Yeah, it's definitely a pain point for me. I've observed when running a local AI on qwen2.5-7b-instruct-uncensored, the performance is *slightly* more reliable, but it's definitely not perfect. Qwen is about the most stable I've used (out of over a dozen _instruct_ tagged models). I mostly just disable stat prompts now and try to construct scenarios in a way that they aren't required. The AI is actually surprisingly good at conjuring skill checks and resolving them without including the details in its response, in a way that keeps the game immersive.