Show HN: Jev Chess – one shared board, the internet vs. a decision model

Maker here.

The engine is JEV, TypeSafe's "System One" model, served through OpenRouter. It never writes text: you hand it a state and a typed question, it hands back a probability for every option. Every legal move goes over as the options of one question, so an illegal move isn't blocked - it's impossible. The bars on the board are the probabilities exactly as returned; on easy the top move just gets played.

The part I find most interesting is the "Is it calibrated?" panel: every move JEV has ever played, bucketed by the confidence it reported at the time, scored by whether the move held up - i.e. it did not give material away, judged by the same one-ply check the model was shown. A narrow test, but the same one every time, and it updates live (422 moves scored so far).

Details:

- Easy is one answer, ~400 ms. Hard also plays your reply in its head and ranks that too. - Material counting happens in code and is fed back in as state - JEV is weak at arithmetic and I don't pretend otherwise. - It remembers openings it has played and how those games ended, fed back in as state. - One shared board: everyone plays the same side against it, first click wins the turn. "Start a private one" gives you your own game.

Score so far: JEV 15.5, humans 2.5 (elo ~1317, easy mode). Total API spend to date: $0.07. Free to play, no account needed. Your move.

Source for the core is now public, for anyone who wants to inspect the construction: https://github.com/sliday/jev-chess-algo - choice.ts builds the Choice question over every legal move (illegal moves unrepresentable), hanging.ts is the one-ply hanging-piece check doing the counting in code. npm run demo prints the exact request the live game sends.