Pre-submission integrity review
A pre-submission integrity review for high-stakes scientific manuscripts. Software resolves every reference at its primary source and checks that each one supports the claim it carries. Then a person reads the full record, re-derives its load-bearing claims, and releases it by hand. Corrections are proposed, never applied. We flag. You decide.
Author-side · Before submission · Confidential by written agreement
“I could check it 500 times and I’m still going to miss something. At some point you become blind to it.”
A page from the sample review, invented to show the form. Read it in full →
Act I · The problem
Across the 2025 and 2026 studies, AI assistants fabricate citations at rates that vary widely by model, domain, even phrasing. That variance is the real problem: from the outside, you cannot tell whether a manuscript’s references are clean or invented.
It is not hypothetical, and the cost is real; the two figures below each expand to their primary source. The one thing you can do before you submit is check every citation at its origin. That is the whole of this review: is every source real, and does it support the claim it carries.
1 in 277
Published biomedical papers found to contain fabricated references in early 2026, in an audit of 2.5 million papers.
Topaz et al., “Fabricated citations: an audit across 2.5 million biomedical papers.” A 2026 research letter in The Lancet, reporting a single audit. PMID 42107362. These were papers that had already passed peer review. Check the record at PubMed.
$392,582
Mean direct-funding cost of a single misconduct retraction, among NIH-funded papers.
Stern et al., “Financial costs and personal consequences of research misconduct resulting in retracted publications.” eLife, 2014. PMID 25124673. Scope: NIH-funded papers, 1992 to 2012, misconduct retractions specifically. Check the record at PubMed.
Every figure on this page expands to the primary record it comes from. If a number here does not match its source, tell us. That is exactly the kind of error this service exists to catch. The full analysis, with every source, is here: Six Times the Rate.
A note from Ryan Gruzen
Science runs on trust in a single citation. Your work builds on someone else’s, theirs on the next person’s, all the way down. When one citation is quietly wrong, everything built on top of it inherits the mistake, and no one downstream can see it.
ASI exists to close that gap: between what a paper claims and what its sources actually support. Source review is where I start, at the smallest piece I can check.
Act II · The review
It tells you what a careful reviewer would raise, and the repair available for each. It never touches your prose or your argument. Three commitments hold everywhere:
Verified at the primary source.
Every reference resolved and read at its origin, field by field. Not “we ran it through AI.”
Every decision left to you.
Corrections are proposed, never applied. We map what a reviewer would raise; you decide every change.
A person releases it.
A person reads the complete record, re-derives the load-bearing findings, and releases it by hand. No automated send, ever.
Entrust your manuscript.
A confidentiality agreement is signed first. Then you send the near-final manuscript and your target journal.
We verify, reference by reference.
Every citation is resolved at its registry (PubMed, Crossref) and read in full: does the paper exist, is the identifier right, does it support your claim at the same scope and population. Title-only matching is never used; journal requirements are checked at the journal’s own live instructions.
You receive a decision record.
Verified citations, the discrepancies found, claim-support gaps, and the repairs available, organized so the one or two decisions that matter are elevated, not buried.
A person reads it, then releases.
A person reads the complete record and releases it by hand. The software is structurally incapable of sending anything on its own.
What it is not
Not a rewrite. It never touches your prose or your argument.
Not a verdict. The readiness assessment is a prediction, labeled as one. Never a score, never a guarantee.
Not publisher-side screening. No paper-mill detection, no image forensics, no plagiarism scan. Different tools, for a different desk.
Not a judge of your science. Merit, ethics, and intent stay with you.
What it is
A primary-source check of every citation and the claim it carries, across the whole manuscript, including references inside tables and figures.
A structured decision record with the repairs available now, from existing literature and wording.
An independent check on every finding the paper rests on. The system that produces a finding is never the system that checks it.
A human release. A person is accountable for every review that goes out.
We verify references, citations, and claim support against their primary sources. We don’t judge the science’s merit. That’s yours.
What you receive
Four artifacts. Two are held until a person releases them; two finalize as your decisions land.
The email body that accompanies your review.
Held until releaseThe structured decision document, and the full verification record.
Held until releaseYour full reference list, set to journal style.
Finalizes with decisionsWhich citation sits at which sentence, and its number.
Finalizes with decisionsBefore anything else
Before you send anything, we sign a confidentiality agreement scoped to your manuscript. Your work is read only to complete your review, never used to train any model, never shared, never reused.
The engagement itself stays confidential: no public attribution, no “powered by.” Particular retention terms go in writing.
Act III · The method, shown
Excerpts in the exact shape of a real review, invented to show the form; no client manuscript appears anywhere on this page. Or read a complete sample review, start to finish: the verdict, every finding, the full reference record, and the review’s own errata. Every detail is invented.
The errata · The review corrects its own work, in the open
“A self-check caught one place this review claimed more than it had verified: an earlier draft said a wrong identifier ‘points to a different paper.’ Corrected: the review confirmed only that the listed identifier is not the correct one for the cited work; it did not verify where the wrong identifier leads, and does not claim to.”
