system-one-models-game-test
Browser arcade games (2048, runner, chess, Flappy Bird, Sudoku) where small decision models choose each move from text options.
Overview
This project hosts five browser games, 2048, a Chrome-dino-style runner, chess, Flappy Bird and Sudoku, in which each move is chosen by a small System One decision model rather than a search algorithm or an LLM. Each turn the game state is described in text, the model picks one of the candidate moves, and the browser shows the result. Games can be watched one at a time, in an arena with many players at once, or head to head in chess.
Seven decision models can be switched from the UI: Laya, Jev-BERTa, the hosted Drex API, Kev-0.8B, GLiNER2.5-Decide, Bev-Decider-0.4B and DecisionMaster, plus random, greedy and oracle baselines. The README stresses that none of the models was trained on these games, so the project is a zero-shot comparison of how well generic decision models act as move pickers. Benchmark scripts play each game repeatedly per player with fixed seeds so runs are comparable.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Given the described game state, which move should be played next? | one of the legal moves listed for the turn |
How it works
- The game engine describes the current position in text.
- The legal moves are passed as candidate options to the selected model.
- The model returns probabilities and the top move is applied in the browser.
- Bench scripts repeat full games per player with fixed seeds to compare models and baselines.
Strengths
- Side-by-side comparison of seven decision models on identical game states.
- Models load once and are shared across all five games.
- Seeded benchmarks make player comparisons repeatable.
Caveats
- The models are not trained on the games, so play quality reflects zero-shot behaviour only.
- DecisionMaster weights are gated and Drex needs a hosted API key.
- No licence is stated in the README.
Getting started
pip install laya chess gliner2 peft bev-decider git+https://github.com/leobitz/jev-berta.git && python server.py, then open http://localhost:8048.At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- local web server + browser
- Has code
- yes
- Stars
- 0
Use cases and integrations
How curated lists describe it
Games where the moves are chosen by a small System 1 decision model instead of a search algorithm or an LLM.
wh000wh000/awesome-jev-live
Frequently asked questions
What does the model decide in these games?
On every turn it picks one move from the listed legal options, such as a 2048 swipe direction or a chess move.
Which models can play?
Laya, Jev-BERTa, Drex, Kev-0.8B, GLiNER2.5-Decide, Bev-Decider-0.4B and DecisionMaster, plus random, greedy and oracle baselines.
Do I need an API key?
Only for the hosted Drex model; the other models run locally after their checkpoints download.
Related projects
- NanoJev★ 2,543
0.6B open Jev replica returning full option distributions with no decoding; one checkpoint plays ViZDoom, Maze and Snake.
ModelresearchPython Fifteen live arcade games where Jev and LLMs play each other, led by WikiRace in which Jev picks only real links.
ProjectbetaTypeScriptOne-page web app where Jev plays chess against OpenRouter LLMs, Stockfish or you, showing the moves it weighed and their probabilities.
AppbetaJavaScriptPac-Man where Jev or other System One models choose each move at junctions, with live odds, fallbacks and a leaderboard.
ProjectbetaTypeScriptBrowser chess lab where Jev, LLMs and people play, each Jev turn being one Choice over the legal moves, with Stockfish analysis.
ProjectbetaPythonNiceGUI desktop chess app for evaluating Jev's move choices against Stockfish, random players or itself, with PGN export.
ProjectprototypePython