The Challenge
Every piece of evidence synthesis rests on its citations, and checking them properly is one of the first things that gets squeezed when a deadline appears. A careful author under pressure extends a claim slightly beyond its source, and the reference list still looks immaculate. AI-drafted text does the same thing, faster and at greater volume. Both need checking, for the same reason.
Reading every source and locating every place it is used is laborious, repetitive work. In a document with hundreds of references, it simply does not happen end to end.
Two Questions, Asked of Every Citation
Does this reference exist?
Title, authors, journal, year, and identifier are parsed straight out of the document and checked against the published record. The answer is graded, never just ticked: high confidence requires the title, author, and journal to agree across independent sources. The system also recognises when something is not a journal article, verifying working papers and technical reports through institutional repositories where a journal lookup would find nothing.
Does the source actually say that?
This question needs the reference itself, so the source papers are supplied and read in full. Every place each one is cited is then located and judged on its own. A reference cited five times is judged five separate times, once for each claim it carries. Each judgement carries the quote it relied on, so it can be checked.
When a citation goes past its source, the system draws the line precisely: it confirms what the source does support, flags the part of the sentence that is an interpretive extension, and writes the claim the reference would properly support, so the author has somewhere to go.
See It Work
The recorded demo runs on one of our own whitepapers: 20 pages, 13 references, 6 source papers supplied and read, 21 citation instances judged individually.
What It Means for You
- Built as a companion, not a gate. It confirms the references that stand and flags the ones that need a person, so a reviewer spends their limited time on what looks suspect instead of reading several hundred references one at a time.
- Graded answers with evidence. Every verification and every citation judgement carries its confidence level and the evidence it relied on, so the human check is fast.
- Works on human-written and AI-drafted text alike. Verification belongs outside the drafting, whoever did the drafting.
- The judgement stays with the author. The machine carries the laborious part: reading every source and locating every use. Deciding whether an extended claim is defensible remains a human call.
For anyone producing reviews, dossiers, or manuscripts, in consultancies or in pharma: a complete citation integrity pass on demand, instead of the check that never quite happens.