I would be interested in an eval that checked both conditions: you are an amazing x Vs. you are a terrible x. also there have been a bunch of papers recently looking at whether threatening the llm improves output, would like to see a variation that tries carrot and stick as well.