Compare the models

Same Position. Two Predictive Approaches.

A statistical human-move model and the state-of-the-art chess-specific human-move transformer, both asked the same question about the same real decision: what move is this player most likely to choose? Click any legal move to see what each approach assigns to it, or switch between the original human move and the controlled engine substitute.

Prepared research studyChecking for the live GPU service… the prepared study below is complete without it.

36 decisions of one prepared player-game (black, rated 1956). substituted in a prepared study featured position

Move 24, Black to move · player 1956 vs 1945 · 31 legal moves

Controlled engine substitute: 24… Qe6

This is a decision-level counterfactual evaluated in the original position: the board never changes; only the move being evaluated does. Later positions of the game are not replayed.

Controlled engine substitute

24… Qe6

Substituted in: Obvious controlled assistance, Selective controlled assistance. The player actually chose h6; the substitute Qe6 is Stockfish's first line, inserted by the controlled experiment.

Regan-inspiredStatistical model (public reconstruction)

Predicted ranknot defined

Move probabilitynot defined

Engine rank of this move#1

P(engine's first choice)17.0%

P(one of engine's top 3)42.4%

Engine's first choiceQe6

Statistical projection from Elo and engine-derived position difficulty. It projects how often a player of this rating matches the engine; it does not give every legal move a probability.

Maia-3 79MChess-specific human-move transformer

Predicted rank#13 of 31

Move probability0.2%

Top predicted moveg6 (45.9%)

A Chessformer-based transformer trained on human games and conditioned on both players' ratings. Not a language model.

Showing Qe6 (the controlled engine substitute). Click any legal move below to see what each model assigns to it.

Every legal move

Sort by

Move	Engine	Maia-3
	#3 +0.07	45.9%
human	—	38.6%
	—	3.7%
	—	3.6%
	—	3.6%
	—	1.0%
	#4 +0.02	0.8%
	#5 0.00	0.6%
	—	0.5%
	—	0.4%
	—	0.3%
	—	0.2%
substitute	#1 +0.09	0.2%
	—	0.1%
	#2 +0.08	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%
	—	<0.1%

Technical details for this position

LLM experiment (next step)

Qwen3-8B-Base (Qwen/Qwen3-8B-Base @ 49e3418fbbbc) will be scored on this exact position with the frozen prompt below (SHA-256 7f4373cf34140539…): every legal move's token likelihood, normalized over the legal moves. It has not been run yet, so no LLM values are shown.

Show the exact prompt for this position

Chess move prediction. The position below is from a rated online game (Lichess, rapid or classical time control).

Player Elo: 1956
Opponent Elo: 1945
Side to move: Black

Position FEN:
4r1k1/p3qppp/1p6/2p5/2P5/1P5P/P1Q2PP1/3R2K1 b - - 0 24

Engine analysis: Stockfish 19, fixed depth 12, MultiPV 5. engine_eval is the evaluation in pawns from the point of view of the side to move; n/a means the move is not among the engine's top lines.

Legal candidate moves:
a7a5 | engine_eval=n/a
a7a6 | engine_eval=n/a
b6b5 | engine_eval=n/a
e7b7 | engine_eval=n/a
e7c7 | engine_eval=+0.00
e7d6 | engine_eval=n/a
e7d7 | engine_eval=n/a
e7d8 | engine_eval=n/a
e7e1 | engine_eval=n/a
e7e2 | engine_eval=n/a
e7e3 | engine_eval=n/a
e7e4 | engine_eval=n/a
e7e5 | engine_eval=n/a
e7e6 | engine_eval=+0.09
e7f6 | engine_eval=+0.08
e7f8 | engine_eval=n/a
e7g5 | engine_eval=n/a
e7h4 | engine_eval=n/a
e8a8 | engine_eval=n/a
e8b8 | engine_eval=n/a
e8c8 | engine_eval=n/a
e8d8 | engine_eval=n/a
e8f8 | engine_eval=n/a
f7f5 | engine_eval=n/a
f7f6 | engine_eval=n/a
g7g5 | engine_eval=n/a
g7g6 | engine_eval=+0.07
g8f8 | engine_eval=n/a
g8h8 | engine_eval=n/a
h7h5 | engine_eval=+0.02
h7h6 | engine_eval=n/a

Predict the move most likely to be chosen by a human player of approximately this strength.

Move:

Maia-3

Primary checkpoint: UofTCSSLab/Maia3-79M @ a107d6ceb7b2
Existing-pipeline checkpoint: UofTCSSLab/Maia3-5M @ b6559de2398d (used inside hybrid systems C/D)
Input: current position + up to 7 previous positions; side-to-move and opponent Elo
Output: softmax of the policy logits over the legal moves

Regan-inspired reconstruction

Position difficulty: 0.8184
Projected P(T1 match): 0.1702
Projected P(top-3 match): 0.4235
Projected P(EV match): 0.1957
Expected capped loss: 37.9 cp
Stockfish lines: Qe6 +0.09 · Qf6 +0.08 · g6 +0.07 · h5 +0.02 · Qc7 0.00

Elo band × difficulty-decile projection fitted on training decisions only (projection.json). A reconstruction based on Professor Regan's publicly documented framework, not his private production analyzer.

The same player-game in the three prepared studies

Game-level signals over all 36 decisions. Higher z(top-1) = more engine matches than projected; higher mean surprisal = moves less expected by the human-move model. These are assistance-sensitive scores, not probabilities of cheating.

Study	Substituted	Regan-inspired z(top-1)	Maia-3 mean surprisal	Hybrid D estimate
Human baseline	none (original)	+0.20	0.977	5.7%
Obvious controlled assistance	40% target · 38.9% realized	+5.00	1.591	40.3%
Selective controlled assistance	20% target · 19.4% realized	+2.60	1.244	16.7%

Hybrid D estimate = the existing “synthetic-equivalent assistance estimate” of System D (unchanged). Substitutions are controlled decision-level substitutions evaluated in the original positions.