The project includes an experimental 4-layer policy-value MLP (256 → 96 → 48 → 16, with value and policy heads) for offline training. Live browser play currently uses deterministic search and a handcrafted board evaluator.

Each 4×4 grid cell is encoded as a 16-dimensional one-hot vector based on log₂(tile value), producing 256 input features that capture the full board state.

1. Teacher Search: The training workflow defaults to deterministic Expectimax at depth 3. Chance nodes model tile spawns (2 at 90%, 4 at 10%), and the depth can be configured when a run starts.
2. Policy-Value Weights: The published checkpoint contains a 4-logit policy head and a normalized value head. The checked-in workflow references separate trainer scripts, so optimizer, loss, clipping, and schedule claims are intentionally not shown here.
3. Value Scaling: The published weight metadata records the target normalization as tanh(heuristic / 800).
4. Live Move Selection: The browser currently selects moves with Expectimax, alpha-beta pruning, caching, and a handcrafted evaluator (space, structure, smoothness, merges, and corner position). The experimental MLP is loaded for training metadata but does not replace that evaluator in live play.
5. Fixed-Seed Checks: The workflow schedules a 50-game validation interval and a final 100-game fixed-seed evaluation before recording a candidate result.

During auto-play, the JavaScript engine uses iterative-deepening Expectimax (depth 1–6) with a dynamic 150–800 ms decision budget. Leaf states are scored by the handcrafted evaluator, not the experimental neural-network output.