AI review that knows why
papers get rejected.

Most AI reviewers learned to sound like a reviewer. Archon's seven-seat panel is built on 14 years and 3,250+ manuscripts taken through peer review, so your editors get the objections a seasoned editor would raise before a single reviewer is invited.

Run it on 10 of your past submissions Under your data terms: nothing retained, nothing trained on.
3,250+ paperstaken through peer review, written into every seat
Seven seatsthat never see each other's reports
The P0 gateno “accept” while a critical flaw is open
Your guidelinesaudited clause by clause, for every submission

More manuscripts, fewer reviewers, more fabrication

The desk is where it lands. Every paper that should have been stopped there costs a reviewer a weekend and a journal its turnaround.

4.5

invitations to secure a single review, nearly double 2018.

Silverchair Peer Review Report, 2026
1 in 277

biomedical papers in early 2026 carry fabricated references. In 2023 it was 1 in 2,828.

Topaz et al., The Lancet, 2026
3.5M

submissions to one publisher in a year, against 728,000 articles published.

Elsevier Journal & Article Ecosystem, 2024

Screening tools catch fraud signals. What editors still do by hand is the judgement: is this paper sound, does it fit, what would a reviewer object to?

Run it on 10 of your past submissions

Does your AI reviewer know what an editor rejects on sight?

Generic AI reviewers are fluent, polite and long. They nitpick, they miss what matters for significance, and they can cite papers that don't exist. The problem isn't the model. It's where the judgement comes from.

 
Generic AI reviewer
Archon panel
Where its judgement comes from
General reading, plus a prompt that says “act like a reviewer”
14 years of learning which errors editors forgive and which end a submission, written into every seat
Who reads the paper
One voice
Seven independent seats, then one editorial synthesis
What you get back
A flat list of comments, big and small alike
A consensus matrix, prioritised P0 / P1 / P2
Your journal's rules
A generic checklist
Your guide for authors, audited clause by clause
Its verdict
Can be talked into “accept”, even by a hidden prompt
Cannot recommend acceptance while a P0 finding is open
Run it on 10 of your past submissions

Fourteen years of experience isn't a dataset you can download

It started fourteen years ago with a question from a university in the south of Chile: can you help us publish 80 papers this year? Since then, Journal Revisions has taken thousands of manuscripts from draft to acceptance, one rejection and one revision at a time.

“In fourteen years of reading reviewer reports, we have learned which errors an editor forgives and which end a submission on sight.”ARCHON white paper, July 2026

A professor knows their field. A reviewer knows the papers they've reviewed. We know the pattern across disciplines: from animal science to astronomy, what makes an editor send a paper out, and what makes them send it back.

That record is what each seat of the panel works from. It is built from our experience, not trained on your manuscripts. The seats don't improvise a reviewer's opinion; they apply rules that took thousands of real decisions to learn.

14years in scientific publishing
3,250+manuscripts taken through peer review
100+universities, across every discipline
7independent seats, one synthesis

Seven reviewer seats that never see each other's reports

Each seat reads the manuscript from one angle and files its findings alone, so the panel can't talk itself into agreement. A synthesis seat then merges them into one recommendation for your editor.

Editor-in-chiefJournal fit, structural completeness, word and reference limits.
→ Seat 1
MethodsEvery reported statistic checked for consistency; methods against the stated design.
→ Seat 2
Domain expertLiterature coverage and contribution; missing papers named with full citations.
→ Seat 3
The readerReads as your journal's audience would: accessible, relevant, useful?
→ Seat 4
Devil's advocateAttacks the core claims. Must file a serious objection or justify why none stands.
→ Seat 5
Guide-for-authors auditorYour guidelines, clause by clause, quoted verbatim. Format only, never content.
→ Seat 6
Field & writing qualityCalibrates to the field; flags AI-typical prose with the passage and location.
→ Seat 7
Editorial synthesisConsensus matrix, conflicts resolved, P0 / P1 / P2 list and one recommendation.
→ Synthesis

The gate: while any P0 finding is open, the panel cannot recommend acceptance. A seat that's been manipulated, or a paper that reads beautifully, can't override it.

Run it on 10 of your past submissions

Also in the API, the checks that should be standard

01

Guide-for-authors audit, on its own

Run at submission: every clause of your guidelines checked before an editor opens the file. The check no other AI reviewer offers.

