LLMs defaulting to old patterns for new standards is the same bootstrap problem as Stack Overflow answers lagging behind spec changes. The training data always looks backward.