toward an escalation decision, and it bounds what the escalation was capable of detecting. Stating it in days from the first dose is the whole requirement, and its absence is one of the two most common findings this agent returns.

Conventionally the window is cycle 1 — 21 or 28 days for most schedules. For a cytotoxic given every three weeks, that matches the toxicity’s natural timing. For a kinase inhibitor taken continuously, for an immune checkpoint agent whose immune-related events arrive weeks to months after the first dose, or for any agent with cumulative organ toxicity, a cycle-1 window systematically excludes the toxicity that will limit the drug in practice. The escalation is not wrong; it is answering a narrower question than the recommended dose implies.

The tell in a manuscript is a discussion section describing grade 3 or higher events at months 3 to 6 that were never candidates for an escalation decision. Without a stated window, neither a reader nor this agent can tell whether those events were excluded by rule or overlooked. Report the window, report whether it differs by dose level or schedule, and report late and cumulative toxicity separately with the duration of follow-up, saying plainly that those events did not inform escalation. Where the design handles late onset by construction — a time-to-event continual reassessment method, or a time-to-event BOIN variant — name it, because a reviewer cannot infer it from the results. Where toxicity is analysed as time-to-event, the competing risks and censoring questions apply to it as they do to efficacy.

Questions about the DLT observation window

My patient had a grade 3 event in month four — does that count as a dose-limiting toxicity?

Only if the protocol’s dose-limiting toxicity window extends that far, and most do not: a cycle-1 window of 21 or 28 days is conventional and will not reach month four. Report the event as a late or cumulative toxicity with its timing stated, and say plainly that it did not inform the escalation decision — it is not thereby unimportant.

Why does the observation window matter if the trial already found its maximum tolerated dose?

Because the window defines what the escalation could see, not what the drug does. A cycle-1 window fits a cytotoxic agent given every three weeks but systematically misses the late, cumulative or immune-related toxicity of a continuously dosed or immunotherapy agent, so a maximum tolerated dose derived under it answers a narrower question than the paper implies.

Evaluability and replacement change the escalation silently

Evaluability criteria and replacement rules decide whose toxicity counts, so a paper that mentions replacement without stating the rule has not reported its escalation. This agent treats a replacement count with no criteria as a finding in its own right, not as a detail.

A patient who withdraws early, progresses before the window closes, or receives less than a specified proportion of the intended dose — commonly at least 75% of intended dose during cycle 1 — is typically declared non-evaluable and replaced. Replacement is not neutral. If a patient comes off study early because of an emerging toxicity, and early discontinuation makes them non-evaluable, the rule removes exactly the patients whose events carried the most information about the dose. That is conditioning on something the treatment influenced, the structure described under selection bias, operating inside the escalation rule rather than at enrolment.

Report the evaluability definition, the number replaced at each dose level, and the reason in each case. “Three patients were replaced” is not a report; a reviewer needs to know whether any of the three left because of toxicity. Note also that a replaced patient remains in the safety population while leaving the dose-limiting toxicity population, which is the first place a paper’s denominators diverge and the point at which the reconciliation in the brief above starts paying.

Which grading version, and why an unstated one is now ambiguous

The adverse event grading version must be named, and in 2026 an unstated version is genuinely ambiguous rather than merely untidy. The US National Cancer Institute lists Common Terminology Criteria for Adverse Events versions 3.0 (2006), 4.03 (2010), 5 (2017) and 6 (2025), with implementation of CTCAE v6 for CTEP and DCP studies targeted for January 1, 2026 and applying to newly opened studies only.

Manuscripts submitted through 2026 and 2027 will therefore be split across CTCAE v5.0 and CTCAE v6, and the terms and grade boundaries are not identical between versions. “Graded per CTCAE” with no version leaves severity uncomparable to earlier trials, to concurrent trials, and to the paper printed next to it. Where a long accrual straddled a version change — which is common in a trial that ran three years — state which patients were graded under which version and whether any events were regraded.

Name the system and the version once in the methods, in the literal form a reader would search for, such as “CTCAE v5.0”. Keep it out of individual table captions, where it drifts. If adverse event terms were coded with MedDRA in addition to being graded with CTCAE, name the MedDRA version too, since preferred terms move between versions and a system organ class summary is not stable without it.

Questions about CTCAE grading and safety denominators

Which CTCAE version should I use if enrollment started before the version changed?

Report both. CTCAE v6 implementation for CTEP and DCP studies applies to newly opened studies from January 1, 2026 only, so a trial that enrolled across the change should state which patients were graded under which version and whether any events were regraded, rather than naming a single version for the whole cohort.

Why do the adverse event percentages in my tables not match my abstract?

Almost always because the tables use different denominators — all enrolled, all treated, all evaluable — computed at different points in the trial, not because either number is wrong on its own. Recompute every reported rate against its stated numerator and denominator before submission; this is the single most common finding in dose-finding manuscript review.

Do I need to report all-cause and treatment-related adverse events separately?

Yes, and consistently. Both conventions are legitimate individually, but a safety section that switches between them without saying so — or a table headed “treatment-related” whose totals match an all-cause table elsewhere — cannot be read. State which scale was used, who applied it, and whether “possibly related” was collapsed into “related.”

