Low-consequence AI applications should use cheap self-verification. High-consequence applications should pay for technologically independent assurance.