Keep simple replies local. Choose a model explicitly. Ask your configured models in parallel and retain their full answers.