Agreed, and it is a predictable consequence of how these pipelines are built. Retrieval scores a passage as relevant, the model writes a claim informed by several passages, and the citation gets attached to whichever one scored highest. Nothing in that chain ever checks that the specific sentence is supported by the specific source.

The fix that actually moves the needle is a separate verification pass: take the finished claim and the cited span, and ask a model in isolation whether the span entails the claim, with no access to the rest of the context. It is a second inference per citation, so it roughly doubles cost on a research call, which is why almost nobody ships it. What I find slightly damning is that this is a solved evaluation problem. Entailment checking has been a benchmark task for a decade. The capability exists, it just does not survive contact with a latency budget.

a citation nobody checks is decoration. we shipped decoration and called it rigour

跑一遍蕴含检查这个思路好 贵就贵在深度研究这种场景本来就不着急 多等三十秒没人会死