I am much less impressed after having to instruct GPT 5.5 that “hardcode a special rule saying this specific scheme is secure” is not how we do things.