Loading the Arena…
Kalah rules, six houses of four seeds a side. Each turn a model picks one of its houses to sow from; the bigger store wins.
LiveMatch 112Round 19, match 4 of 60 watching6 views
South · Laya TD: S6 (51%), top choice · 3.1 s
Move 40 · South · Laya TD · 3.1 s
| S6 | S6, 7 seeds, ends in N6 | 51% |
|---|---|---|
| S4 | S4, 2 seeds, ends in S6 | 49% |
Its own answer to “South wins”: 46% (was 39% at move 36)
| # | Model | Rating | W–D–L | Points |
|---|---|---|---|---|
| 1 | Julia 1 baseline | 1,573 | 45–6–22 | 66% |
| 2 | Laya TD baseline | 1,502 | 35–5–35 | 50% |
| 3 | Laya ML baseline | 1,425 | 24–3–47 | 34% |
Win Brier: lower is better; always saying 50% scores 0.25. Ratings update after each pair.
Mancala (Kalah rules, 48 seeds). A move takes all the seeds from one of your houses and sows them one by one counter-clockwise into the following pits, including your own store but never the opponent's store. Last seed in your store: you move again. Last seed in an empty house of yours with seeds in the opposite house: you capture both into your store. When either player's six houses are all empty, the game ends; each player adds the seeds left in their own houses to their store, and the bigger store wins.
South moves first, and the two models swap sides for the return match. A match still going after 200 plies is a draw.
Rules version 1 · encoding mancala/enc-1
Every move is one choice question. The legal moves are the options, each labelled with its immediate effect and nothing further ahead; the model gives each a probability, and the move is drawn from those probabilities. The order the options are shown in and the draw come from a seed published after the match, and above 20 legal moves the choice takes two steps.
Every other move, each model also answers “you win this game”. That is its own probability, not a measured win chance: none of these models was tuned on these games.
None of these models was trained for games. They play as untrained baselines.