Time-Machine Chess

Method. Each bot played 150 self-play games at its serving temperature. The same analyzer computed identical, move-sequence-based metrics on the bot games and on random samples of the historical era corpora (Lumbra's Gigabase, OTB games) — no ECO tags, no cherry-picking. Resignation and draw agreement — the social layer policy networks lack — are both adjudicated from the model's own win-probability head, with era-specific manners tuned against the historical record: the Soviet school resigns promptly and agrees draws readily; the Classical era is moderate in both; Romantics play on toward the mate and almost never agree to a draw.

Honest residuals. One systematic gap remains. First moves lean 1.e4 beyond the historical rate: the bots are fine-tuned from Maia-2, which was trained on modern online chess — one epoch of era data moves its opening prior a long way, but not all the way. The two larger gaps documented at launch — draw rates 8–10 points below history, games ~25 plies too long — are now modeled rather than missing: the bots resign and agree draws with era-specific manners driven by their own win-probability head, tuned against the record. Draw rates land within ~3 points and average game length within ~5 plies of history in all five eras. The draw-culture curve is itself a finding: it rises for 150 years (12% → 25% → 28.75% → 31.75%) and then plateaus in the engine age — Sofia rules and fighting-chess culture flatten it exactly as games stretch to their historical maximum length — and the bots track every step. Tuning those social constants revealed they interact: prompt resignation eats would-be draws, so the pairs were tuned jointly. The remaining first-move lean doesn't affect the in-period style signal above.