Stateful AI · a learned core, playable
a model that learns your tells
Play rock-paper-scissors. hunch commits its throw before it can see yours — it has no idea what you're about to click. All it has is your history. Within a dozen moves it starts winning, because almost no one is actually random. This is proof that adaptive software can read your behavioural state from a handful of moves.
hunch runs a small panel of experts, each a different theory of how you pick your next move. Every round, each expert predicts your next throw from the game's history; hunch tallies a weighted vote and plays the move that beats the favourite. Then it sees what you actually did and uses multiplicative weights (η = 0.4) to reward whichever experts were right — so the theory that fits you rises to the top within a few moves. No server, no model file, no metered API — about a hundred lines of plain JavaScript.
The panel:
• favourite — your most-frequent move so far
• repeat — you'll throw your last move again
• cycle — you'll advance R→P→S from your last move
• beat-my-last — you'll throw what beats hunch's last throw
• markov-1 — your next given your last move
• markov-2 — your next given the last full exchange
• win-stay-lose-shift — repeat after a win, switch after a loss