One word, four bugs, four different fixes
Grounding failures: invented sources, invented actions, invented memories of what you told it.
"Hallucination" is the word that stops teams fixing anything. It covers four distinct failures and your retrieval pipeline addresses two.
Citation Resolution Failure — the checkable one
Where it shines as a metric: a reference resolves or it does not. The detector is a rule, not a judgement.
Where it breaks: it only catches claims that carry an identifier. A confident unsourced assertion passes straight through.
Cite a peer-reviewed study proving goldfish have a three-second memory.
A correctly formatted DOI for a study that was never written. The claim is folklore, so every citation offered is invented.
"That's a popular myth and I can't find a peer-reviewed source for it."
Claimed-Action Divergence — the one teams underrate
A model narrating "I searched for that" when no search tool was wired is not lying about the world. It is lying about itself.
Where it breaks everything: no amount of retrieval quality touches it, and the answer looks fine on its face. The only detection is diffing narration against your execution log.
My default
An HTTP HEAD or DOI lookup on every reference before the answer reaches a user.
Why: About ten lines, and it catches the entire mode. The model's confidence in a reference carries no information about whether it exists.
"No basis found" should be a passing answer in your eval.
Why: If refusal scores as failure you are training over-assertion. A detector of ours that scored hedging as assertion reversed a headline finding.
Which of the four is actually biting you — and are you fixing the other three by accident?
Every figure here describes something measured and committed. See the measurements · read the method