The first thing a lawyer does with an AI answer is try to break it. They open the source, find the passage, and check that it says what the answer says it says. If it does not, the tool is finished with them, and no amount of accuracy on the other cells will bring them back. So the citation is not a feature of the answer. It is the thing that makes the answer usable at all.
Most tools handle this by asking the model to include citations in its output. The model reads the documents, writes an answer, and writes a page number or a quote next to it. Sometimes the quote is real. Sometimes it is close. Sometimes it is a plausible sentence the document never contained. There is no way to tell from the outside, which means every citation has to be checked, which defeats the point of having one.
Doclayer does it the other way round. The model never writes a citation. It chooses one.
The server cuts the evidence first
When a review runs, Doclayer does not hand whole documents to the model and hope. For each question, the server searches the selected Documents and pulls out candidate passages: exact excerpts, each with its document, its page and its position. Every excerpt gets an opaque identifier, a short token that means nothing outside that request.
The model receives the question and the excerpts with their identifiers. Its job is to answer the question and to say which identifiers support the answer. That is the whole contract. The output is checked strictly: an answer that names an identifier the server did not send is rejected, not repaired.
Only the excerpts the model selected, and only if the server sent them, become citations in the cell. When you open the citation, you are looking at text the server cut from your document before the model ever saw it. The model could not have invented it, because it never had the chance to write it.
What this rules out
It rules out the citation that points to a page which does not contain the quoted words. It rules out a citation to a document that was not in the review, or to the model's general knowledge of what NDAs usually say. And it rules out the quiet failure where the answer is right and the citation is wrong, which is the one that gets people into trouble, because the answer looks fine.
It does not rule out everything. The model can still choose the wrong excerpt, or read a correct excerpt and draw the wrong conclusion from it. Those failures are visible, though, because the excerpt is right there beside the answer. A reviewer who opens the cell sees the mismatch in a few seconds. That is the failure mode we chose to keep: the one a person can catch.
Insufficient evidence is a result
The same design gives us an honest answer for the case where the documents do not answer the question. If none of the excerpts support an answer, the model is expected to say so, and the cell shows insufficient evidence. It is not a blank, and it is not a no. It means: the documents you selected do not contain this, look elsewhere. In diligence that is often the most useful cell in the row, because it tells you what to ask the other side for.
What it costs
The obvious cost is retrieval quality. If the search misses the right passage, the model cannot cite it, however good the model is. We spend most of our effort there. Retrieval runs a lexical search and a semantic search over each document and fuses the two candidate lists by reciprocal rank rather than by adding scores that were never on the same scale. Chunks keep their page and layout position from parsing, so an excerpt can always be traced back to where it sits on the page.
The less obvious cost is that this approach is slower to build than "ask the model for a page number", and it produces a product that refuses to guess. In a demo, a tool that always has an answer looks better. In the third week of a real review, the tool that says insufficient evidence when it should is the one still being used.
Checking it yourself
Run a review on a few documents you know well, and try to find a citation that does not say what the cell says. Open every cell if you like; that is what the design is for. The free trial covers ten Documents, which is enough to form a view. How reviewers use the result is on the monday.com page.