AI citation verification
Does your citation actually support your claim?
You wrote the claim. You attached the sources you thought supported it. Agent Bayes reads the full text behind each citation and scores how well the evidence carries the sentence as written, then tightens the phrasing where it does not.
Free to start, no card. Verification is a paid-plan feature.
The reason you are here
The reference is real. The question is whether it says what you think
You read the paper in March. You paraphrased it in June. Somewhere between the two, the claim picked up a "demonstrates" where the source only "suggests", and a scope the authors were careful to limit.
Rereading every cited passage to check the fit is exactly the work that gets skipped under deadline, and it is exactly the work a reviewer does first.
The failure mode here is not a fake citation. It is a real citation attached to a sentence that reaches slightly further than the evidence.
How verification runs
Your sentence, your sources, checked against each other
Three steps, and the first two are things you were doing anyway.
Write the claim in your own words
You are not asking the agent to generate the sentence. It is your argument, phrased the way you intend to publish it.
Attach the citations you believe support it
Pull them from search results, from the agent's retrieval, or from your own labelled collections of passages.
Ask whether the evidence actually says that
The agent reads the full text behind each citation, scores the alignment, and rewrites the claim where the sources do not carry it as phrased.
Where the citations come from
Attach evidence as you read, verify it later
Search your corpus semantically, label the passages worth keeping, then attach them to the claims they belong to.
What you can rely on
What verification actually checks
It reads the full text, not the abstract
Verification works from the complete passage behind each citation you attached, which is the level at which overstatement actually shows up.
A score you can act on
Each claim gets a confidence score reflecting how well the cited evidence supports it as written, so you know which sentences to revisit and which to leave alone.
A rewrite that keeps your meaning
Where the phrasing overstates or only partly reflects the sources, the agent tightens the claim into alignment while preserving what you were trying to say.
Run it on a selection
Select the nodes you care about rather than the whole map. Useful before a supervisor meeting, and cheaper than verifying everything.
What Agent Bayes does not do
- It does not certify your work. A high alignment score is evidence that your phrasing matches your sources, not a guarantee that a reviewer will agree with the argument.
- It cannot check a source it cannot read. Verification works against the indexed full text, so a citation to a paper outside your knowledge base cannot be scored.
- It is not on the free plan. Indexing, search, the reader and the research agent are, and verification needs a paid tier.
The middle one is the practical limit worth planning around. Verification is only as good as the coverage of your own library.
Before you ask
Questions researchers ask first
How is this different from the citations the agent writes itself?
That is the other direction. When the agent writes a claim, it starts from retrieved passages, so the claim follows the evidence. Verification runs on claims you wrote, where the sentence came first and the question is whether your sources hold it up.
Will it just agree with me?
It scores alignment rather than confirming it, and a low score with a tightened rewrite is the common outcome for a claim that reaches slightly further than its sources. That is the point of running it.
Does it change my text without asking?
You get a score and a proposed rewrite. Applying it is your decision, and version history keeps the original either way.
Which plans include this?
AI verification is a paid-plan feature. The free plan covers indexing, semantic search, the reader and the research agent, so you can see how the citations work before deciding whether verification is worth a subscription.
What does it cost to run?
It is metered in credits like every other operation, costs are estimated and reserved before a run, and only actual usage is charged. Every expense is logged in a ledger you can inspect.
Related
If you are also wrestling with
Create your account
Check a claim you are not quite sure about
Start on the free plan to index your papers and see how citations attach to claims. Add verification when you want the alignment score.
- No credit card needed
- Costs estimated before every run
- Your original text is never lost