Tabula

Pick a model for each side, press start, and watch two models play a real game on a realistic board. Each one gives a one-sentence reason for its move, typed out live in its panel.

The engine owns the rules

The most important design decision is that the model never makes a move directly. On each turn the engine rolls the dice and computes every legal play. The model is only asked to choose an index from that list and explain why. A model can therefore never make an illegal move however it formats its reply, which matters in a game as fiddly as backgammon.

The prompt shows the board from the model’s own side, with pip counts, the opponent’s blots and each legal play pre-scored for hits, resulting pip count and exposure. If a call fails, times out or cannot be parsed, the engine’s heuristic plays that move instead. It never hands the move to a different model, so no model is credited with another’s play.

Ratings, not just win counts

Finished games are stored in SQLite and every model earns a Glicko-2 rating with a ± uncertainty that shrinks as it plays. Beating a strong, well-established model counts for more than beating an uncertain or weaker one. A matchmaker suggests the most informative next pairing, so a growing roster gets rated fairly without every model playing every other.

More than who wins

Every decision records latency, tokens, a notional API cost, whether the model actually answered, and how often its pick matched the engine’s best play. Dice make results noisy, so the per-move quality signal builds up much faster than wins and losses do.