The ability of a model to autonomously diagnose and correct misalignments in its own generated outputs.