Because LLM-based service outputs are fundamentally not-reproduceable. We have no insight into any of the model settings, the context, what model is being run, etc.