Do Rosario, Younes & Pires (2026) define 'grounding violation' as distinct from 'factual hallucination,' and name 'parametric leakage' as the mechanism that turns an ungrounded answer into confident fabrication

ragnlp-evaluationhallucinationgroundingai-safety

Building on an earlier confidence-target analysis of hallucination as miscalibrated guessing (cited by the paper as its starting point, not yet independently read by this vault), a RAG-evaluation preprint sets out terminology that separates three things usually collapsed into one word. A "factual hallucination" is simply a wrong answer. A "grounding violation" is the system answering without support from its knowledge base — a property of how the answer was produced, independent of whether the answer happens to be correct. "Parametric leakage" is named as the mechanism behind most grounding violations: the answer is drawn from the backbone model's training knowledge rather than from anything actually retrieved.

The paper states the relationship between the three directly: a grounding violation is what turns a correct-but-unverified lucky guess into a dangerous kind of wrongness in the retrieval setting, because "whenever parametric memory does not happen to contain the right answer, an ungrounded response becomes a confident fabrication." This is a definitional separation — distinguishing a correct guess with no support from a sourced answer — rather than a quantitative finding about any specific system, and it is recorded here as terminology rather than as a measured rate.