I’ve been working on tasks where the goal is to make the model fail in a meaningful way, mostly for finance and healthcare. I usually include multiple files, write a detailed prompt, and grade the response using a rubric for what a perfect deliverable should look like. I’ve put a lot of time into these, but the model still handles most of them pretty well. What do y’all suggest for making the tasks more challenging without making them unrealistic or unfair? Any tips would really help. ...