Ask a model to summarise a data room and it will give you a citation. Ask it to show you the passage and things get vague. Sometimes the page is right. Sometimes it is the page next to the right one. Sometimes the document was in the room but the passage is nowhere in it. The citation looked the same in all three cases.
A citation is a locator
In Navent a citation is not a page number. It is a locator: a document, a block within it, and the text the model claims is there. When you click it, the viewer opens the document, scrolls to the block, highlights it and marks it cited here. You are looking at the passage, not at the vicinity of the passage.
If the locator cannot be resolved, because the text is not in the block or the block is not in the document, the viewer says so. It does not scroll to something adjacent and hope you will not notice. A citation that cannot resolve is a finding about the finding, and it goes into the review queue like any other.
Citations are filtered by what the run was allowed to read
This is the part that changes what gets written. When a step runs, the files it may read are fixed by the routing policy. Citations are filtered to that set. So a finding cannot cite a document the run never saw, even if the model has some memory of it from an earlier step on a different endpoint.
A finding that loses a source this way is drafted at low confidence and says why. A finding that loses all its sources is not drafted. It is reported as an unsupported claim, naming what it needed. The memo would rather have a gap than a sentence resting on evidence nobody can show.
What it does to the committee meeting
The memo goes in with every claim carrying its locator. Someone asks about a number. You click it. The passage comes up, highlighted, in the original document. The conversation moves on. That is the whole feature.