You can actually skip 2, if you model yourself as part of the system as well (i.e. have an Ideal Observer who changes from “has XYZ knowledge of the system” to “observed ABC about the system” for all possible ABC, with respective amplitudes) – but then you're doing a lot of computation you're never going to use for prediction. (It's more elegant, though – even more so if you model yourself as a human, though this'll take more computational power than we have.)