The players. Eikos-27B, an open-weights model (MIT), against Jev, TypeSafe's decision API. Both are general decision models: they read a situation and pick one option from a list, with a probability for each option. Neither searches ahead like a chess engine.

Every move. The side to move gets the board (diagram, FEN, the moves so far, material) and one question, "Which move is best?", with every legal move as an option. Each option is described by the rules: piece, squares, capture, check, mate, stalemate, repetition, and which pieces are left hanging. Both models get exactly the same request.

Fair openings. Each game starts from a different named opening from Lichess's public opening list, in a fixed order. Every opening is played twice, once with each model as White. No game is ever repeated, so every game counts.

Result and ranking. Games end by the rules of chess (checkmate, stalemate, insufficient material, threefold repetition, fifty-move rule), or as a draw at 200 half-moves. Score is wins plus half the draws. The Elo gap comes with a 95% interval. A game where a model does not answer is left out.