Citation auditing for documents that carry liability
Every citation, checked. The fabricated ones, flagged.
RefProof audits every reference in a document and flags the ones that do not exist, or do not say what the author claims. Upload a brief, report, or paper; get each citation verified against Crossref, PubMed, and court records.
Request early access In development. Waitlist opens the first pilot slots.Courts have repeatedly held that a duty of care extends to foreseeable third parties[1],
[1] verified: holding matches the opinion
and at least one circuit has extended this duty to software vendors[2].
[2] no such case in any registry; flag
Recent scholarship supports a broader reading of proximate cause in this context[3].
[3] article exists; argues the opposite
Nobody checks citations at scale. The consequences are stamped on real documents.
LLM-drafted text has flooded courts, journals, and corporate reporting with references that are fabricated or misrepresent their source. Manual verification takes minutes per reference; documents have hundreds. Three places it ends badly:
U.S. District Court · Motion
Plaintiff relies on the settled rule of Varghese v. China S. Airlines, which this Court cannot locate in any reporter.
In courtrooms
Sanctions for briefs citing nonexistent cases are now routine, and they attach to the lawyer who signed the filing, not to the tool that drafted it.
Journal manuscript
Attention Mechanisms in Clinical Triage: A Review
Anonymous et al.
Prior work established the effect in three cohorts [12] and the finding was widely replicated across settings.
However, subsequent review found reference [12] does not exist in any indexed database.
In journals
Editorial offices are retracting papers over reference lists that do not survive scrutiny. One verified study counted roughly 146,900 fake citations in a single corpus. [a]
Meridian Advisory
The regulator has signaled approval of this structure (Bulletin 14); in fact Bulletin 14 cautions against it.
In firms
Due-diligence reports and consulting deliverables inherit the same risk: a confident claim resting on a source that says something else, or nothing at all.
[a] We practice what we audit: every number on this page traces to a source we have verified, and we publish nothing we cannot support. The three documents above are constructed illustrations, not real filings.
How an audit runs
Three passes over your document, in order. Each pass narrows what the next one has to do. You watched them run on the brief above.
Extract every citation
Upload a PDF or DOCX. Structure-aware parsing pulls each reference with the exact claim it is attached to, including footnotes, inline cites, and mangled bibliography entries.
Check that each source exists
Fan-out lookups against authoritative registries: Crossref, PubMed, arXiv, and CourtListener court records, with fuzzy matching for garbled titles, misspelled authors, and wrong years.
Check that each source says what is claimed
Where full text is available, RefProof retrieves the source and runs claim-against-source entailment with an LLM. A real citation that misstates its source gets flagged just like a fabricated one.
Read the audit report
One verdict per citation with the evidence behind it: the registry match, the retrieved passage, and the reasoning. Flags are for human review; you decide what to fix.
Three verdicts. No maybe.
Every citation in the report resolves to exactly one of these, each backed by evidence you can inspect.
The source exists
Matched in an authoritative registry, and where full text is available, the cited claim is supported by what the source actually says.
The source does not exist
No plausible match in any registry we check, after fuzzy matching for typos and formatting damage. The signature failure mode of LLM-drafted text.
Real source, wrong claim
The reference exists, but the passage it points to says something different, weaker, or opposite. The hardest errors to catch by hand, and the most embarrassing.
Built for documents that cannot afford fake references
Registry-grade existence checks
Crossref for journal articles, PubMed for biomedical literature, arXiv for preprints, CourtListener for case law. Each verdict names the registry it was checked against.
Claim-level verification
Existence is the easy half. RefProof also reads the source and checks whether it supports the specific sentence it is cited for.
Evidence you can hand over
Every flag comes with the registry response or retrieved passage behind it, so the report is reviewable by a partner, an editor, or a court.
Minutes, not days
A reference list that takes an afternoon to check by hand runs as one automated audit, so verification can happen on every draft instead of only after something goes wrong.
An audit, not a truth oracle
RefProof flags problems for human review. It does not certify a document as correct, and we will not pretend otherwise.
API for platforms
The same pipeline behind the web app, exposed for editorial systems and legal-tech platforms that want citation checks inside their own workflow.
Where we are, plainly
RefProof is in development. There is no public product yet, no customer logos to show you, and no invented metrics on this page. What exists today: the extraction and existence-checking pipeline, running against live registries.
We are opening a small early-access group: law firms, editorial offices, research-integrity teams, and anyone else whose documents carry their signature. Early users get free audits during the pilot and a direct line to the people building it.
Join the waitlistQuestions we get
What kinds of documents can I audit?
Anything with references: legal briefs and filings, journal manuscripts, theses, due-diligence and consulting reports, white papers. Upload a PDF or DOCX; RefProof handles footnotes, endnotes, inline citations, and bibliography styles that have been through too many edits.
Which sources do you verify against?
Crossref, PubMed, and arXiv for scholarly and biomedical literature, and CourtListener for court records and case law. Coverage grows by vertical; if your field relies on a registry we do not check yet, tell us and it moves up the list.
What if a source exists but is paywalled?
Existence can still be verified against registry metadata. Claim-support checking needs the text itself, so where full text is unavailable the report says exactly that, rather than guessing. A verdict you cannot trace to evidence is not a verdict.
Does a clean audit mean my document is correct?
No, and be suspicious of any tool that says otherwise. A clean audit means every citation was found and, where checkable, supports its claim. Your argument, your data, and your conclusions remain yours to defend. RefProof removes one specific, well-documented failure mode.
What happens to my documents?
Documents are processed to produce your report and are not shared or used to train models. For legal work in particular we know confidentiality is not optional; pilot users get our data-handling terms in writing before uploading anything.
How do I get access?
Email founder@refproof.xyz with a line about what you need audited. We are onboarding a small pilot group in order of fit, and pilot audits are free.