Someone made jev play Atari games

A developer released an open-source project that evaluates TypeSafe's Jev model, version jev-1.13.0, as a policy in Gymnasium and Atari environments, converting game observations into structured JSON so Jev selects a legal action through a single Choice question. The project supports CartPole, Pong, Breakout, and Ms. Pac-Man, with Montezuma's Revenge marked experimental and development paused, and records rewards and decisions without training or updating model weights, using a seeded random policy as a local baseline. Atari runs use the Arcade Learning Environment with Python 3.14, a committed uv.lock, and a trial wrapper defaulting to 64 decisions per game, 2 emulator frames per decision, and a 0.25 sticky-action probability.

Evaluate TypeSafe Jev https://docs.typesafe.ai/introduction as a policy in Gymnasium and Atari environments. Game-specific adapters convert observations into structured JSON; Jev selects a legal action through a single Choice question. This project evaluates a fixed model. It records rewards and decisions without training or updating model weights. A seeded random policy provides a local baseline. Three games, one shared action-selection question. Joysticks and probability bars show Jev's recorded decisions at 2× game speed. See recording details https://github.com/taodav/jev deep rl/blob/main/media/README.md for the source runs, rendering command, and smaller gameplay-only GIF. Environment → structured state → Jev Choice → legal action → environment ↓ ↓ RAM, frames, requests, responses, rewards → NPY / JSONL / GIF | Environment | Observation adapter | Status | |---|---|---| | CartPole | Named position, velocity, angle, and angular velocity | Supported | | Pong | RAM or RGB object detection | Supported | | Breakout | RAM: ball, paddle, lives, and brick map | Supported | | Ms. Pac-Man | RAM: maze, actors, food, and local navigation | Supported | | Montezuma's Revenge | RAM and reference room geometry | Experimental; development paused | RAM and known room geometry provide privileged information compared with pixel-only Atari benchmarks. Adapter limitations are documented in observation adapters https://github.com/taodav/jev deep rl/blob/main/docs/observations.md and the Montezuma prototype https://github.com/taodav/jev deep rl/blob/main/docs/montezuma.md . Install uv https://docs.astral.sh/uv/getting-started/installation/ , then from the repository root: uv sync --locked --extra atari --extra recording Local run: no credentials or API requests. uv run --locked --no-env-file --extra atari --extra recording main.py \ --env ALE/Breakout-v5 --policy random --max-steps 500 \ --log runs/breakout-random.jsonl --save-npy runs/breakout-random.npy uv run --locked --no-env-file --extra recording npy to gif.py \ runs/breakout-random.npy --scale 2 The project uses Python 3.14 and a committed uv.lock . Atari runs through Arcade Learning Environment https://ale.farama.org/getting-started/ . No ROM files or credentials are included in this repository. Raw experiment recordings stay local; media/ contains selected demonstration GIFs and a showcase preview. CartPole can run without the Atari extra: uv run --locked --no-env-file main.py --env CartPole-v1 --policy random Create .env from the blank template only if you do not already have one : cp -n .env.example .env Set TYPESAFE API KEY in your local .env or process environment. The .env file is ignored by Git. uv --env-file .env loads it at runtime; the Python scripts do not open it. CLI entry points suppress SDK/HTTP debug output and raw exception payloads. Response capture redacts runtime credentials and excludes headers/cookies. Start with a bounded live trial: uv run --locked --env-file .env --extra atari --extra recording run jev trial.py \ --game breakout --max-steps 64 --output-dir runs/breakout-trial Game choices: pong , breakout , mspacman , montezuma , or both Pong and Breakout . The wrapper uses RAM observations, one episode, no HTTP retries, and exports a GIF automatically. Existing outputs are never overwritten; use a new output directory for each trial. Runs stop at game over, truncation, or the requested decision limit. | Setting | Default | |---|---| | Model | jev-1.13.0 | | Emulator frames per Atari decision | 2 | | Sticky-action probability | 0.25 | | Seed / episodes | 7 / 1 | | General CLI decision limit | 500 | | Trial-wrapper decision limit | 64 per game | | HTTP retries | General CLI: 2; trial wrapper: 0 | The general CLI exposes frame skipping, sticky actions, model, seed, episode count, and replay controls. Run main.py --help for details. Pong uses RGB by default in the general CLI; add --pong-state ram for RAM input. The simulator pauses during each API request. Playback uses emulator time, so network latency affects runtime but not the recorded game speed. Each decision normally uses one API request; general-CLI retries can add requests. Check current TypeSafe pricing and limits https://docs.typesafe.ai/models before large runs, and use logged token usage to measure cost. - .npy : RGB frames, raw observations, exact structured states, actions, rewards, episode summaries, and redacted API responses. - .jsonl : transition records; .responses.jsonl : exact redacted API requests and response bodies, captured before validation. - .gif : game-time playback exported from the recording. All local outputs belong under the ignored runs/ directory. NPY files contain pickled dictionaries; only load your own trusted recordings. Details and examples: recordings, replay, GIF conversion, and offline debugging https://github.com/taodav/jev deep rl/blob/main/docs/recordings.md . Join GIFs side by side, stopping when the shortest finishes: uv run --locked --no-env-file --extra recording join gifs.py \ runs/pong.gif runs/breakout.gif runs/mspacman.gif --output runs/combined.gif uv sync --locked --extra atari --extra recording uv run --locked --no-env-file ruff check . uv run --locked --no-env-file ruff format --check . uv run --locked --no-env-file --extra atari --extra recording python -m pytest -q Tests use local emulators and mock HTTP transports. No live TypeSafe requests or API key are required. See CONTRIBUTING.md https://github.com/taodav/jev deep rl/blob/main/CONTRIBUTING.md for adapter development and AGENTS.md https://github.com/taodav/jev deep rl/blob/main/AGENTS.md for agent instructions. The core modules are adapters.py and the game decoders state , policies.py action selection , runner.py episode loop , recording.py NPY , and responses.py redacted capture and replay , under jev rl/ . Built with Gymnasium https://gymnasium.farama.org/ , ALE https://ale.farama.org/ , and TypeSafe https://docs.typesafe.ai/ . RAM decoding references and Montezuma room geometry draw on OCAtari https://github.com/k4ntz/OC Atari . Its license and the installed TypeSafe skill's license are recorded in third-party notices https://github.com/taodav/jev deep rl/blob/main/THIRD PARTY NOTICES.md .