Purplelink
← Paper Review

Example report

What a Paper Review report looks like

This is an example report for a fictional machine-learning manuscript, shown the way you would read it, with a note after each section on what it is for and how to act on it. A real report quotes your own text, and also comes as a Markdown file and an annotated PDF with each comment on the page it refers to.

Critical blind spots

  • Hyperparameter search asymmetry favors the proposed method.
    Quoted: "We tuned learning rate and batch size on the validation set; baselines used reported defaults."
    Surfaced by: Methodology Critic, Statistical Skeptic (consensus).
    Fix: Match search budget for each baseline; report the protocol in the methods section.

  • Sample-size justification missing for the human evaluation (Experiment 2).
    Quoted: "We recruited 24 annotators ... yielding sufficient statistical power."
    Surfaced by: Statistical Skeptic.
    Fix: Replace the claim with a sensitivity analysis: the smallest effect 24 annotators can detect at 80% power. Report confidence intervals for the preference rates, not p-values alone.

What this section is for. These are the problems a reviewer is most likely to raise. Each one quotes the passage it is about, names which of the seven reviewers raised it, and says what would fix it. Findings raised by more than one reviewer are marked as consensus and listed first.

Data-to-claim contradictions

  • Claim: "Our method beats all baselines by a wide margin."
    Evidence: Table 2 shows the proposed method tied with Baseline C within reported standard errors on 4 of 7 datasets.

What this section is for. These are places where the text claims more than the tables and figures show. Reviewers catch them quickly, and the fix is usually a change of wording, not new work.

Rectification checklist

  • A Add hyperparameter search protocol to methods (page 5, Section 3.2). Match search budget across all baselines.
  • A Replace "wide margin" framing in abstract and Section 5 with what the data actually shows.
  • B Add per-seed variance to all results tables.
  • B Reference [17] only weakly matches its CrossRef entry. Check the title and DOI.
  • C Figure 3's y-axis breaks 0-60 then jumps to 90; redraw on a continuous scale.

What this section is for. This section lists the same findings as a to-do list. A items are the ones a reviewer would reject over. B items weaken the paper. C items are polish. Work through the A items first.

True novelty estimate

The core architectural contribution is novel and not covered by [11] or [23]. The empirical comparison, as currently framed, oversells the gap. Marginal advance. The proposed mechanism is interesting, but the published evidence does not yet warrant the strength of the abstract's claim.

What this section is for. This section gives a plain read of what is new in the paper and whether the evidence supports how strongly it is claimed. Use it when you revise the abstract and the cover letter.

Reference verification summary

47 of 52 references verified against CrossRef. 3 had weak title matches (re-check [12], [17], [31]). 2 had no DOI and could not be located by title. Verify these are correctly cited.

What this section is for. Every reference is looked up in CrossRef. Weak matches and unresolvable entries are listed so you can check them, which catches typos and citations an AI tool invented.

Citation support audit

Verdicts: 18 supported, 4 partially supported, 2 not supported by abstract, 1 contradicted, 3 source unavailable.

Claim Cited ref Verdict What the source's abstract says
"Method X reduces error by 40%." [14] Contradicted Abstract reports a 4% reduction, not 40%. Verify the figure.
"Prior work proved this is optimal." [9] Not supported by abstract Abstract describes a heuristic; optimality may be in the full text.

5 citations with weaker claims were not audited (per-run cap).

What this section is for. For the claims that lean hardest on a citation, the cited paper's abstract is fetched and compared with what you wrote. "Not supported by abstract" means you should check the full text. It does not mean the citation is wrong.

How this report was made

A chat assistant reads your PDF once and answers in one pass. Paper Review looks every reference up in three scholarly databases, recomputes your statistics without a language model, and has seven separate reviewers read the paper before one pass reconciles what they found.

One chat prompt

One model, one pass

  1. Your PDF
  2. One model
  3. One answer
  • References are not looked up one by one, so a wrong or invented one can pass.
  • Retractions are caught only if the model happens to know them.
  • p-values and means are not recomputed.
  • One point of view, written in one go.
  • No marked-up PDF.
  • Whether the manuscript is kept or used for training depends on your account settings.

Paper Review

Four stages, many checks, run for every paper

  1. Read the paper apart

    The PDF is split into text, figures, tables and the reference list.

  2. Check it, in parallel
    • Lookups Every reference
      • CrossRef
      • OpenAlex
      • Semantic Scholar
      • Each reference matched to a real record
      • Retractions, withdrawals and expressions of concern flagged
      • Key claims checked against the cited source's abstract
    • Rule-based, no AI Numbers and structure
      • p-values recomputed from t, F, chi-square, r and z (statcheck)
      • Reported means tested for consistency (GRIM)
      • In-text citations matched to the reference list
      • Figure and table numbering checked
    • Vision Every page and figure
      • Pages rendered and read as images
      • Figures compared with what the text claims
      • Missing error bars, broken axes, color-only encoding
    • Panel Seven reviewers, separately
      • Methodology
      • Statistics
      • Data integrity
      • Editor
      • Equations
      • Literature
      • Numerical realism
  3. Reconcile

    Findings are merged, and issues raised by more than one reviewer rank first. A coherence check compares the abstract, hypotheses and results with each other.

  4. Deliver

    A report with a fix-first checklist and an annotated PDF, by email. Your files are deleted within about a day.

The reviewers and the final pass use Claude Opus through Anthropic's API, which does not train on your manuscript.