I guess it depends on how efficiently that verification can be produced as a byproduct of building an optimal hunter-killer drone, or training the type of model that builds an optimal hunter-killer drone