I wonder if you could take advantage of the fact that the LLM is more likely to follow instructions that humans might miss. For example include instructions somewhere in the repo that says you must use a certain phrase in all pull requests, and then you just check the PR for that phrase.
Or maybe require the PR to contain something that is generated by running code, which the LLM may not be able to do without some effort on the user's part.
Or maybe require the PR to contain something that is generated by running code, which the LLM may not be able to do without some effort on the user's part.