About Beat the Predictor
Rock-paper-scissors is a game of pure chance — if both players are random. People are not. This predictor watches how you play and learns: nine small models track your favourite move, your last one to four moves, and how you react to winning and losing, and the ones guessing best lately get the most say. It commits to its move before you click, and plays whatever beats the move it expects.
Everything it claims is tested. Against 10,000 simulated random players it wins exactly the third that chance allows, and against scripted habits it climbs from 43% to 99%. After 30 rounds you get its win rate with an exact binomial p-value against the one-in-three line, and the habit it caught, corrected for the number of habits it tried so it does not cry wolf. It runs entirely in your browser; your moves go nowhere unless you share them.
Before each move the predictor seals its choice and shows the SHA-256 fingerprint above the buttons. After you play, it reveals the move, the secret and why it chose it. Paste the revealed text into any SHA-256 tool and you get the same fingerprint — proof it decided first.
How to Play Beat the Predictor
Click Rock, Paper or Scissors (or press R, P or S). The predictor has already locked in its move, chosen from everything you did in earlier rounds; the round is scored the instant you click, and the tape underneath keeps the history. The headline is its win rate. Chance alone gives it 33.3%, so anything well above that means it has found something predictable in how you play. From 30 rounds the verdict panel turns that into an exact p-value, and the pattern panel names the habit it exploited.
How the Predictor Works
It is an ensemble of variable-order Markov models. Four of them look at your last one, two, three or four moves; four more look at the last one to four rounds including its own replies, which lets it learn how you react to a win, a tie or a loss; one simply counts your favourite move. Each model predicts your next move from what you did the previous times the same situation came up, with older rounds fading so a change of habit is noticed. Models are weighted by how well they have guessed lately, the weighted guess wins, and the predictor plays the move that beats it.
Reading the p-Value
If you were choosing truly at random, the number of rounds the predictor wins would follow a binomial distribution with a one-in-three chance per round. The p-value is the exact probability of it winning at least as many rounds as it did under that assumption. Below 0.05, the page says it read you; below 0.01, that a random player would give it this many wins about once in a hundred games or less. Thirty rounds is the minimum for a verdict because shorter games cannot separate a real habit from a lucky streak.
The Habits It Looks For
Fifteen specific tendencies are tested: favouring one move; repeating or cycling forwards or backwards; and, after a win, a tie or a loss separately, staying put or switching up or down. The classic human tell is win-stay, lose-shift, and especially switching after a loss to the move that just beat you. Because testing fifteen habits gives fifteen chances to find one by luck, each p-value is multiplied by the number tested (a Bonferroni correction) before the page says a pattern was caught. On random players that false alarm fires in about 3.5% of 100-round games.
How to Actually Beat It
The only reliable way is not to have patterns: flip a mental coin you cannot steer, or use something outside your head, like the second hand of a clock. Out-thinking it works for a while, because switching strategy resets what it knows, but its models forget old rounds and it catches the new habit within a dozen rounds. Staying near 33% over a hundred rounds is genuinely hard, which is what makes it a good test. Winning more than a third of rounds yourself means you have found a pattern in it — and it will be learning from that too.
Related tools: Can You Fake Randomness? to test your coin flips with the runs test, the Coin Flip and Dice Roller for real random outcomes, and the Reaction Time Test. Browse every Fun & Novelty tool for more.
Method: ensemble of nine Markov predictors (orders 0–4), evidence decay 0.96, accuracy decay 0.85, exponential weighting. Statistics: exact one-sided binomial test at p = 1/3; habits tested with exact binomial tests and a Bonferroni correction. Validated on 10,000 simulated random 50-round games (33.31% win rate) and eight scripted strategies. Your moves stay in your browser unless you share or save them.
Frequently Asked Questions
Does the predictor cheat by seeing my move?
No. It chooses its move from the history of earlier rounds before you click, and the page never feeds your current move into that choice. Subscribers can check this for themselves: in think-out-loud mode it shows a SHA-256 fingerprint of its sealed move before you play, then reveals the move and the secret that produce exactly that fingerprint, which you can re-hash in any SHA-256 tool.
Why is 33% the number to beat?
Against someone playing truly at random, every strategy wins a third, ties a third and loses a third — no algorithm can do better, because there is nothing to learn. So the only way the predictor climbs above 33% is by finding real structure in your choices. We checked this on 10,000 simulated random games: it won 33.3% of 500,000 rounds, as it must.
What does the p-value mean here?
It is the probability that a player choosing at random would let the predictor win at least as many rounds as it did against you. A p-value of 0.01 means that would happen in about one game in a hundred, so the result is hard to blame on luck. It is computed exactly from the binomial distribution — no approximation — and the page only says the predictor "read you" when p is below 0.05.
How does it learn my habits?
Nine small models watch different things: your overall favourite move, your last one to four moves, and your last one to four rounds including its own replies (which is how "after losing you switch to what beat you" gets caught). Each keeps counts of what you did next in each situation, with older rounds slowly fading. The models that have been guessing well recently get more say, and the combined guess decides its move: the one that beats what you are expected to play.
Why is it so hard to play randomly?
Because people avoid repeating themselves, balance their moves out on purpose, and react to the last result — winners tend to stay, losers tend to switch, often to the move that just beat them. Studies of real rock-paper-scissors play have found exactly these win-stay, lose-shift tendencies. True randomness has streaks and repeats that feel wrong to produce, which is the same reason our Can You Fake Randomness test catches most people.
