Recreating Historical Chess Style with Policy-Only Fine-Tuning and Quantitative Validation

Chess engines optimize for move quality and, in doing so, converge on a single contemporary style. Chess style, however, is historically situated: opening repertoires, draw culture, and resignation habits all changed measurably between 1840 and 2019. We present Time-Machine Chess, a family of five era-specific chess models produced by one epoch of fine-tuning Maia-2 — a human-move-prediction network — on balanced corpora of dated over-the-board games (1840–1885, 1900–1939, 1950–1985, 1990–1999, 2010–2019). Models are served policy-only (no search), with a temperature schedule that preserves period opening diversity, and a rule-based “social layer” that derives era-appropriate draw agreements and resignations from the model’s own win-probability head. We validate the era claim three ways: (1) self-play reproduces historical opening, draw-rate, and game-length gradients within ~3 percentage points and ~5 plies respectively; (2) used as a generative classifier on held-out historical games, the models identify a game’s era at 42% (single game, 20% chance) and 88% (20-game batches), with a confusion structure that itself tracks chess history; and (3) strength measured against a calibrated engine ladder places all five models within a 181-point Elo band (~1580–1760), supporting the design goal that eras differ in style rather than strength. Total training cost is under one laptop-hour; serving requires no GPU.