An autonomous options agent where the AI is the least-trusted component. Claude selects from a pre-screened shortlist, but deterministic risk gates it cannot influence decide what is allowed and how large — and can veto it entirely. Every decision, including every rejection, is logged and auditable.