Small moves. Real decisions.
Watch Jev play Snake. Explore every run.
Every replay comes from a saved game.
Last decision
Recorded model decisions.Advance the replay to inspect a decision.
Benchmark results
Mean food collected · eight-food target| Difficulty | Direct (Jev) | Guarded | Baseline |
|---|
What is the model actually doing?
Jev receives structured board coordinates and selects left, straight, or right. It does not see this canvas. Direct mode offers all three non-reversing moves. Guarded mode removes immediately fatal moves; a sole safe move is executed by code and labeled. The baseline uses deterministic pathfinding and makes no model calls.
Games advance only when decisions return. Replay speed is a viewing control, not measured real-time performance. A game stops at eight foods, 200 moves, a collision, or 50 moves without food. Each difficulty/controller pair has eight seeded games. A best replay is a selected example, not typical performance.
To run live locally: python -m snake_demo.server from the repository root, with TYPESAFE_API_KEY in your environment or local .env. The key stays on the server. The live button starts paid model calls when you choose a Jev controller.