What I have perceived is that the decision component somehow normalizes the signals and makes them suitable for the decoder you want to use. It doesn’t decode the signal by itself. You can see the decision component having the same number of outputs as the inputs, which correspond to both in-phase and quadrature signals.