Full game report

Inspect a game move by move

Open one of the three prepared studies (precomputed; no server needed), or run live analysis on your own PGN. For every move you see how strong it is, its model-estimated human plausibility, how hard the position was, and how the whole game compares with expectation.

Prepared studies are precomputed reports bundled with this site; they work without the live server.

One original game, three prepared reports. Test-split player-game, 1800-2199 band, >= 28 eligible decisions, baseline z(top-1) closest to the group median. Not selected on model output. The two substitution reports analyze counterfactual engine moves at some of the analyzed player's decisions, each evaluated in the original position; the game record and all later positions stay those of the original game. They are not replayed games.

Download AIF fileExperimental AIF-compatible export. Compatibility with Professor Regan's current IR analyzer has not been tested.

AIF stores the chess game together with detailed engine analysis for each position.

About the AIF export

Format: Analysis Interchange Format (AIF), the input format of Regan's IR analyzer
Contents: PGN-compatible game · engine ID/settings · per move: FEN, legal moves, engine evaluations, depth, MultiPV lines, principal variation
Our data: MultiPV 5 at one fixed depth (12); shallower depth columns are “n.a.”; node counts not recorded (“n.a.”)
Specification: K. W. Regan & T. Biswas, AIF v0.5, described in “A Computer Chess Analysis Interchange Format”, Gödel’s Lost Letter and P=NP, 20 Jan 2015.
Compatibility: Not tested with Regan's current analyzer. Internal storage is JSON/Parquet, not AIF.

Prepared report · precomputedOriginal game moves · no injected assistance600+0Result 1/2-1/2

White (1945)vsBlack (1956)

ply 17 / 119 · ←/→ move · Shift for decisions

PlayedBxc4
Stockfish #1Nc6
Maia-3 most probable moveBe7

Original game record

10.

11.

12.

13.

14.

15.

16.

17.

18.

19.

20.

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

46.

47.

48.

49.

50.

51.

52.

53.

54.

55.

56.

57.

58.

59.

60.

Position inspector · Black (1956)

9… Bxc4

Engine quality

Mistake

Human plausibility

Medium

Position difficulty

Low

Pattern over time

Too early

Stockfish rates this move as a mistake and Maia-3 assigns it a moderate predicted probability (25.7%) for a 1950-rated player in a straightforward position.

Technical details

Stockfish

Rank of move: 6+ (outside the 5 analyzed lines)
Centipawn loss: 105.0 cp
Evaluation before move: +1.14
Best − second-best gap: 45 cp
Near-optimal alternatives: 1 within 25 cp
Search: depth 12, MultiPV 5
Display rule: Excellent ≤ 10 cp or rank #1; Good ≤ 50; Inaccurate ≤ 100; Mistake ≤ 300

Stockfish lines

eval in pawns, side to move

1	Nc6	+1.14	Re1 Qd7 a3 O-O-O
2	h6	+0.69	Re1 Nc6 Bf4 Qd7
3	Be7	+0.65	Re1 Nc6 Bf4 Bxc4
4	a6	+0.65	Bf4 Nc6 Re1 Qd7
5	Qf6	+0.58	Qh5 Nc6 Na5 Nxa5

Maia-3 human model

Predicted probability: 0.2571
Maia rank: #2 of 38 legal moves
Surprisal: −ln(0.2571) = 1.36 nats
Expected surprisal (entropy): 1.86 nats
Elo conditioning: player 1956 · opponent 1945
Display rule: High ≥ 30%; Low < 10%

Maia-3 most probable moves

predicted probability at this rating

Be732.0%
Bxc425.7%played
Nc617.9%
Bc57.0%
Bd65.1%

Statistical view (Regan-inspired approximation)

Expected: P(top-1 match): 47.3%
Observed: no top-1 match (0)
Deviation for this move: −0.47
Position difficulty: 0.554 (hard ≥ 0.732)
z-score: computed per game over all decisions, not per move

Sequence context

Rolling engine agreement (8): 0%
Rolling move loss (8): 105.0 cp
Rolling surprisal (8): 1.36 nats
Window expected agreement: 47% over 1 decisions
“Unusual” rule: observed − expected ≥ 0.25 over ≥ 4 decisions

Decision-level model (separate per-move model, SHAP)

Resemblance of this single move to a synthetic substitution: 0.000. Contributions in log-odds. This is a separate model trained on individual decisions; it is not a decomposition of the game-level score.

← toward human patterntoward assisted pattern →

Stockfish rank of the movevalue 6−4.26
centipawn loss of the movevalue 105−4.18
best-vs-second-best gapvalue 45−0.08
Maia-3 predicted probability of the engine-best movevalue 0.179−0.08
position difficultyvalue 0.554−0.07
human surprisalvalue 1.358−0.06
Maia-3 predicted probabilityvalue 0.257+0.01
player ratingvalue 1956+0.01

Synchronized with the board

The game at a glance

Three views of every scored move. Highlighted columns are moments where an engine-best move also had a low Maia-3 predicted probability for this rating, in a hard position. Selecting a column moves the board and inspector to that decision; the strip below confirms the selection.

Engine qualitytaller = stronger move (Stockfish)

Human plausibilitytaller = higher Maia-3 predicted probability at this rating (square-root scale)

Position difficultytaller = harder to find the best move

move 9each column = one scored move · click to inspectmove 59

Highlighted: engine-best + low Maia-3 probability + hard position

1 highlighted moment

Selected: 9… Bxc4 · engine mistake · Maia-3 probability medium (25.7%) · difficulty low

Black (1956) · 36 scored moves

Overall assessment

Why it matters: one number summarizing how much the whole game resembles the controlled substitution examples — a resemblance measure, not a verdict.

