OpenAI and other labs are trying to do this within existing structure by explicitly training a specific chain of authority in the inputs: https://github.com/openai/model_spec/blob/main/model_spec.md...

However, you may immediately see how using same input space essentially relies on the model itself to do the judgement which can't be ultimately trusted