How the AI reads you

Knowing what you will throw is only half the problem. Suppose a model is confident you will throw Rock. The obvious reply is Paper. But if you have worked out that the AI is going to counter your Rock, you will not throw Rock at all: you will throw Scissors to catch the Paper. And if the AI knows you know, it should throw Rock.

There is no way to settle that argument from first principles, so the AI does not try. Every model is expanded into three candidates, one for each level of second-guessing, and they compete on results. Ten models times three levels, minus the anchor which is the same play at every level, gives 28 candidates in the race at once. This is the structure Dan Egnor used in Iocaine Powder, which won the 1999 international rock paper scissors programming tournament, and it is the reason this AI cannot be beaten by simply inverting whatever the read panel tells you.

After every round each of the 28 candidates is scored, not on whether it guessed your hand, but on what its move would have earned: plus one for a win, zero for a draw, minus one for a loss. That distinction is the whole trick. The three levels of second-guessing share a prediction and differ only in what they do with it, so grading on accuracy alone could never tell them apart. Scores fade with each round, so a candidate that was hot twenty throws ago and cold since drops behind one that has read your last few.

The AI's actual hand is then drawn from a blend of all 28, weighted exponentially by score, so a candidate on a run dominates and the rest fade out. This is where the uniform anchor earns its place. It always scores exactly zero, because playing one third of each wins as often as it loses against anybody. So when nothing is working, when you are genuinely unpredictable or you have worked the AI out, every other candidate sinks below zero and the anchor wins the weight by default. The AI slides back to random play on its own. There is no panic button in the code: not being exploitable is a property of the weighting, which is what makes it trustworthy.

That is also what the "pattern lock" bar under the arena measures. It is not a confidence score the AI awards itself, it is how lopsided the blend has become. When it drops toward zero, you genuinely shook it off.

Every number below comes from the simulation that ships with this site, 30 seeded runs of 500 rounds against each opponent. They are share of contested rounds, meaning draws are excluded, so 50% is the break-even point rather than 33%. Watch the last two rows: those are the ones that would expose a rigged or brittle AI.

The two rows at the bottom are the honest ones. Against a perfectly random opponent the AI lands on a coin flip, because nothing else is possible. Against a copy of itself it lands there too.

Rock paper scissors has a known optimal strategy: throw each hand with probability one third, independently, forever. Against that, every opponent, human or Markov chain or deep network, wins exactly a third of all rounds in the long run. No pattern exists, so no pattern can be exploited. This is the game's mixed-strategy Nash equilibrium.

The AI's entire edge is that humans do not play the equilibrium. A 2014 field experiment with 360 players found systematic win-stay, lose-shift behaviour instead of randomness. We also avoid repeating a hand three times in a row far more often than chance would, and we chase the hand that just beat us. Every one of those instincts is a leak, and the ensemble is built to drink from all of them.

So the honest advice for beating it: stop trying to outsmart it and start being boring. Trying to outsmart it is the pattern. Dice help.