Generate controlled synthetic training or demonstration data to improve a model or agent without treating self-generated data as independent ground truth.
