Role Architectures: Applying LLMs to consequential tasks — LessWrong