Attribution: all-cause, treatment-related, and who decided

Attribution decides which events appear in which table, and a safety section that switches between all-cause and treatment-related reporting without saying so cannot be read. Both conventions are legitimate; mixing them inside one manuscript is not.

All-cause reporting counts every adverse event occurring in a treated patient regardless of cause. Treatment-related reporting counts only the events an assessor judged related, which is a graded clinical judgement recorded on an ordinal scale and then collapsed to a binary for the table. Three things routinely go unreported: the scale used, who applied it — site investigator, sponsor medical monitor, or an independent safety committee — and whether the collapse treated “possibly related” as related. That last choice alone can move a related-event rate by several patients in a cohort of thirty.

The pattern to look for is a table headed “treatment-related adverse events” whose totals match an all-cause table elsewhere in the same paper, or a results sentence quoting a rate that no table produces. Because dose-limiting toxicity assessment is itself attribution-dependent in most protocols, an unstated attribution method propagates from the safety tables into the escalation decisions. Note too that “treatment-emergent” is a timing concept, not a causal one; using it interchangeably with “treatment-related” is the same category error the causal language agent flags in efficacy sections.

Worked example: a denominator that does not reconcile

Denominator drift is the most common finding this agent returns, and it presents as arithmetic rather than as an error anyone would defend. A worked case shows how ordinary it looks.

A first-in-human report enrols 34 patients, treats 32, and describes 28 as evaluable for dose-limiting toxicity after four replacements. Table 1 gives baseline characteristics for 32. Table 2, headed “treatment-emergent adverse events”, reports grade 3 or higher neutropenia in 7 patients, 21.9%. Table 3, “dose-limiting toxicities by dose level”, reports 3 events among 28. The abstract states that grade 3 or higher neutropenia occurred in 25% of patients.

Recompute: 7/32 is 21.9%, so Table 2 uses the treated population; 25% is 7/28, the evaluable population. Neither figure is wrong on its own, and no reviewer can see the switch without doing the division. Two rates for the same event in the same trial now enter the literature, and a later meta-analysis will pick one. Recomputing every percentage against its stated numerator and denominator is the check, and it is precisely what code execution in a custom agent is for.

Precision is the second half of the same problem. With three to six patients per dose level, one event moves a rate by 17 to 33 percentage points, so a percentage carried to one decimal claims a precision that six patients cannot support. Report counts with the denominator in the row or the caption, and where an interval is given, use an exact binomial one — see confidence intervals — rather than a normal approximation that can return a lower bound below zero.

Justifying the recommended phase 2 dose

The recommended phase 2 dose is the paper’s main deliverable, and the absence of dose-limiting toxicity at the highest dose tested does not justify it. This is now the position of the regulator as well as of reviewers.

The US Food and Drug Administration issued the Project Optimus guidance, Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases, as final guidance in August 2024. It is “intended to assist sponsors in identifying an optimized dosage(s) for human prescription drugs or biological products for the treatment of oncologic diseases during clinical development prior to submitting an application for approval for a new indication and usage.” The premise of the document is that a maximum tolerated dose derived from a small escalation is not by itself an adequate basis for the dosage carried forward.

A defensible justification names at least five things: pharmacokinetic exposure at each dose level and whether it plateaus; a pharmacodynamic or target-engagement measure and the dose at which it saturates; cumulative and late toxicity across the whole treated period rather than the window alone; dose intensity actually delivered, meaning reductions, interruptions and discontinuations by dose level; and any efficacy signal by dose level, where one exists. Where the recommended dose is lower than the highest dose tested, say why — that is a decision, and unexplained it reads as chosen after seeing the data, which is the shape the overclaim check is built to find. Where the argument runs from preclinical exposure, translational dose relevance is the companion check.

Not every escalation reaches a dose-limiting toxicity at all, and that absence is not automatically good news. For a targeted agent, a monoclonal antibody, or an immune checkpoint inhibitor, escalation sometimes stops because a pharmacodynamic or target-engagement marker has saturated, because manufacturing or feasibility set a ceiling, or because a protocol-specified maximum was reached — not because toxicity appeared. A manuscript that reports “no maximum tolerated dose was reached” without saying which of these stopped escalation is describing a ceiling as though it were a null result, and the same five-part justification above still applies to it.

Common questions about the recommended phase 2 dose

Reviewer 2 says my recommended phase 2 dose isn’t justified — what are they asking for?

More than the absence of dose-limiting toxicity at the top dose tested. A defensible justification names pharmacokinetic exposure by dose level, a pharmacodynamic or target-engagement measure, cumulative and late toxicity across the full treatment period, dose intensity actually delivered, and any dose-level efficacy signal. Add whichever of these your data support, and say so explicitly rather than restating the escalation result.

Does the recommended phase 2 dose have to be the highest dose tested?

No, and under the FDA’s Project Optimus guidance, final in August 2024, it increasingly should not be, where pharmacokinetic exposure plateaus or target engagement saturates below the maximum tolerated dose. Where the recommended dose is lower than the highest tested, state why; left unexplained, it reads as a choice made after seeing the results.

