Chapter 01 / The model
How can one network suggest moves and estimate outcomes?
Policy proposes. WDL evaluates. Neither is a proof.
Two questions, one network
At a position, the agent needs to ask two different questions: which moves are plausible? and how good might this position be? CM-14 supplies both signals from the recorded game history.
Policy
A ranking over candidate move tokens. It suggests what to try.
Value
White win, draw and White loss estimates. They help compare positions.
Rules
An external chess library defines what is legal. The model does not get the final say.
Learning from human games
Imitation learning asks the model to predict moves from recorded human play, including games by relatively strong players. It learns patterns in those examples, rather than optimizing directly for an enjoyable opponent.
Andrei finds the current agent challenging but approachable: it can show a useful trick and still leave chances to win. Human-game training may contribute to that character, but it is a hypothesis, not a controlled result. Model errors, search, helpers and time controls also shape the experience. LittlePly is a playing partner, not a validated teaching system.
A move is one token
CM-14 uses one token per complete UCI move. e2e4 names a source and destination; a promotion such as e7e8q also names the new piece. A beginning token precedes the history.
An actual legal sequence
BOS → e2e4 → e7e5 → g1f3
After these moves, Black is to play. Earlier moves are input; the next move is the prediction target.
CM-14 reads full history from the standard starting position. Its 512-token context is a limit, not permission to silently drop older moves.
Policy probability is not winning probability
A high policy score means a move looks likely under the learned distribution. It does not mean that move wins with the same probability. WDL is a separate three-way normalization.
For a fixed-White value, one useful score is P(White win) + 0.5 × P(draw). White prefers larger values; Black prefers smaller ones. Mixing perspectives breaks search even if every prediction looks reasonable.
Why “fixed White” matters
If the model predicts White win / draw / White loss as 0.6 / 0.3 / 0.1, its White score is 0.75. Black’s corresponding score is 0.25. These numbers are a teaching example, not recorded telemetry or a calibration claim.
Legal play does not prove legal prediction
The harness filters to legal moves. That makes a played move legal even if the model’s raw top choice was impossible. We distinguish raw top-prediction legality from legality of submitted moves.
WDL estimates are fallible too. A confidently wrong evaluator can steer deeper search into the wrong branch.