I did not understand Wei Dai's explanation of how UDT can reproduce updating when necessary. Can somebody explain this to me in smaller words?

(Showing the actual code that output the predictions in the example, instead of shunting it off in "prediction = S(history)," would probably also be useful. I also don't understand how UDT would react to a simpler example: a quantum coinflip, where U(action A|heads)=0, U(action B|heads)=1, U(action A|tails)=1, U(action B|tails)=0.)