How a manuscript is reviewed

A protocol, not a prompt

A review is not a single question put to a language model. It is a sequence of passes defined in a protocol file that ships with the service and carries a version number. The protocol guides how the manuscript is examined; it never predefines the findings. Two runs over the same paper follow the same stages, apply the same severity rubric, and are held to the same convergence criteria.

The six passes

Each pass reads the manuscript with a different question in mind. The complete raw output of every pass is preserved, not only the final review.

Comprehensive review

A whole-paper pass over the extracted text. It establishes what the study claims, what design it used, and where the argument carries its weight, before any detail is examined.

Visual page windows

The manuscript is re-read as overlapping windows of rendered pages, so figures, tables, axes, captions and typeset equations are seen as printed rather than as extracted text. The overlap makes every hinge page appear in two windows, so nothing falls between them.

Study completeness audit

A check for what is absent: missing reference lists, unresolved citation placeholders, placeholder repository links, absent sample sizes, and reporting items the study design requires.

Full-paper synthesis

Findings from the earlier passes are reconciled against each other and against the paper as a whole, so that a detail flagged in one section is judged against what the rest of the manuscript says.

Adversarial pass

The draft review is attacked. Every finding must survive the question of whether it is actually supported by the manuscript, or whether it is a plausible-sounding criticism that the text does not justify.

Validation until convergence

Validation passes repeat until the review stops changing. A pass that keeps discovering material issues is a pass that has not finished reading.

When a review is finished

A run stops when all of the following hold, or when the configured iteration limit is reached.

• No new material critical or major issue is identified.
• Unsupported claims and disagreements are resolved, or explicitly retained with a reason.
• Every rendered comment has page-grounded evidence or an explicit visual region.
• Two consecutive validation passes make no material change.

Nothing is drawn where the evidence does not match

A comment is placed on the PDF only when it is anchored to an exact quote long enough to identify a location on that same page. If the quote cannot be matched, the finding is not drawn. It is not silently dropped either: it moves to an audit of unplaced comments, with the reason it could not be anchored. A reviewer can therefore tell the difference between a finding the tool could not place and a finding it did not make.

This refusal is the point. An annotation drawn in approximately the right place is worse than no annotation, because it invites the author to fix the wrong sentence.

Internal and external verification are kept apart

Manuscript-internal checks, whether the abstract agrees with the results, whether table values match the text, are separated from external factual verification. Cited DOIs are resolved against Crossref, so a reference that does not exist is reported as a reference that does not exist, and not as a stylistic remark.

What each pass looks for

Internal consistency

Duplicate or contradictory conclusion sections; repeated values for the same quantity that disagree across abstract, methods, results, tables and appendices; arithmetic in PRISMA and search totals; stated list counts against the number of items enumerated; sample sizes across characteristics, outcomes, safety tables and figure captions.

Statistical and scientific method

Whether the estimator matches the estimand; causal assumptions, including exchangeability, positivity, consistency, interference and measurement; variance estimators, robust corrections, small-sample degrees of freedom and multiple-comparison handling; whether hand-coded implementations are justified where validated packages exist; whether the conclusions are stronger than the design supports.

Tables, figures and style

Text claims against table and figure values where they can be extracted; point estimates, confidence and prediction intervals, axes, legends, labels and captions; unreadable, cropped or insufficiently explained figures; table notes and appendices that contradict the body; vague or inflated wording replaced with precise claims.

Severity rubric

Critical

Findings that make the manuscript unverifiable, internally contradictory, or scientifically unreliable.

Major

Findings that require substantive correction before the result can be interpreted confidently.

Moderate

Findings that materially improve reproducibility, clarity or interpretation, but do not invalidate the core result.

Minor

Editorial and presentation issues that improve scientific communication.

Failure classes it is built to catch

These are concrete defects that reach published papers. The protocol names them so that a pass cannot consider itself complete without having looked.

• Two contradictory conclusions in the same manuscript.
• The same interval reported with different bounds in different places.
• All citations rendered as unresolved placeholders.
• No reference or bibliography section at all.
• Table values that disagree with the forest plots drawn from them.
• PRISMA totals that do not add up.

Every run is auditable

Each review returns a JSON evidence bundle alongside the Markdown review and the annotated PDF. It records the document manifest, the raw output of every pass, the comments that could not be placed and why, and the result of each external verification. A finding you disagree with can be traced to the pass that produced it and the page it was anchored to.

What this is not

This is not peer review, and it does not decide whether a paper should be published. It does not reproduce an analysis, rerun code, or access data that is not in the file you upload. It reads one manuscript at a time and reports what that manuscript does and does not support. Treat the output as a demanding reader's notes taken before submission, not as an editorial decision.

Read your manuscript the way a reviewer will

One review is included when you create an account. No card required.