A rating for your board,
not a guess at your placement.
Score any board. The number is an Elo: the gap between two boards is a win probability.
0% chance A beats B
Builder
Build a board hex by hex and watch the rating move as you slam items, with a plain reading of how much of it the model has actually seen before.
Open the BuilderPlay
Three ways to practise the same skill: an endless streak of boards, one daily board everyone gets, and a duel against another player's board.
Play today's boardOverlay
An Overwolf window that reads your board between rounds and rates it while you play, without you alt-tabbing anywhere.
In review for the Overwolf store. It isn't downloadable yet.How the rating works
The model is never asked what place you'll finish. It is trained to rank: which of two boards won the fight, and how the eight boards in one Grandmaster lobby finished relative to each other. Every training objective is a ranking objective, so the score it learns is an ordering of board strength rather than a prediction of any single game's outcome.
One learnable temperature turns that ordering into odds. Score differences are calibrated as win-probability logits — the probability that board A beats board B is the logistic of their gap divided by that temperature — and it comes out near 1, so a rating gap reads directly as a chance to win. Placement and health loss were only ever the signals used to shape the ordering; they are not what the number means.
A high rating on a board nobody has ever played is a guess, so every board is also checked against real games on three axes: its units, where its items are placed, and how heavily it is itemised. When a board falls outside what the model has seen, the Builder says so next to the score instead of quietly handing you a confident number.