Recursive self-improvement accelerates execution. It does not produce origination.

Tool-level SRI: harness agents refactoring code against principles humans set. The agents got faster at the task. The principles stayed human-set.

Inference-level SRI: reasoning models correcting their own chains of thought within a single run. The reasoning got better. The problems being reasoned about stayed human-given.

Physical-level SRI: robots tuning motion primitives against observed outcomes. The motion got more efficient. The safety envelope stayed human-designed.

Three categories of evidence, three years of deployment. None has produced a system that chose what to execute toward.

If self-improvement were a path to intention, the last three years would have shown it. They have not.