Finding a source does not prove a conclusion

A retrieval system returns the passages closest to the question. Proximity is not support. A document can be about exactly the right subject, mention the same terms and still not sustain the claim being written.

The failure that worries us most has this shape: the citation exists, it is relevant, and it does not say what the sentence claims. The reader sees a reference and assumes the support, because checking it means opening the document.

We have adopted two habits. The first is to anchor every claim to a specific fragment, not to a whole document: if the anchor does not appear, the claim is withdrawn before it reaches the text. The second is to evaluate separately two things that are usually measured together: whether the correct passage was retrieved, and whether the retrieved passage supports the conclusion. The second question needs human reading and admits no shortcuts for now.

The cost is visible: the system answers less and abstains more. In exchange, when it does answer, the relation between source, meaning and conclusion can be examined without reconstructing it from memory. In the sets we work with, a well-explained abstention has turned out to be more usable than an answer with a citation that holds nothing up.