Limiting models to learn from trusted adapters only is a nice fix for the semantic drift caused by sheer param size. Getting alignment back to 'what' models learn is the real step forward.