Jev attempts chess
A System One model picking one legal move at a time, with no search behind it. It never plays an illegal move, and it plays badly — that gap is the whole demo.
Jev’s own guidance puts chess-like planning outside what a fast decision model should be asked to do: positions that need lookahead, a search tree and an evaluation function belong to a dedicated engine, or to a large reasoning model that can think in extended steps. Jev answers in one classification and keeps nothing between calls — no plan, no memory of the last move, no idea what the reply will be. This page wires it up to a real board anyway, so the shape of that limit is something you can watch rather than something you have to take on trust.
Jev vs Minimax
Setup
Black runs the depth-limited minimax below. A few hundred lines of material counting, and the thing Jev is here to lose to.
The seed fixes the opponent’s choices, so the same seed and mode replay the same game. Changing the blunder threshold re-marks the moves already played without asking Jev anything again.
The decision
No moves yet
Play a move to see the position summary Jev received, the legal moves it chose between, and what the referee made of its pick.
Game log
No moves yet
Every half-move lands here. Jev’s are marked with what they cost against the referee’s best line.