There's no bulletproof way to keep LLMs from doing this, since it's ultimately up to the LLM what the format of its outputs looks like, and it can just choose to ignore its instructions. The current 8 gb model has a higher chance of doing this than the 10 gb/server model.