When a training process (the base optimizer) produces a learned model that is itself an optimizer with its own internal objective—one that isn't guaranteed to match the objective it was trained on.