How Science AI Journal reviews your paper

1. Upload

Upload a PDF, Word (.doc/.docx) or LaTeX (.tex) file. We review manuscripts from any scientific discipline. The 7 calibrated agents' examples are not matched to your field: on 28 October 2026 they came only from machine-learning, physics, life-science and research-software reviews (OpenReview, SciPost, PeerJ and JOSS), the same for every paper.

• Images embedded in a Word file are extracted for the Figures & Tables agent; with LaTeX you can attach the figure files.
• You keep all copyright; your manuscript is never added to the calibration corpus.
• 30 credits per review, from one-time credit packs — no subscription.

2. Parallel pre-flight: prior publication detection

Before any review agent runs, the title and abstract fan out in parallel to 7 external sources — CrossRef, PubMed, arXiv, bioRxiv, medRxiv, Europe PMC and Unpaywall — and to our own 7.9M-paper local library, with a 12-second timeout per source.

We look for fuzzy title and abstract overlap. Matches above 60% word overlap are flagged as likely prior publication and shown in your report, with the matching record linked when there is one. The check is informational: the review continues either way, and the judgement is yours.

3. Nine specialised agents, in parallel

All 9 start at the same time, so the wall time is roughly the slowest single agent rather than the sum of them. Each language-model agent carries only the rubric it needs, and the 7 calibrated ones only the calibration examples chosen for their own concern — one narrow prompt per concern rather than one prompt asked to hold every concern at once. We have not published a measured agreement rate against human editorial decisions; what we have measured is on the benchmarks page.

• Methodology — study design, power, CONSORT/STROBE/PRISMA compliance.
• Formulas & Equations — derivations, dimensional analysis, algebra.
• Originality — novelty claims checked against the 7.9M-paper local library.
• Literature Coverage — missing seminal refs, over-reliance on self-citation.
• Reproducibility — code availability, dataset access, methods sufficiency.
• Clarity & Language — IMRaD adherence, hedging, undefined acronyms.
• Figures & Tables — readability, colourblind safety, caption completeness.
• Research Gap — the gap the manuscript states, checked against up to 10 comparable published papers retrieved for its topic; it reads those papers, not calibration examples.
• Prior Publication — a deterministic lookup (not a language model) across 7 external sources and our local library.

4. Synthesis

A synthesis pass then integrates every agent's report into a single editorial recommendation — accept, revise, reject — with a numeric score and a line-by-line reviewer report. The full report is yours: every agent's findings, not a summary, copyable as Markdown, JSON or plain text. If our publication route opens, the plan is for that report to travel with the paper rather than sit behind a curtain.

What we will not claim

We do not replace human peer review where the stakes demand it — drug trials, regulatory submissions, grant panels. We do not outperform a careful, well-resourced human reviewer on nuanced theoretical work. We do not generate novel scientific insight. We review.

What AI Review is for: a competent, fast, transparent first pass before a manuscript goes to a journal — so the human referees who read it next spend their time on the science rather than on the fixable problems.