What is the difference between retrieval and verification?

The confusion is understandable because the two operations look alike from outside. A system searches, finds documents, and produces an answer citing them. What has actually happened is that documents matching the query were located and an answer was generated in their presence. Whether the answer follows from those documents is a separate question that nothing in the pipeline asked.

The gap shows up in three ways. The retrieved document is topically relevant and does not contain the specific fact asserted. It contains a related fact that has been generalised into the claim. Or it contains the fact with a qualification — a date, a population, a condition — that did not survive into the answer. All three produce a well-sourced sentence that the source does not support.

Verification is a different operation and it is checkable. Does this exact claim appear in this exact passage, and does the passage carry conditions the claim omits. That question has a yes or no answer, can be performed by a person in seconds when the passage is supplied, and can be performed mechanically for anything expressible as a match. It also fails, which retrieval never does.

The practical consequence for anyone building on retrieval is that citations must point at passages rather than documents. A citation to a forty-page policy tells a reader where to start looking and does not let them check anything, so the check does not happen. A citation to the paragraph converts verification from a project into a glance, and the difference in whether it actually gets done is enormous.

There is a second consequence for evaluation. A retrieval system is usually measured by whether relevant documents were found, which is the wrong measure if the output is a claim. The measure that matters is whether each claim in the answer is supported by the material cited, which is a lower number and a more useful one. Systems that report high retrieval quality and produce unsupported claims are not contradicting themselves; they are measuring the operation they perform rather than the one users care about.

None of this is an argument against retrieval, which remains the single most effective way to ground an answer. It is an argument for not treating grounding as verification. The material being present makes the right answer available; it does not make the produced answer right, and only a check on the specific claim closes that gap.