Every human carries their own internal logic to the board. This engine reads a corpus of games, turns each move into a
feature vector (python-chess), and (1) distills each player into a style fingerprint, and (2) trains a tiny
torch ranker to answer: which of the legal moves would THIS player choose? — evaluated on games it never saw.
MVP kernel of a larger idea: gameplay is a verifiable signature of how a person thinks. This is a tiny corpus (a dozen games) and a tiny net — the point is the pipeline and that the learned model beats greedy/heuristic play at reproducing a specific human's choices. Scale the corpus (beginner→GM) and the same machinery grows a font of move-logic.
← HRL gallery · the same relaxation-labeling lineage, now reading minds through moves.