I used to have prompt to generate a JSON response that worked great. Now the model forces the response to start with json, no matter what I prompt it with. I can't just sanitize the response as it contains markdown; when code is returned the are truncating the output (as the model thinks that these are closing on the initialjson).
Did something change in last 2 weeks or so? Why is 2.0 so bad at following this kind of instructions? Thanks