By probe, I mean observe the internal activations. There are methods that can suggest if it's hallucinating or not, and ones that can delete individual pieces of knowledge from the model.