Fix the task before optimizing the model

The game’s real win condition

Snake starts with three occupied cells on an 18 × 18 board. Open garden has 324 playable cells, The long path has 316 and Twin hedges has 312. Completing those boards means collecting 321, 313 and 309 apples respectively.

The earlier small-apple target rewarded a useful beginner task, but it did not answer the full-board goal. We corrected the game and experiment rules, including the case where no food can be placed because every playable cell is occupied.

Long planning horizons

Finding the next apple is easy compared with avoiding a trap thousands of moves later. A Hamiltonian cycle visits every playable cell in a safe order. Expert routes and safe shortcuts give training a much better starting point than random collisions.

Verify the environment too

The recordings were checked against the website’s TypeScript rules, comparing a state hash after every action. Across the three public courses that covers 43,508 states. A win requires the expected body length, no remaining food and the game’s actual won state.

These are finite validation results for three fixed course geometries. They do not prove safe play on arbitrary obstacles, resized boards or every possible seed.