Is this something where you have a guaranteed unknown known language model to practice operating in edge cases where LLMs fail because they don't contain all knowledge? Or is it a way to train LLMs to recognize when they're in a situation where they don't have the knowledge they need?