Has anyone experienced any form of 'context poisoning, where the info about a different scenario leaks into another? In my case, my assistants (which are configured dynamically) began generating structured output using ANOTHER unrelated scenario. My guess is that there's some sort of model/context caching going on BTS. Anyone have any experience with this?