Then you look for largest probability*utility , which you generally do by trying to find a way to demonstrate A>B which you can do in many cases where you can't actually evaluate either A or B (and many cases where you can only evaluate A and B so inaccurately that outcome of comparison of evaluations of A and B is primarily dependent on inaccuracies).

Furthermore, a "probability" is a list due to loss of statistical independence with other variables. edit: the pieces of information are very rarely independent, too. Some reasoning that 3 is more likely than other digits would not be independent from 2 being a bad choice.

edit: also, holy hell, trillionth prime does end with 3.