AI systems answer fluently, and fluency is regularly mistaken for correctness. The only load-bearing difference between the two is the trace: which context produced this answer, and where did it come from? Without the trace, every review turns into an excavation.
Why citing after the fact fails
The common reflex is to decorate answers with sources afterwards. Anyone who answers first and searches second finds evidence for the answer rather than for the truth. Real traceability happens earlier, in the selection of what the system reads at all.
There is a technical problem on top. When the model writes the citation itself, that citation is text like any other and can be just as wrong as the rest. A footnote that looks plausible and points at a passage which does not support the claim is worse than no footnote, because it simulates checking and lulls the reader into skipping it.
Traceability is not a feature you retrofit. It is a property of the architecture.
Selection produces evidence
A system that selects its context deliberately knows at every moment what it has read. The evidence falls out as a by-product, because the selection has to be logged anyway. It only becomes useful, though, when it points at a passage rather than at a document. Telling someone it is somewhere in these eighty pages is not auditability, it is homework.
What a usable trace contains
A trace that only names the sources used tells half the story. What the system considered and then rejected is often the more interesting half. So we record three things: which candidates entered the selection, which of them went into the window, and which rule removed the rest. When an answer is wrong, the fault sits in the discarding more often than in the wording, and without the rejected candidates that is exactly what you cannot see.
This is not free. Traces occupy storage, and they contain extracts from sources that may be confidential. Keeping them means deciding for how long and who may look. We therefore keep them short and in the same place as the data they came from, rather than copying them into a separate system.
What traceability does not deliver
A trace does not make an answer true. It shows what the answer rests on, which is a different thing. If the source is outdated, wrong or misread, the trace leads cleanly to a false statement. The gain is that the error becomes findable instead of staying invisible. Selling traceability as a guarantee of quality is selling something it is not.
In the audit
The difference shows most clearly when somebody from outside asks. A traceable system answers the question of how it knows something in seconds and hands over the passage. An unevidenced one does not answer it at all and repeats its claim instead. In regulated environments that difference decides whether a system may be deployed at all.