Tom wrote: "How would Bayes' Theorem go about estimating the chance that a video poker machine is gaffed? I understand that it would use the results of play on the machine, but if it needs one's initial estimate of the probability, isn't it dependent on a factor that, it's assumed, has no value?"
Almost all gambling involves an initial guess at the EV. One example I can think of that does not would be video poker with autohold (and a non-gaffed machine). Another example would be a kiosk promotion where you have equal weighted choices, also rare, most are weighted or even fixed. But the EV of most video poker (no autohold) varies widely depending on how on plays in the casino environment. That doesn't mean the initial guess as no value, it's either a high quality guess or a low quality guess, you use the actual results as they come in to fine tune your final estimate of what the true EV is. The point is that there is feedback, a control loop, you don't just blindly dump all your bankroll on the assumption that the initial guess at EV is 100% correct. In the real world, results do matter. You're either winning or losing and it's either do to variance or an incorrect EV guess. You should figure out which it is and if it's due to an incorrect EV guess, you should really make an adjustment, hence an adjustment based on results.