Time-Machine Chess

What is Time-Machine Chess?

A free, open-source website where you play chess against the theory of a past era: the gambit-happy Romantics of 1840–1885, the Classical masters of 1900–1939, the Soviet school of 1950–1985, the database-armed 1990s, and the engine-hardened 2010s. Each era is a neural network fine-tuned on thousands of real games from its period — and each is validated against the historical record it learned from, from opening fashions to draw culture.

Why is each era's Elo only about 1600–1750 when its great players were far stronger?

Three honest reasons. First, the bots imitate an era's recorded practice, not its champions: the training games mix World Championship matches with ordinary master games, and the model learns the blend. Second, the bots play without calculation — a policy network samples what a player of the era would likely do, with no search to catch tactical errors, and the sampling that reproduces period variety sometimes picks the second or third most human move. Anderssen calculated; his imitation doesn't. Third, by design: every era is conditioned at the same fixed skill level so that the eras differ in style rather than strength — and our measurements (120 games per era against a calibrated engine ladder) confirm all five land within one rating class, roughly 1580–1760.

Think of each bot as an era's habits at strong-club strength, not a resurrected champion. Comparing across centuries is slippery anyway — Elo ratings only began in 1970. Full measurements on the validation page.

How can a neural network play like 1850?

Each era bot starts from Maia-2 (University of Toronto), a model trained to predict what humans actually play rather than what's best. We fine-tune it on 10,000–12,000 dated over-the-board games from one era, then serve it with no search at all: the bot simply plays what a player of its period most plausibly would — sacrifices, fashions, blind spots and all.

Can I find out which era I play like?

Yes — the era classifier reads your games (paste a PGN, or give a lichess or chess.com username) and scores every move you played under all five era models. You get an era mix ("you are 28% Romantic"), a verdict in era voice, and the single most characteristic move you ever played for each era. On held-out historical games it identifies the right era at 42% from a single game (chance is 20%) and 88% from twenty games.

How do you know the bots really play in period?

We measure it. Each bot played 150 self-play games, and the same metrics were computed on those games and on the historical era archives: the King's Gambit appears at 18% for the Romantic bot (14% in the 1850s record) and 0% in the engine-age eras (matching its real extinction); draw rates land within ~3 percentage points of history in every era; game lengths within ~5 moves. The full tables, confusion matrices and honest residuals are on the validation page.

Why did the Romantic bot decline my gambit?

Probably because the Romantics would have declined it too. Honour culture was real — against the actual King's Gambit (1.e4 e5 2.f4), players of 1840–1885 accepted 81% of the time, and so does the bot's spirit. But not every gambit deserved acceptance: against the Vienna Gambit, the same masters declined 92% of the time, because taking the pawn concedes the centre. The bot read the same books. Honour demanded you accept a gambit — not a bad one.

How do the bots decide when to resign or agree a draw?

Resignations and draw agreements aren't chess moves, so a move-prediction network never learns them. We model them as era-specific manners driven by the model's own evaluation of the position, tuned against the historical record: the Soviet school resigns promptly and agrees draws readily; the Romantics play on toward the mate and almost never shake hands. After tuning, each era's draw rate and game length match its history — the bot really can resign like it's 1974.

Why is there no era before 1840?

The past kept poor records. Only around 670 over-the-board game scores survive from before 1840 — systematic recording only began with the first chess magazines in the late 1830s — and that is far too few to train on. The Romantic era's 10,700 surviving games are effectively the entire recorded dawn of chess history, and the Romantic bot is trained on nearly all of them.

Is Time-Machine Chess free? Can I see the code?

Free, no account, and nothing you play or paste is stored. The code — including every training and validation script — is MIT-licensed on GitHub. It's built on Maia-2 (CSSLab, University of Toronto); the historical games come from Lumbra's Gigabase and are not redistributed.