It's possible it's supposed to be normative though, obviously people don't run decision theory directly on the output of Solomonoff induction using a pre-progra

It's possible it's supposed to be normative though, obviously people don't run decision theory directly on the output of Solomonoff induction using a pre-programmed or learned utility function, but it's what they should do (but they can't, because it's not even computable). It's the theoretical perfection that all practical systems should attempt to approximate.