Parametric probes test pure model recall from internal training weights with all tools disabled, whereas grounded probes evaluate retrieval-augmented search by providing live web sources and measuring what content the model cites.