peterwooden/chess-gpt-board-snapshot-0002

Chess GPT board-snapshot policy 0002

This is a public, browser-native tournament candidate trained on complete chess board snapshots rather than SAN move histories. It is Model 1 in the paired 0002/0003 laptop experiment.

Use in the ChessGPT arena

Load this repository as:

peterwooden/chess-gpt-board-snapshot-0002@ecdf3c42046c01abdd351d1327b77d18388c4306

The unqualified repository name also works, but the pinned reference above is the exact package revision independently re-downloaded and verified for this release.

The tournament runner reads browser/manifest.json. The manifest describes a self-contained JavaScript entrypoint, an ONNX policy, and the stable promotion move vocabulary. ONNX Runtime Web 1.27.0 is supplied by the runner and is not part of the submitted byte count.

The network produces 4,272 move logits. The adapter reconstructs the current position from SAN history, maps the runner-supplied legal SAN moves to those logits, masks everything else, and returns deterministic argmax SAN. A softmax probability vector is therefore unnecessary for tournament move selection.

Architecture and training

• —Six-layer Transformer encoder, width 336, eight attention heads.
• —One post-encoder feed-forward expert and 10,586,256 trainable parameters.
• —January 2026 frozen Lichess standard-rated data: first 100,000 games, 6,619,804 available positions.
• —Trained for 26,255 AdamW updates / 3,356,140 positions on Apple M4 MPS, stopping at the three-hour training limit.
• —Seed 20260729, batch size 128, float32.
• —Ratified dense-operation accounting: 11,029,491,768,153,600 lineage FLOPs, 1.10% of the 10^18 tournament limit.
• —Training cost: $0; the run used the local laptop.

April 2026 was held out from optimization. On its first 2,000 games / 131,491 positions, validation loss was 2.83490, raw next-move top-1 accuracy was 26.858%, legal-masked next-move top-1 accuracy was 29.387%, and the legal-move rate was 100%. Human next-move agreement is not the same as playing strength. In the separate tournament-shaped local match—50 frozen April openings, colors reversed, 100 games—the two published candidates tied exactly: six wins each, 88 draws, and 50/100 points each.

Integrity and limits

The canonical browser package is 42,585,883 bytes, below the 100,000,000-byte cap. browser/manifest.json pins every submitted artifact by byte count and SHA-256. The checkpoint SHA-256 is a06aeb7626559cc5b653d3d40b4b9122b1972ff3fece47fdc465d10145674190.

The model has no search and no engine evaluation. It imitates moves in a limited public game slice and can make strategically weak choices even though the adapter always returns a legal move. The package is deterministic and contains no pretrained weights, engine labels, outside games, or synthetic training data.

Full configuration and measured run evidence are included under training/.