User repository on GitHub
TypeSafe'in Jev karar modeli gerçek bir online 2048 sitesinde oynuyor — hamle başına tek API çağrısı, tek anahtar.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jev2048 — erhanmeydan
A Playwright agent that lets TypeSafe's Jev decision model play a real online 2048 site, one API call and one keypress per move, printing calibrated probabilities for all four directions as it goes.
- Stars
- 2
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- erhanmeydan
- Last verified
- Oct 2026
About this project
TypeSafe'in Jev karar modeli, gerçek bir online 2048 sitesinde oynuyor — hamle başına tek API çağrısıyla. Jev bir dil modeli değil. System One sınıfı bir karar modeli: ona durumu ve seçenekleri verirsin, sana kalibre olasılıklarla birlikte seçeneklerden birini döndürür. Seçenek uzayının dışına çıkamaz, yani bu bağlamda halüsinasyon yapması mümkün değil. Bu proje o fikri en saf hâliyle gösteriyor. 2048'de karar uzayı sabit ve dört tane:
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- All four directions are always offered, with no heuristic hint and no filtering of no-op moves, so the choice is entirely the model's.
- A measured full game ran 127 moves to a score of 884 for about 140,467 input tokens, which the author puts at roughly $0.006.
- It reads gameState and waits for the DOM to catch up instead of sleeping, because 2048 defers rendering to requestAnimationFrame.
- Probabilities arrive rounded to two decimals, so validation allows a 0.02 argmax tolerance; three of forty replies deviated by exactly 0.01.
Watch out
This is not a 2048 solver. Jev decides one move at a time with no search, lookahead or expectimax, and the score sits well below search-based bots that routinely reach 2048 and 4096 tiles — the project is demonstrating constrained choice on a real site, not playing well. It needs a classic DOM-based 2048: play2048.co fails because it moved to a PixiJS/WebGL canvas with no readable board, and screenshots are not an option since Jev takes text input, while the Cirulli GitHub Pages mirror just redirects in the browser. The board and recent moves go to TypeSafe, one request per move. The README is written in Turkish. 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.
A demo. The author is explicit that it is not a solver: Jev makes single-move decisions with no search, lookahead or expectimax, so the measured full game ended at 127 moves with a score of 884 and a highest tile of 64, against search-based bots that routinely reach 2048 and 4096. What it shows is a decision model playing a real site with one call per move.
The agent never generates text, it only chooses. A general-purpose browser agent would need a second language model to fill in forms; here the decision space is fixed at four directions, so one key covers reading the board, pressing a key and choosing. Because the model answers from a supplied option set it cannot leave that space, which is the property the project exists to show.
Only on classic DOM-based ones. Tested alternatives are 2048.org and 2048game.com alongside the 2048.io default. Two known failures are documented: play2048.co, which moved to a PixiJS/WebGL canvas with no readable board, and the Cirulli GitHub Pages URL, which redirects to play2048.co in a browser even though curl shows the classic HTML.
Jev reports probabilities rounded to two decimals, so in near-ties the returned choice can differ from the reported argmax by one rounding unit. Measured over 40 requests, three replies deviated and every one was exactly 0.01, never more. validate_choice therefore uses a 0.02 tolerance, which absorbs the reporting noise while still rejecting a genuine inconsistency such as 0.60 against 0.05.
uv sync, uv run playwright install chromium, cp .env.example .env with your TYPESAFE_API_KEY, then uv run jev2048. Flags cover the site with --url, a move ceiling of 500 with --max-moves, giving up after 6 fruitless moves with --stuck-limit, --delay to slow it down while you watch, --headless, and --json for one JSON record per line. The key stays in .env, which is gitignored. MIT licensed, and the README is written in Turkish.
The board and the recent moves, one request per move, and nothing else. A move that does not change the board is not removed from the option set: it is written into recent_moves as changed: false and fed back as state on the next turn, so the model learns it without the harness making the choice for it. There is no heuristic score hint, no filtering of no-op directions and no second model anywhere in the loop.
board.py is pure data holding all parsing and validation, knowing nothing about the browser, the network or the environment. model.py is the TypeSafe client that builds the request, validates the response and returns a Decision. browser.py wraps Playwright to read the board and press a key, agent.py is the read-ask-press loop, and cli.py is presentation only. Because the logic is pure, uv run pytest runs 52 tests that open neither a browser nor a network connection.
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
1GitHub stars
A FastAPI plus React lab where Jev picks each 2048 move through one choice decision per turn, running many games in parallel with a live best score and a browser editor for the exact state payload Jev sees.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
KyleKreuter/jev2048TypeScript
KyleKreuter/jev2048TypeScriptMore Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython63GitHub stars
Claude Code kit that guards PreToolUse calls with a free pre-filter plus one Jev question for ambiguous cases, and adds sub-agent sizing, rechecks, search, daemon speedups, and monitoring.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
jonathanavis96/jev-kitPython
jonathanavis96/jev-kitPython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPythonFrom 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.