Can AI review a phase 1 dose-escalation manuscript?

Yes, for the definitional and reconciliation failures that dominate review of these papers. PerfectPaper’s early-phase safety agent checks the dose-limiting toxicity definition, the observation window, evaluability and replacement rules, the grading version, attribution, and every safety denominator, and recomputes each reported rate. It complements expert review rather than replacing it.

How do I check dose escalation reporting before submitting?

Work down the nine numbered checks in the brief above against your own manuscript, then confirm every rate divides. Read the methods for a dose-limiting toxicity definition with four parts — grade threshold, organ systems, exceptions, window — and confirm the paper contains a table giving patients treated and dose-limiting toxicities at each dose level, as CONSORT-DEFINE item 13a asks.

The escalation account, dose level by dose level

The escalation account is a small table, and its absence is the single reason a reviewer most often cannot follow a dose-finding paper. It is CONSORT-DEFINE item 13a rendered as rows.

The minimum columns are dose level, dose and schedule, number assigned, number treated, number evaluable for dose-limiting toxicity, number with a dose-limiting toxicity, and the decision the escalation rule then took. Include de-escalations, dose-level backfills and any intra-patient escalation, each with the rule that permitted it. If the design is model-based, report the model’s estimate of the toxicity probability at each decision point together with the prior: a continual reassessment method or BOIN run whose prior skeleton and target toxicity rate are unreported cannot be reproduced or audited, and a target rate of 0.30 behaves very differently from one of 0.20.

Keep expansion cohorts out of the escalation denominators unless you say otherwise. Where a dose-expansion cohort appears in the same manuscript, state the point at which its patients were pooled with the escalation patients for safety reporting, and give both figures. Papers that report an escalation and an expansion under one safety table are the reason the reconciliation step above ranks as high priority in the brief.

What reviewers write when dose reporting is thin

Reviewers rarely write “the safety reporting is inadequate”. They write the specific version, and these are the comments that hold papers in revision.

“Dose-limiting toxicity is reported but never defined; please state the grade threshold, the organ systems, the exceptions and the assessment window.” “It is not clear whether Table 2 reports all-cause or treatment-related events.” “The denominators in Tables 1, 2 and 3 differ and no reason is given.” “Four patients were replaced; please state the replacement criteria and the reason in each case.” “The adverse event grading version is not specified.” “The recommended phase 2 dose is stated without pharmacokinetic or pharmacodynamic support.” “Please report the number of patients treated and the number of dose-limiting toxicities at each dose level.” “Late toxicity beyond cycle 1 is described in the discussion but did not inform escalation; please make that explicit.” “The abstract reports a rate that I cannot reproduce from any table.”

Almost every one of these is answerable from data the authors already hold, which is why finding them before submission is worth more than answering them afterwards in a response to reviewers. The exception is the recommended-dose objection, which sometimes cannot be answered at all.

When the obvious remedy is unavailable

Some findings in a dose-finding manuscript cannot be fixed by supplying the missing data, because the data were never collected; the remedy is then to bound the claim rather than to restate it. Reconstructing a definition after the fact is not one of the options.

The protocol never specified a dose-limiting toxicity definition tighter than the one reported. State what was applied in practice, state that it rested on investigator judgement, and remove any language implying the maximum tolerated dose was formally determined. “Highest dose administered” is an accurate term where “maximum tolerated dose” is not.

No pharmacokinetic sampling was performed. The recommended dose then cannot be justified on exposure, and the honest form is a recommendation labelled provisional with a stated route to confirming it. That belongs in the limitations section written as a specification, naming what would settle the question, not as an apology.

Per-level numbers are too small for any rate. Report counts and stop. Three patients support “0 of 3” and do not support “0%”, and no statistical power argument rescues a design that was never intended to estimate a rate.

The escalation deviated from the stated design. Report the deviation and the decision actually taken at each point. A design reported as followed that the trial did not follow is a more serious finding than a deviation reported plainly, and reviewers who spot the discrepancy in the dose-level table treat it that way.

Accrual straddled two grading versions. State the split by patient or by date. A single-version claim that the timeline contradicts will be checked.

How this differs from the built-in review

The standing team includes cohort accounting and table integrity specialists that reconcile numbers across a manuscript, and they will find a total that does not add up. What they cannot supply is the domain knowledge that a dose-limiting toxicity is a defined term with a window and evaluability criteria, or that a recommended phase 2 dose needs more support than the absence of toxicity. Those are conventions of dose-finding rather than arithmetic. That gap is the general case for custom reviewer agents: the standing team knows how manuscripts work, and a custom agent supplies what one subfield takes for granted.

PerfectPaper reads the methods, every safety table and the results text of a dose-finding manuscript together, recomputes each reported rate against its stated numerator and denominator, and reports where the recommended phase 2 dose rests on nothing but the absence of dose-limiting toxicity.

Pairs with trial registration and endpoint pre-specification. 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.

• Response criteria and endpoints — use it for confirmed responses, censoring rules and waterfall consistency.
• 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.