Guide
What an AI citation proves, and what it does not
A citation proves the quoted text exists in the source at that location. It does not prove the passage supports the claim, that the claim is complete, or that it is correct. Those are separate checks.
For anyone shipping ai output with citations. Written by the Encevra engine team. Reviewed 2026-10-12.
Three separate questions
Is the quote real and located? Does the passage support the specific claim? Is the claim correct and complete given everything else in the material? Noetis answers the first mechanically, makes a model judgment on the second, and leaves the third to your reviewer.
How this shows up in the API
Each finding carries the quoted passage with its location and a support status from the claim review. A 'supported' status is a model judgment, not a guarantee. An unsupported or unlocated claim is withheld from the released brief or the run is returned incomplete.
Native citation features
Some model providers offer native citation features that also return source locations. If you are evaluating options, compare against them fairly.
Labels to show your users
Keep source checks and interpretation checks visibly separate so a real quote is never read as proof of the claim.
- Source verified: the quote was found in the submitted text
- Interpretation requires review: the passage exists, its meaning still needs a person
- Insufficient evidence: the quote could not be located
- Conflicting evidence: passages in the material disagree
- Extraction incomplete: some text could not be read
Frequently asked questions
- Can a model mark its own quote as verified?
- No. Location is checked mechanically against the stored text, not taken from the model.
- Does agreement between two models prove a claim?
- No. It is a useful review signal, not proof.
- What about missing clauses?
- A gap describes what was searched. It never cites text that is not there.