Decision lab / Live Snake benchmark
System1 × Laya × Jev
One game. Three decision engines. Every move measured.
Real engines / local arena
NumPy CPU · MLX GPU · Remote API
ConnectingLoading real engine status…00:00.0
Fairness, teaching & what the planner does
- The shared planner labels moves “Blocked,” “Unsafe,” “Safe,” or “Best.” It already knows a safe route to food. This is a test of acting on these hints, not discovering Snake strategy from a raw board. A direct planner is sufficient to play.
- The optional shield picks the highest-scored safe direction for every engine. Each changed move is counted below its board. Turn it off to observe raw choices. System1’s review flags are recorded; this game executes its top choice even when it asks for review.
- System1 uses a small taught TF-IDF skill on CPU. Laya uses its English 421M checkpoint through compiled MLX on GPU. Jev makes a real TypeSafe API request for every decision. There is no saved decision cache or substitute competitor.
- Each engine gets the same seeded starting board and three questions. Paths can diverge, so later food locations can differ. Loading and warm-up are excluded. P50 and P95 include input encoding and result formatting; Jev also includes network time. E2E includes planning and the game update. Concurrent engines share this Mac.
- JSON exports include engine details, teaching results, the protocol and every executed move. Replay uses that recorded trace and makes no model requests.