I agree. It's a cool little experiment, but it seems postdictive, not predictive.
Computers aren't predictive when it comes to simulations, if their predictions are indistinguishable from the real thing because they are simply continuations of the simulation. The Weisner-Gross bot in the paper linked by OP runs the simulation into the future and maximizes its score of "entropy", calculated by an equation provided by the researchers. Tom7's NES bot runs the emulator into the future and maximizes whatever memory values the example input demonstrated were valuable. They both return to the present and choose "optimal" paths that achieve their objective by rejecting choices that played out poorly. Perhaps in a simulation where the results of the input were nondeterministic, or in a fuzzy simulation of an external environment out of the program's control, it would be valid to call a program predictive.