Can a language model beat a simple search bot at Congkak?
Models and bots play a mancala game with a relay rule that makes single moves long and hard to predict. A referee checks every move. A script recomputes every number on this page from the replay files.
Loading results.
Against the search bot
Can it predict its own move?
Before every move, the model says how many seeds its store will hold when the move ends. The relay rule makes that hard to work out. The baseline is a one-line rule that ignores relays and captures.
Head to head
Each pairing plays an even number of games, swapping who moves first, so the first-move advantage cancels out.
Every player
Win rates carry a 95% interval: the range the true rate plausibly sits in, given how many games were played. Two players whose ranges overlap have not been shown to differ.
The game in one paragraph
Each player has seven houses and a store, with seven seeds in every house. On your turn you lift all the seeds from one of your houses and drop them one at a time into the houses that follow, your own store included, never the other player's store. If the last seed lands in your store, you go again. If it lands in a house that already had seeds, you lift them all and keep sowing (the relay). If it lands in an empty house on your side, you capture the seeds in the house facing it. The game ends when the player to move has nothing to sow; more seeds in store wins.