Start with tactics

1. Open World. Try a move as White in a generated practice position.
2. Choose Rule teacher to see the reference choice.
3. Train a neural network to rank legal moves.
4. Review held-out positions in Results, then activate a passing model.

Rank the available choices

The rules engine supplies legal moves, captured and moving piece values, check, mate, strongest immediate reply capture, promotion, castling, centralisation, promoted material gain and immediate stalemate. The NN learns the teacher’s authored score from these ten features.

The NN scores every legal move on each turn, including Queen, Rook, Bishop and Knight promotions. It learns from authored scores and fixed promotion exercises. These exercises are reported separately from unseen-game checks.

Run uses fixed weights. Retrain between games using opening, middle-game and endgame examples, then review checks before activation. Changing weights after every move would make comparisons unstable; one move rarely proves whether its decision was good.

It learns a ranking, not the rules or camera perception. Computing these features already uses chess knowledge. There is no NN search or hidden teacher override after activation.

Validation selects the checkpoint. Tactics checks compare separate generated positions. Complete-game checks compare terminal games against the same authored Black opponent. Draws meet the non-loss goal; these checks do not establish a chess rating.

Train the selected World goal

Archived eight-input tactics checkpoint: Download the older checkpoint. Its historical 24/24 tactics result predates promotion-aware inputs. Open project preserves its weights, but retrain with Train selected goal before using the updated NN. Saved-weight checks are available only for the current model format.

Tactics: 72 fitting positions, 16 validation positions and 24 final positions. Complete games: 3 fitting opening seeds, 2 validation seeds and 4 final seeds, all played to terminal outcomes. One scorer per goal, up to 60 epochs and a 60-second worker limit. Two attempts per training round, including checks and cancellations. When used, choose Start another training round above. It retains active weights and the recorded failure.

Ready · 0 / 2 attempts used.

Tactics activation: at least 18/24 near-best choices, at least four more than the initial NN, and no incumbent regression. Complete-game activation: at least 3/4 non-losses, no regression versus initial or incumbent, and at least 80% near-best validation decisions. A failed run can be retrained with exact failed observations and fresh final seeds. No candidate activates automatically.