Rock paper scissors

How the computer plays, exactly

The AI keeps a table keyed by your last two throws, and in each of those situations it counts what you played next. When your current two-throw context has at least two observations, it takes your most frequent follow-up and plays the counter to it. Until then, and for the whole opening stretch of a session, it throws uniformly at random, so the first rounds are a fair coin. This is a second-order Markov predictor, the classic first weapon of RPS bots, and every ten rounds the message line tells you how many distinct patterns it has collected on you.

One thing worth stating plainly because people ask: the computer commits to its throw before learning what you just clicked. The prediction uses only your history up to the previous round. It cannot peek, and every loss you take is a loss to your own habits, which is somehow worse.

Why you start losing, and what actually works

Humans have three durable tells: repeating a throw that just won, switching after a loss, and cycling rock, paper, scissors in order when trying to be clever. All three are exactly the kind of two-throw context this model eats. If your win rate sags as rounds accumulate, you are not getting worse, you are getting legible.

The optimal strategy against any pattern-learner is genuine randomness, which holds it to an even split forever, and humans are famously terrible at producing it on demand. The practical tip: outsource the decision. Glance at a clock and take the seconds mod 3, or assign throws to the letters of a sentence you are thinking of. Any external noise source beats your intuition, and you can watch the difference in the win rate box within twenty rounds. There is also a second, sneakier option for people who understand the model: since the AI counters your most frequent follow-up in each context, you can deliberately feed it a pattern for a dozen rounds, then play the move that beats its counter. This works, briefly, until your betrayal itself becomes the new pattern. That arms race is the actual game.

Reading the scoreboard

The win rate shown is wins over decided rounds, ties excluded, so 50 percent means dead even. Against this model, sustained rates meaningfully above 50 mean either disciplined randomness or successful double-crossing; sustained rates below about 40 mean the table has your number. Ties hovering near a third of all rounds is normal and tells you nothing by itself. At 25 rounds the game offers a shareable tally, which is roughly the point where the result starts reflecting strategy rather than luck.

Honest limitations

The model is deliberately simple: it looks exactly two throws back, no further, and it does not weight recent behavior over old behavior, so a habit you abandoned twenty rounds ago still haunts the table until enough new data outvotes it. It also does not model its own influence on you, and it never bluffs. Session data lives only in the page: reload and the AI forgets everything, tally included, which doubles as a mercy rule. No history is stored or sent anywhere. And if you want the purest test of the premise, play fifty rounds trying your hardest to be random by feel, then fifty using a clock. Most people never fully trust how big the gap is until they see it.