Feedback means using outcomes to adjust later actions. Treat the taught AlphaEvolve search and surrounding intentmaking/sensemaking cycle as two feedback levels. The important difference is what can change: inner feedback operates under the current evaluator, or scoring rule; outer feedback can question that rule or the scientific goal itself. Better candidate scores therefore need not repair a rule that tests the wrong thing. This is a bounded working analogy, not a universal model of science. Its sources are Google DeepMind’s May 14, 2025 developer report and Bäuerle and colleagues’ May 7, 2026 v1 interview-and-observation study of 11 selected mathematicians, not a controlled test of oversight effectiveness.