Right, but under this construction you progress very slowly and naively in reconstructing the demand distribution. You protect against adversaries who want to harm prices but not adversaries that will prevent progress, i.e. consensus on optimal values w.r.t the chosen objective function.
I don’t necessarily agree that we know the true cost of an opcode too, but the goal of using a learning protocol with limited feedback would be to approximate the true cost with respect to the current distribution of users/buyers.