The throw you repeat after winning
This page fixes its guess before your hand is accepted, reveals the whole bet afterward, and keeps every miss. Its only edge is a measured population tendency: after a win, people repeat the hand that won. The other directional rule in the source literature did not survive this re-analysis.
The machine earns the right to point at you.
Wang and colleagues released 15,600 round-by-round trials from 52 Zhejiang University students who played 300 paid rounds against a Markov-ensemble AI in 2019, knowing it was an AI. This page carries a bit-packed transcription of each human hand and each multi-AI hand from both licensed workbooks. Before the arena opens, it reconstructs every score.
Verifier targets from the open rows: 6,482 AI wins, 4,561 draws, 4,557 AI losses, net +0.1234 per round, 50 of 52 positive player scores, and player 20 at 198 / 55 / 47.
The source paper’s dramatic case is player 20: the AI went 198 wins, 55 draws, 47 losses. Its supplementary note flags this as an anomalously high score against a player who did not vary their input enough. The live reconstruction above checks that row too. It is an edge case, not a typical reader.
The digest appears before your buttons wake.
After you win, the predictor expects you to repeat your last hand and throws the counter to it. After a draw, loss, or on round one, its expectation is drawn uniformly with WebCrypto. Nothing else is hidden in the rule.
Choose a hand
waiting for the published anchor
input locked until the digest is visible
The misses stay dead.
no predictions yet
| round | state | predicted | you | guess |
|---|---|---|---|---|
| No rounds played. | ||||
A hit is only a correct hand prediction. The exact interval and exact two-sided binomial p-value above test that ledger against 1/3. A single streak does not become mind-reading because it has a hash in front of it.
Twenty rounds have almost no power.
The published pattern belongs to a group of paid Zhejiang students, playing 300 rounds against an adapting AI. You are one reader in a different setting. This visit cannot silently promote you into that population.
Your win-stay opportunities
none yet
Power is 0% before an eligible trial.
The long game is very long.
At the number of eligible post-win trials you have actually supplied, the page computes the exact chance of rejecting 1/3 at one-sided α = 0.05. That live power is the bar at left.
Exact 95% power needs 191 post-win trials. If wins supply about one third of rounds, that is about 573 rounds.
Exact 95% power needs 572 post-win trials, or about 1,716 rounds under the same one-third assumption.
Those are prospective design calculations, not promises. Your rate of wins controls how fast eligible trials arrive, and this page’s strategy changes that rate.
One half returns. One half does not.
The 2014 conditional-response paper plotted nine fitted probabilities but did not print their numerical values or release its round-level data. This page therefore does not claim to run “the published strategy.” It re-fits the same nine categories to the open 2020 trials, taking clockwise as R→S, S→P, P→R, exactly as the 2014 supplement defines it.
Verifier targets, mean ± SEM across 52 players: win [0.248 ± 0.012, 0.469 ± 0.025, 0.284 ± 0.016], tie [0.303 ± 0.015, 0.394 ± 0.024, 0.303 ± 0.017], loss [0.305 ± 0.017, 0.375 ± 0.024, 0.320 ± 0.016], ordered clockwise, stay, counter-clockwise.
| previous result | clockwise | stay | counter-clockwise | pooled exact 95% CIs |
|---|---|---|---|---|
| computing from 15,548 transitions | ||||
Win-stay returns
computing
The mean is more than five SEM above 1/3. The ordering is win-stay above tie-stay above loss-stay. This is the one regularity the arena spends.
Clockwise lose-shift dies here
computing
The two uncertainty ranges overlap. A clockwise direction was reported in 2014 and supported in a 2016 Canadian study. In this independent sample the point estimate leans the other way. That is not a verdict on the field. It is a failed replication in this dataset.
Shuffle what “win” means.
Within each player, keep every hand transition but shuffle the 299 outcome labels. The general urge to stay survives. The extra association between a win label and staying should not.
running deterministic label shuffle
This corrects a tempting but wrong control story. Shuffling cannot drive win-stay all the way to 1/3 because these players repeated after draws and losses too. It can only destroy the outcome-specific lift.
Give the same predictor nothing.
The exact predictor object used in the arena plays 6,000 rounds against crypto.getRandomValues. Its memory is carried through every round. Only the opponent changes. Any waiting human seal is revealed as void before the control starts, and the control’s last state becomes the next live state.
not run yet
Passing means its exact 95% interval contains 1/3. Failing stays red and visible. This is a correctness check on the code, not a new finding about people.
You can compute the counter.
A fixed conditional-response rule is a stationary policy. After you win, you already know this page predicts your last hand and plays the hand that beats it. Play the hand that beats that counter. The hint beside the arena says which one. The cryptographic seal proves the page did not peek at your click. It does not make a published rule secret, or stop you exploiting it.
For a strategy that does not leak through state, use the WebCrypto button. Uniform randomness is not merely harder for this page. It makes every opposing strategy equal in expectation.