02

Every reference verified

Table stakes, done properly: resolved across Crossref, PubMed, OpenAlex and Semantic Scholar. Retracted, unverifiable and misrepresented sources are flagged, never guessed.

03

Reviewer-report check

Human reviews checked for citations that don't exist and objections the manuscript doesn't support, before the decision letter.

Your editor receives findings to act on, not a score to trust

Prioritised, not piled up

P0 blocks acceptance, P1 needs fixing, P2 is polish. Your editor reads the P0s first instead of wading through forty comments of equal weight.

Every finding anchored

Each one names the seat, the location in the manuscript and the evidence: the DOI that resolved to a different paper, the clause of your guidelines, the claim the design can't support.

A recommendation, never a decision

Accept, minor, major or reject, with the reasoning shown. The decision, and any disagreement with the panel, stays with your editor.

// Illustrative: structure and field names are indicative
{
  "manuscript_id": "JRN-2026-04418",
  "recommendation": "major_revision",
  "accept_blocked": true,
  "open_P0": 2,
  "seats": ["eic", "methods", "domain", "reader",
            "devils_advocate", "format", "field_wqc"],
  "references": {
    "checked": 64, "unverifiable": 3, "retracted": 1
  },
  "guide_for_authors": { "clauses": 38, "failed": 5 },
  "findings": [{
    "priority": "P0",
    "seat": "devils_advocate",
    "location": "Discussion, para 2",
    "evidence": "Causal claim; design is cross-sectional"
  }]
}

Illustrative. Full schema and documentation are shared with publishers during onboarding.

Built for the rules your journal already has

“Our manuscripts are confidential.”

They stay that way. Nothing is retained after the report, and nothing is used to train a model, under a data-processing agreement in your name. It's a private tool in the sense publishers' AI policies now define, not a public chatbot.

“Our policy bans AI-written reviews.”

This isn't one. The panel supports the editor; its output isn't sent to authors as a review. Each journal switches seats on or off to match its own policy, down to the guidelines audit alone.

“AI hallucinates and nitpicks.”

Every reference is resolved or flagged, every finding points to a location, and P0 / P1 / P2 keeps the trivia out of your editor's way. It does not judge a paper's significance. Your editors and reviewers do.

“Authors hide prompts in manuscripts.”

Manuscript text is treated as material to review, never as instructions. Hidden prompts are flagged as an integrity finding, and the P0 gate means no single seat can force an “accept”.

The same panel for big and small publishers

Large publishers

A component, not a platform: take the seats you're missing into the integrity stack you've built, configured per title and consistent across the portfolio.

Society & university presses

An editor's assistant on day one: senior judgement at the desk without adding staff, so reviewers only see papers that are ready for them.

Small & diamond-OA journals

The integrity desk you can't hire: a volunteer editor gets the same reference checks and panel as the largest houses.

Who the Reviewer API is not for

Journals looking to automate acceptance or replace their reviewers. The panel prepares the decision; it never makes it.

Run it on 10 of your past submissions

Frequently asked questions

How is this different from our integrity screening tools?
Screening tools look for fraud signals: paper mills, image manipulation, duplicate submissions. The panel assesses substance: soundness, fit, what a reviewer would object to, and compliance with your guidelines. It sits alongside the tools you already use.
Does it replace peer reviewers?
No. It helps editors decide what goes out for review and what to ask reviewers to focus on. Significance and the final decision stay human.
What does “built on 14 years” actually mean?
The rules each seat applies (what counts as a fatal flaw, what an editor forgives, how a methods section is judged) come from Journal Revisions' fourteen years taking 3,250+ manuscripts through peer review at 100+ universities. It is our method written into the panel, not a model trained on anyone's manuscripts.
Will it work with our submission system?
It's an API, designed to plug into your existing submission workflow. Tell us which system you use and we'll walk through the integration.
How is it priced?
By volume and by the seats you switch on. Contact us for a proposal.

See what the panel would have caught in your journal

Send us 10 manuscripts you've already decided on. We'll return the panel's reports, and you compare them with what your editors and reviewers concluded.

Run it on 10 of your past submissions Under a data-processing agreement · nothing retained · nothing trained on