CHAMPIONSHIP 2026  /  OPEN EXPERIMENT

Behind the moves.

Replay the 2024 championship and inspect the sourced 2026 model experiment.

← Play the three-move challenge
02 / THE DECISION BOARD
Loading verified historical moves… Lichess broadcast ↗
REAL CHAMPIONSHIP ARCHIVE

A world title on the line

Tap or drag a piece to lock your local choice.

♟Gukesh DBLACK
♙Ding LirenWHITE
109 / 116
Choose a move directly on the board, or use the list alongside it.
TRY THE GRANDMASTER TESTHISTORICAL GAME

Which move did Ding choose?

12 December 2024. The final game of the World Championship. Choose the move you think Ding played, then reveal his decision.

◎
Historical decisions, genuine chess.

This is a retrospective test of your prediction, not a claim about model performance.

⤷Sourced to the 2024 Lichess championship broadcast.
THE STANDARD — FORECAST → PUBLISH → OBSERVE → VERIFY.HISTORICAL DATA: LICHESS · DEC 2024
03 /

The models.

Three current frontier systems, a predeclared sample of championship positions, and observed human decisions. Model identities come from provider documentation. Predictions appear here only if a real API attempt has been recorded.

0PUBLIC API ATTEMPTSMODEL VERSIONS VERIFIED 19 NOV 2026
Loading the model roster…
NO MODEL RANKING UNTIL PREDICTION PRIORITY IS VERIFIED.Inspect public evidence ↗
01 / IN THE MOMENT

Before the move.

Championship moves come from a documented public broadcast. Predictions must be published before the observed decision to qualify as prospective evidence.

02 / TWO DIFFERENT QUESTIONS

Choice versus strength.

Did a model anticipate the player’s exact choice? Was that move strong? Exact matching and engine evaluation measure different things.

03 / THE EVIDENCE

Open to inspection.

Our public repository contains the collection protocol, chess rules, tests, and published observations. We disclose gaps before displaying any model accuracy.

04 /

Predictions without hindsight.

Forecasting a grandmaster demands a chronology that survives scrutiny. The live board is an observation tool; the model scoreboard will use a separate, documented evidence standard.

Read the complete protocol ↗
01 / INPUT

Models share a recorded position, legal moves, player identity, prompt, and explicit version.

02 / PRIORITY

A forecast needs independent evidence of publication before the human move. Local timestamps are insufficient.

03 / OBSERVATION

Chess positions are replayed from Lichess PGN. Feed retrieval time is distinct from the physical move time.

04 / DISCLOSURE

No historical test is passed off as a prospective AI result. Engine evaluation and provider provenance are separate verification tasks.