AI Review, agent by agent
The 9 agents behind AI Review
AI Review splits a manuscript review into 9 specialist checks that run at the same time. Of those, 8 are language-model reviewers — 7 of them calibrated with real examples from open peer review for their own specialty, and Research Gap checking the stated gap against comparable published papers instead — and Prior Publication is a deterministic lookup across 7 external sources and our 7.9M-paper local library. A synthesis step then writes one report. A full review costs 30 credits.
What each one checks
One narrow job per agent
Each language-model agent carries only its own rubric, so it can point at specifics — a section, a table, a sentence — instead of generalising. The Prior Publication agent uses no language model at all.
Methodology
Language model · calibrated on open-review examplesWeight ×2.0 in the overall scoreAudits study design, statistical power, and analytical choices against field-specific rigour standards (CONSORT, STROBE, PRISMA).
What it looks for
- Study design against reporting standards such as CONSORT, STROBE and PRISMA
- Sample-size justification, statistical power, and whether the test suits the data
- Red flags: data leakage, pseudoreplication, HARKing, causal claims from correlational data
Formulas & Equations
Language model · calibrated on open-review examplesWeight ×1.2 in the overall scoreVerifies mathematical derivations, checks dimensional analysis, and flags algebraic errors.
What it looks for
- Dimensional consistency of each equation's terms
- Undefined variables and one symbol used for two quantities
- Values in the text that disagree with the equations or tables, with a suggested correction
Originality
Language model · calibrated on open-review examplesWeight ×1.5 in the overall scoreChecks the manuscript's novelty claims against papers in our 7.9M-paper local library and flags possible overlap and self-plagiarism signals.
What it looks for
- Novelty claims set against similar papers in the local library
- Incremental versus genuinely new contribution, and overclaiming such as "first ever"
- Self-plagiarism, salami-slicing and duplicate-submission signals in the text
Literature Coverage
Language model · calibrated on open-review examplesWeight ×1.2 in the overall scoreEvaluates citation completeness, missing seminal references and self-citation balance, with a live OpenAlex snapshot of the field (volume, top venues, peak year) for context.
What it looks for
- Missing foundational and directly competing work
- Self-citation balance and recency for the field
- Miscitation and weak sources for key claims
Reads the reference list it can see. With LaTeX, attach your .bib or .bbl so it reads that too.
Reproducibility
Language model · calibrated on open-review examplesWeight ×1.0 in the overall scoreInspects code availability, dataset accessibility, and sufficiency of methods detail for independent replication.
What it looks for
- Where the data and code live: a repository, DOI or accession number, not "available on request"
- Software versions, hardware, random seeds and dataset splits
- Trial registration, protocol and analysis plan for clinical work
Clarity & Language
Language model · calibrated on open-review examplesWeight ×0.8 in the overall scoreAssesses readability, structural flow, and adherence to scholarly writing norms.
What it looks for
- An abstract with background, objective, methods, results with numbers, and a conclusion
- Hedging: overclaims such as "proves", and needless over-hedging
- Consistent terminology, over-long sentences, undefined acronyms and section flow
Figures & Tables
Language model · calibrated on open-review examplesWeight ×0.8 in the overall scoreChecks figure quality, caption completeness, and appropriateness of visual encodings.
What it looks for
- Self-contained captions: units, error-bar type, n per group, the test used
- Chart type for the data, colour-blind-safe palettes, resolution
- Numbers in the text that disagree with a figure or table, and image-integrity flags
Sees the figure images embedded in a Word file, or the figure files you attach to LaTeX; from a PDF it works from the captions and text.
Research Gap
Language model · reads comparable published papersWeight ×1.0 in the overall scoreQuotes the gap the manuscript says it fills, checks it against comparable published papers retrieved from our library (with their mined methods, data and results where we hold them), and scores whether that gap is real and addressed.
What it looks for
- The sentence where the manuscript states the gap, aim or question it addresses, quoted
- Whether comparable published papers already answer that question, citing only the papers it was given
- Whether the paper's design and results actually address the stated gap, with concrete fixes
Reads the manuscript plus up to 10 comparable published papers our library holds for its title; it never cites a paper it was not given.
Prior Publication
Deterministic lookup · no language modelReported on its own, not averaged inA deterministic lookup, not a language model: fans out in parallel to 7 external sources — CrossRef, PubMed, arXiv, bioRxiv, medRxiv, Europe PMC and Unpaywall — and our local library to detect prior publication and duplicate submission.
What it looks for
- An existing record whose title and abstract closely match yours
- Searched in parallel: 7 external sources and our 7.9M-paper local library
- A likely or possible prior publication is flagged in the report, with a link to the matching record when the source gives one
Reads the title and abstract only.
From 9 agents to one report
How the agents become one report
- 1
They start together
All 9 agents begin at the same time, so this stage takes about as long as the slowest single agent, not the sum of all of them.
- 2
The specialist reviewers are calibrated first
Before it reads your paper, each of the 7 specialist reviewers is shown up to 60 real examples from open peer review, retrieved for its own specialty from an index of 5K+ peer-review texts collected from 15 open-review platforms — reviewers' critiques first, the same examples whatever the manuscript, mixed across outcome labels, and no platform above 40% of an agent's examples unless the others run out — and told to hold its concerns to that standard. On 28 October 2026 every agent's examples came only from four of those platforms: OpenReview, SciPost, PeerJ and JOSS. The Research Gap agent is not calibrated on reviews: it is given up to 10 comparable published papers that our library holds for your title, with their mined methods, data and results where we have them.
- 3
It gets context on the field
When the local library has them, each specialist reviewer also receives research gaps that similar papers have already stated in its area, plus a live OpenAlex snapshot of the field: how much has been published, the top venues and the peak year.
- 4
Scores become one weighted score
Methodology counts ×2.0 and Originality ×1.5, the two heaviest weights; Prior Publication is reported on its own and not averaged in. Any language-model agent scoring 2 or below caps the overall at 4.0. The verdict follows the score: Accept from 8.5, Minor Revision from 6.5, Major Revision from 4.0, Reject below that.
- 5
A synthesis step writes one report
One report: the verdict, a recommendation that cites what the agents found, the issues to address first, the paper's strengths, any formula corrections and the reproducibility gaps.
Run all 9 on your manuscript
Upload a PDF, Word or LaTeX file. A full review costs 30 credits, charged once for every agent and the synthesis. Credits come in one-time packs of 50 and never expire; there is no subscription.
Questions