I agree with this; that's why I was careful not to use the examples you mentioned. Quoting incorrect training knowledge would be an unavoidable issue if your pr

I agree with this; that's why I was careful not to use the examples you mentioned. Quoting incorrect training knowledge would be an unavoidable issue if your probe can only say "it's quoting something", and as far as I know it can't do better than that. But I have seen issues with prompts where a creative-writing prompt and just asking a question look similar, and in that case it could help to know which one it thinks it's doing. Gemini itself has a funny verification button where it more or less Googles every sentence the model writes and tells you if it seems like it made it up or not.