Did it really beat an 1825-rated bot?

That is the author's reported Iteration 3 outcome after two failed architectures: unguided generation lost to 1000-rated bots and rigid heuristics peaked near 1350. Treat it as one experiment's log, backed by the repo's tests and demo recording.

Why not just let Jev pick the move?

Iteration 1 tried that and got hallucinations, passive shuffling, and single-ply blunders. The working design only shows Jev engine-evaluated candidates with exact scores and principal variations, never raw legal moves.

Without a TypeSafe key, does the bot still play or only search locally?

Only the strategic layer needs one. The backend on port 8765 runs the 0x88 alpha-beta search on its own, and a circuit breaker drops to local evaluation when the Jev calls time out or fail, so a missing key costs you candidate selection and priority calls rather than the game.

How is the system tested?

Extension unit tests cover move generation, FEN tracking, and turn guards, while backend tests cover the engine and decision pipeline, run as npm test and npm run server:test.

What is the deterministic engine actually capable of?

It uses a 0x88 board representation, generating 7.3 million nodes per second and reaching depth 6 to 12 in under 600 ms, with alpha-beta negamax plus iterative deepening, quiescence search, null-move pruning and late move reductions. A dual 32-bit Zobrist transposition table and a 36-line opening book sit on top, with tapered evaluation interpolating between opening and endgame tables by material phase.

How much freedom does the model actually get?

A dynamic trust window. The fusion engine lets the selected candidate deviate from the engine's best move by up to 40 centipawns, scaled by model confidence, under an absolute blunder ceiling of 70 centipawns. Forced mates, single legal moves and theory matches bypass that entirely. The window is what keeps creative strategic play from turning into a tactical blunder.

How do I run it, and what license is it?

Node.js 18 or newer with npm or pnpm, a Chrome or Chromium browser, and optionally a TypeSafe key. Copy server/.env.example to server/.env and set TYPESAFE_API_KEY with PORT 8765, HOST 0.0.0.0, JEV_MODEL jev-latest and DEBUG_JEV true, then npm run server for the backend and dashboard on localhost:8765, and npm run build before loading dist/ as an unpacked extension. MIT licensed.

CommunityJev Tools & CLIs

Chess With Jev

Hybrid chess experiment pairing Jev strategic candidate selection with a deterministic alpha-beta search engine inside a Chrome extension that plays on Chess.com with live telemetry.

See highlights

Open github.com

Stars: 1
Language: TypeScript
Published by: Community
Author: Deepender25
Last verified: Oct 2026

Deepender25/chess-with-jev-

User repository on GitHub

View on GitHub

An empirical experiment fusing TypeSafe AI's non-autoregressive Jev model (System 1) with a high-performance 0x88 alpha-beta search engine (System 2). Includes live Chrome extension tracking, trust window fusion, and real-time telemetry, capable of defeating 1825 ELO bots.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

This project is an empirical investigation into hybrid decision architectures for strategic game environments. It explores the integration of TypeSafe AI Jev (a non-autoregressive System 1 model) with a high-performance deterministic chess search engine (System 2) operating in a live browser environment.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Jev picks only from 3-5 engine-verified candidates with exact evaluations, so illegal moves never reach the board.
Fusion engine lets Jev deviate up to 40 centipawns by confidence, with a hard 70-centipawn blunder ceiling.
Deterministic engine uses 0x88 boards, iterative deepening to depth 6-12, quiescence, and a 36-line opening book.
Board observer settles state after 300ms and coalesces requests; a circuit breaker falls back to local eval.

Watch out

Reported outcomes come from the author's own iteration table and demo recording against engine bots, not an independent replication. 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.

More Jev resources like this

Jevest

GitHub Action that reviews pull requests with a read-only finder agent plus a refuting verifier, using Jev only to triage, route, label, and gate merging.