Methodology
AI hallucinations are also a verification problem
Why a convincing answer, a citation, and an independent check are three different things.
Related reading: our original Medium post. This company-site edition reflects the current product scope.
Fluency does not establish correctness
An answer can be clear, detailed, and wrong. Asking another model whether it sounds right may help identify candidates for review, but it does not replace evidence from the source that establishes the fact.
Verification starts by deciding what the claim means. Which entity, period, unit, and relationship is being asserted? Without that scope, even a correct-looking number may answer a different question.
A citation is a lead, not a verdict
A URL can point to a real document while the answer misstates it. Publication metadata can establish a title or date without establishing that the paper supports a scientific conclusion. The scope of the check must match the scope of the claim.
Groundtruth uses authoritative sources for the facts a dataset is designed to check. The row should retain enough provenance to inspect that relationship. Source availability and completeness are part of the result, not details to hide.
Preserve uncertainty
A source timeout does not prove an assertion false. A missing record may reflect incomplete coverage. A question outside a verification rule remains unsupported by that rule.
Keeping these situations separate from contradictions gives a team a more useful failure analysis. It also prevents a confident grade from disguising a weak reference answer. Human review remains necessary for ambiguity and unsupported cases.
Use verification to guide the next experiment
After reviewing the baseline, choose a narrow intervention and evaluate it on separate data. Preserve the old outputs, the new outputs, and the rules used to compare them.
This is a way to measure bounded factual behavior. It does not guarantee that an AI system will never hallucinate, and a result on one proof set is not a universal model reliability score.