gpt 5 system prompt explicitly tells it to be careful with trick questions and arithmetic and it still fucks up. these system prompts are cheating, i wonder if they are customized based on the prompt. does some classification model identify the prompt as a trick question and then selects that system prompt?