Do you have a more detailed trace that shows the system reminders? Do you know in which order it was fed into the LLM call that resulted in the bad reasoning?