Before release, every review checks itself and discloses any place an earlier draft over-stated. A review that publishes its own corrections is the one you can trust. We would rather show a catch than claim we never need one.
Findings · Each one traceable, each with a repair
Substantive
“The abstract hedges responsibly; the conclusion drops the hedge and asserts causation. Repair: harmonize the language to the hedged register the limitations section already commits to.”
Findings arrive in tiers, each with the same structure: the claim as written, what the source actually is, why it matters, and the repair available now. No new research required.
The rule behind the method
The verification software is AI-assisted, and we assume it can err. So the discipline is independence: every load-bearing finding is re-checked by a separate pass that does not share the first one’s inputs, and a person re-derives it before release. Where something cannot be confirmed, the review says “could not confirm.” A false “verified” is the one failure we treat as catastrophic.
The rule exists because of a real failure. Early on, a text-extraction bug dropped citations and produced a false “out of order” claim; the self-check missed it because it re-ran the same buggy extraction. The lesson became the rule: independence is a property of the input chain, not of who checks.
It is also the honest answer to “why not just use a chatbot?” A general AI assistant cannot catch its own fabricated citations. This method is built to.
A pattern we see
A pattern common enough to name. An author runs a near-final manuscript through a leading AI assistant, and it comes back with a verdict:
The assistant’s verdict
“Fully submission-ready. A clean bill of health.”
What a source-level review finds
A source that does not support the claim it is cited for.
An author listed on a reference who had not contributed.
Identifiers that resolve to the wrong work, or to nothing.
An abstract in the wrong format for the target journal.
The assistant graded the prose. It could not check whether the sources were real. That check happens at the registries where the sources live, one reference at a time, and it is the entire job of this review.
“I found the feedback especially helpful and constructive.”
Unattributed by design. Engagements are confidential, and no client is named here without written consent. Attributed words will appear when, and only when, a client chooses to be named.
We don’t ask you to trust this review. We ask you to check it.
Where this stands
ASI Source Review is early: one operator and an AI verification tool, a handful of manuscripts, a method documented so you can check it. Early means a full human read, not a place in a queue.
It also means no acceptance-rate claims, no accuracy figures, no track-record numbers, because none are demonstrated yet. What we can show is the method, and it is on this page.
On the record
Long-form pieces on what discipline looks like pointed at scientific discovery. Written in the open, every source cited.
Before you need us
A short pre-submission citation checklist: the questions we ask of every reference list, in a form you can run against your own manuscript. Free, yours whether or not we work together, and it makes your submission stronger on its own.
Open the checklistKnow a colleague with a submission coming up? The checklist is easy to pass along, and it asks nothing of your name.
More free tools, same spirit: the claim-support self-audit, the retraction-risk reckoner, and a printable grant-deadline pass.
The questions worth asking
Still holding questions? Good. These are the ones authors actually ask, answered plainly.
Yes, by written agreement signed before anything is sent: scoped to your manuscript, no secondary use, no training, no attribution.
A general AI assistant cannot catch its own fabricated citations. This is built on the opposite principle: the tool that produces a finding never checks it, every citation is resolved at its registry, and a person confirms every load-bearing conclusion before it goes out.
Journal screens are automated, at-scale, and run after you submit (text overlap, AI-text detection). This is the opposite end of the desk: author-side, before submission, human-released, and focused on the one question those screens skip: is each citation real, and does it support its claim.
One manuscript: the structured decision record shown above, with every citation verified at its source, the discrepancies and claim-support gaps found, and the repair for each. Includes the initial review plus one re-review of your revision; further rounds priced the same way. Corrections are proposed, never applied.
You are the author, and nothing is ever applied to your manuscript. If you believe a finding is wrong, say so: it is re-checked against the same primary source and either corrected or explained. The review's own errata practice applies to itself first.
No. It is a review, not a rewrite. It does not touch your prose or your argument. It tells you what a careful reviewer would raise; you make every change.
A named engagement, not a queue. Timelines are set against your actual deadline when you request the review, and we tell you honestly what we can commit to.
Per manuscript, set in conversation and sized to the manuscript and its deadline. This is a professional, human-released engagement, not an automated scan, priced in the neighborhood of what it costs to publish the paper itself. Weigh it against the cost of getting a submission wrong: a rejection cycle in months, or a retraction whose mean direct cost, in the one careful audit we know of, ran to hundreds of thousands of dollars. A pre-submission source check is inexpensive insurance.
A person is accountable for every review and releases it by hand. Software verifies at the source registries; the person reads the full record and confirms each load-bearing conclusion before anything is sent. We would rather show the method than a bio: every review documents how its checks were run, so you can re-run them.
It is real, and it is early. Our first engagements are with senior academic authors, and they are confidential, so you will not find a wall of endorsements here. What we can show is the method, including how it corrects its own work. We put the method first for a reason: a testimonial asks you to trust someone else's judgment; the method invites you to use your own.
The invitation
One manuscript is the base unit: the full review, plus one re-review of your revision. If you have a high-stakes submission ahead, send word. The confidentiality agreement comes first, the manuscript after. What you get back is a record you can check yourself, line by line.
Request a review