Selected default: CM-5, full-history move prediction. Browser play uses the verified CM-5 weights and matching full-history encoder. The Lichess bot is stopped during a separate Mac runtime migration; CM-5 with iterative Rust 1/3/5/7-ply search is the requested serving target, not a verified active service. Browser search remains three plies.
Historical research model: CM-7, with state-conditioned input and a separate iterative seven-ply live helper. Current evaluation · Current source and reproduction guide. Earlier reports and full-history packages below keep their historical provenance.
Reading the evidence
Three distinct evaluation layers.
A move played by the system is not always a move predicted by the model.
Historical full-history model objective
The original evaluation below concerns a non-reasoning, causal next-move model. Each move is encoded as source square, destination square and promotion tokens. A three-token causal next-token objective implements move prediction from the complete history beginning at the standard start. There are no thought tokens, chain-of-thought traces, or search-supervision targets. Full histories are retained; over-length games are rejected rather than left-truncated. The optional mate helper is external search, not learned reasoning.
Tokenizer source · Data preparation · Training objective · Play harness
1. Model-only diagnostics
Training loss uses raw 128-class logits without a legal mask. Teacher-forced source/destination/promotion diagnostics condition later tokens on gold prefixes. Triple legality is not free-running game strength. Promotion accuracy is dominated by no-promotion tokens.
2. Legality-filtered play
The play policy chooses the highest joint-probability legal move. Raw greedy legality is logged separately. The 591 legal raw-greedy queries out of 651 model queries in the match are a different metric from the 64.64% teacher-forced triple-legality diagnostic.
3. Mate-helper overrides
A bounded all-defenses proof checks for forced mate within three plies. The helper supplied 5 of 656 total system decisions; the other 651 decisions queried the model, of which 591 raw-greedy queries were legal. All three wins contained helper overrides. No model-only ablation establishes that the helper was necessary for those wins. Those moves have an amber star in the viewer, the original PGN comments, and an expandable proof. They must not be credited as model discoveries.
Scope and reproducibility
All 20 games are included. Ten fixed openings were played with both colors. There is no Elo estimate, model-only match ablation, or held-out test result. Model-only weights, filtered raw data and tokenized datasets are available, with checksums and reproduction boundaries.
Adding experiments
New records must validate against the versioned schema. Publish successful, failed and inconclusive results with their protocol. Freeze code and artifact digests before interpreting outcomes; never replace missing historical source with current files.