Skip to main content

An AI plays a cyberpunk strategy game. You watch its mind.

Pick a mission, hit Start, and watch a PPO-trained Qwen 2.5 model run a partially-observable strategy environment turn by turn. The dashboard shows the prompt the model sees, the raw text it generates, the parsed action, and how every reward component reacts. Race the trained model against a hand-coded baseline and pure randomness to see whether learning actually transferred.

Pick a mission

    Press Space to play / pause, R to restart, 1-6 to jump missions.

    How the game works

    1. Build alliances. Negotiate with Ghostwire, Iron Vultures, Civic Shield, or Black Orchid.
    2. Gather intel. Scout sectors to learn patrol patterns and find the right extraction point.
    3. Trade resources. Convert credits ↔ intel ↔ influence ↔ energy when stocks run low.
    4. Deploy assets. Pre-position before running an operation; energy is the cost.
    5. Run the right op. Each mission has a specific code. Wrong code = penalty.
    6. Secure extraction. Right sector, right keywords in the message, in 12 turns or less.

    What the dashboard shows

    • Model's mind — the prompt sent to the LLM, its raw text, and the parsed JSON action.
    • Reward forensics — per-turn reward broken into alliances, resources, operation, message, extraction.
    • Mission gates — five locks; once all five are green, extraction succeeds.
    • Turn log — every action with reason and reward delta. Scrollable, replay-friendly.

    Triple Race mode

    The trained policy, the heuristic, and a random agent all play the same mission in lock-step lanes. Cumulative reward, gates unlocked, and final outcome are tracked side by side. Spoilers: the random agent rarely wins.

    Neon Syndicate is a partially observable, long-horizon strategy environment for OpenEnv. The trained model in this demo is a 0.5B parameter Qwen 2.5 fine-tuned with TRL PPO. Trace honestly reflects when the model output cannot be parsed and the heuristic takes over.