User repository on GitHub
Chess where both players are TypeSafe's Jev model: every move is a typed Choice decision
- Forks
- 3
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
TypeSafe Chess — TholeG
Two Jev players playing chess with no engine: code owns the rules and the tactical facts, and the model returns one Choice per legal move with a probability for each, plus a Score for the position and a Noul for sharpness.
- Stars
- 10
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Author
- TholeG
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is TypeSafe Chess — Dimesio.
About this project
Two virtual chess players, both powered by TypeSafe's System One model Jev. No engine, no generated text: each move is a typed decision that the model returns with a full probability distribution, and the code does the rest. Optionally, a Monte Carlo Tree Search uses those same distributions as policy and value, AlphaZero-style, to play noticeably better.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- An illegal move is impossible by construction: chess.js generates the legal moves and each becomes one Choice option, so the model can only pick from them.
- Optionally MCTS uses the same call as policy and value, PUCT over a tree where every node costs one API call and there are no random rollouts.
- A Stockfish regret harness scores the fast move and the MCTS move on identical middlegame positions, reporting centipawn loss per move.
- The UI shows visits, prior and Q for the top candidates and says whether the search confirmed or overruled Jev's first instinct.
Watch out
Cost is the first thing to check, because the two modes differ by an order of magnitude: a fast move is roughly 2-3k tokens and about 350 ms end to end, while an MCTS move with 16 evaluations runs around 40-50k tokens and 3 seconds, and a single MCTS game reaches about 2M tokens. Strength numbers carry a stated caveat - the elo ladder plays a lite WASM Stockfish at fixed depth, so the figures are relative rather than FIDE, and eight games give roughly plus or minus 200 Elo. On the measured side, the regret script scores the fast move and the MCTS move on identical positions and reports centipawn loss: the sweep table puts fast mode at 168-190 cp average loss and the best MCTS configuration at 25 cp, over two rounds of 20 tactical positions at oracle depth 12 with 16 evaluations. Worth knowing that the first MCTS version lost to the single-call player, because Jev's Score is an absolute judgment that calls every child position clearly better once one side is materially ahead, leaving the tree nothing to rank on; the fixes are code-side tactical facts. The README continues past the point this listing's source text ends, cut off inside the tuning-loop discussion. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
There is no text to parse and no illegal move to handle. chess.js generates the legal moves and the server turns each into one Choice option, so the model can only ever pick from the legal set and an illegal move is impossible by construction. Alongside it, the server computes the facts a rules engine knows for sure - the piece, captures, checks, castling, promotion, whether the moved piece can be captured on arrival and whether that square is defended - and passes those as the option description.
It made it worse at first, and the repo says so. The initial version lost to the single-call player because Jev's Score is an absolute judgment: once one side is materially ahead it says clearly better for every child, so the tree had nothing to rank on and its choices were noise. The fix kept what code can compute in code - exchange estimates from a capture-only lookahead and mate-in-one flags in the option descriptions, plus a leaf value blending Jev's read with the settled-material delta. After that, the sweep puts the best configuration at 25 cp average loss against fast mode's 168-190 cp.
The README gives a cost guide rather than a bill. A fast move uses roughly 2-3k tokens; an MCTS move with 16 evaluations uses roughly 40-50k tokens and about 3 seconds. Defaults come from MODE, MCTS_SIMULATIONS and MCTS_CONCURRENCY. For benchmarking, one MCTS game runs to about 2M tokens, which is why the elo ladder is only run in small batches - a two-game match at 16 simulations and 6 parallel evaluations cost about 47k tokens per move against 2.6k for the single-call player.
Reasonably, because the repo benchmarks against something external. Two Jevs beating each other proves little, so it includes a UCI wrapper around the stockfish npm WASM package and three scripts: regret for per-move centipawn loss on fixed positions, match to play a Jev player against Stockfish limited by UCI_Elo, and elo to fit a rating by maximum likelihood with a 95% interval against a ladder. The README flags that Stockfish's limiter is calibrated for the full engine at tournament time controls, so lite-WASM numbers at fixed depth are relative.
It needs Node.js 20+ and a TypeSafe API key. Clone the repository, run npm install, export TYPESAFE_API_KEY, then npm start, which builds the React client and serves it at http://localhost:3000. The API key stays on the server, and the browser only talks to local endpoints.
The current game state as JSON plus, in one request, a Choice question with one option per legal move, a Score question for the position evaluation, and a Noul for whether the position is tactically sharp. Fast mode takes about 350 ms end to end, and the model can only pick one of the legal moves generated by chess.js.
Jev supplies the policy, a prior probability per legal move from the Choice answer, and the value, from the Score answer mapped to minus one to plus one. mcts.js wires those into a PUCT tree search where every new tree node costs one Jev call, there are no random rollouts, and the budget is counted in API calls with transpositions cached for the whole game.
More Jev resources like this
30GitHub stars
Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.
Jev Demos & Experiments#community#jevbest#catalog-import
21GitHub stars
Browser top-down driving sim that sends sensor JSON to Jev every 200ms with four typed driving questions in one call, then applies confidence gating and reflex rules to steer.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
vinilana/live-jevJavaScript
vinilana/live-jevJavaScript12GitHub stars
Chrome extension that labels each X post with Substance, Humor, Chit-chat, Promo or Junk plus AI-written and off-topic flags, collapsing hidden posts to one line.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bohutang/siftJavaScript
bohutang/siftJavaScript11GitHub stars
Chrome extension that labels each X post with intent plus rage-bait, synthetic, and undisclosed-ad signals from Jev, with threshold-only rendering and local caching.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
manifoldor/xtagsJavaScript
manifoldor/xtagsJavaScript3GitHub stars
Browser agent where Jev makes every navigation decision and agent-browser performs clicks and typing, stopping when the named goal page is reached or blocked.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
kyrylosyzonenko/jev-browseJavaScript
kyrylosyzonenko/jev-browseJavaScript3GitHub stars
Local three-panel playground for trying TypeSafe judgments, with 110 editable examples across 22 categories including A/B comparisons, running one request per Run without bulk automation.
Jev Demos & Experiments#community#jevbest#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.