Similarity to controlled substitution

Typical of original validation games

higher than 47% of original validation player-games

Estimated assistance intensity

5.7%

of scored decisions (synthetic-equivalent)

Ground truth

No injected assistance

Original game moves (presumed human). That the game contains no real assistance is not established.

Why is the estimate above 0%? The regressor has estimation error and cannot go below zero, so original games receive a positive estimate on average (8.2% across original test player-games). The estimate also reflects resemblance to the synthetic benchmark, not detected assistance.

scores from models trained and validated on the controlled substitution benchmark

All model scores

Model	Score	Above original val. games	Synthetic-equiv. estimate
A0Baseline z(top-1) only	z = +0.20	56.7%	—
AStatistical baseline	0.412	64.2%	11.4%
BStockfish-XGB	0.351	55.7%	8.1%
CStockfish+Maia-XGB	0.244	48.8%	6.0%
DStockfish+Maia+Sequence-XGB	0.223	47.3%	5.7%

Scores (0–1) measure similarity to controlled synthetic assistance; they are not probabilities of cheating. A0 is a z-statistic. “Above original val. games” is the share of original (presumed-human) validation player-games with a lower score. The synthetic-equivalent estimate is the regressor's estimate of the fraction of eligible decisions that would have to be engine-substituted to produce a comparable pattern.

Deviation from expected · Regan-inspired statistical view

Compared with expectation

Why it matters: strong play alone is not suspicious — the question is whether it is far beyond what players of this rating achieve in positions this difficult.

Agreement with the engine's first choiceProvenance: MoveLens — Our approximation. Our simplified, empirical implementation inspired by a Regan concept. It does not reproduce his model or calibration.
expected 37% · observed 39% · 36 moves
Within the normal range
z = +0.20
Agreement with an engine-best move (ties count)Provenance: MoveLens — Our approximation. Our simplified, empirical implementation inspired by a Regan concept. It does not reproduce his model or calibration.
expected 41% · observed 58% · 36 moves
Within the normal range
z = +1.90
Agreement with the engine's top threeProvenance: MoveLens — Our augmentation. Added by MoveLens. Not part of Regan's documented framework.
expected 63% · observed 69% · 36 moves
Within the normal range
z = +0.69
Average move errorProvenance: MoveLens — Our augmentation. Added by MoveLens. Not part of Regan's documented framework.
expected 39 cp · observed 8 cp · 36 moves
Within the normal range
z = +1.90
Hard-position performanceProvenance: MoveLens — Our approximation. Our simplified, empirical implementation inspired by a Regan concept. It does not reproduce his model or calibration.
expected 23% · observed 33% · 15 moves
Within the normal range
z = +0.94
Ordinary-position performanceProvenance: MoveLens — Our approximation. Our simplified, empirical implementation inspired by a Regan concept. It does not reproduce his model or calibration.
expected 75% · observed 75% · 8 moves
Within the normal range
z = +0.01
Human-typicality of the movesProvenance: MoveLens — Our augmentation. Added by MoveLens. Not part of Regan's documented framework.
expected 1.07 · observed 1.06 · 36 moves
Within the normal range
z = −0.05

“Expected” is what players of this rating achieve in positions of this difficulty, learned from original training games. Results beyond about two standard deviations (|z| ≥ 2) are uncommon for genuine human play — but one game is a small sample.

Technical details

Formula: z = (observed − projected) / √(κ · Σ pᵢ(1 − pᵢ)); for move loss Var = Σ σᵢ² and the sign is flipped
Projection: Elo band × difficulty decile table fitted on original training-split decisions
κ (variance inflation): t1 1.289 · t3 1.324 · loss 1.864 · t1_hard 1.082 · maia 1.503 · ev 1.463
κ estimated on: original validation player-games (so that their z has unit variance)
Provenance: Regan-inspired approximation; the projected-vs-actual z-test concept is Regan's, this implementation is ours

Agreement with the engine's first choice: T1-match analogue (first-listed Stockfish line, depth 12)

Agreement with an engine-best move (ties count): EV-match analogue (value equal to the best line)

Agreement with the engine's top three: top-3 match (MoveLens statistic)

Average move error: mean unscaled centipawn loss, capped at 300 — not Regan's ASD

Hard-position performance: top-1 match on the hardest third of positions (not EWN)

Ordinary-position performance: top-1 match on the easiest third of positions

Human-typicality of the moves: Maia-3 surprisal vs its expectation (entropy)

Metric	n	Projected	Observed	z
Top-1 engine match fraction of eligible decisions matching the first-listed Stockfish move (cf. Regan's T1-match)	36	37.2%	38.9%	+0.20
Equal-value engine match move's value equals the best line's value at depth 12, ties included (cf. Regan's EV-match)	36	41.2%	58.3%	+1.90
Top-3 engine match fraction matching one of Stockfish's top 3 lines (ChessAssistLens statistic)	36	63.4%	69.4%	+0.69
Mean move loss unscaled centipawn loss capped at 300 - NOT Regan's scaled ASD; positive z = less loss than projected	36	39.0 cp	8.3 cp	+1.90
Top-1 match, hard positions difficulty in top tercile of training positions	15	22.8%	33.3%	+0.94
Top-1 match, ordinary positions difficulty in bottom tercile of training positions	8	74.9%	75.0%	+0.01
Human surprisal (Maia-3) expected = Maia entropy; positive z = less human-typical than projected (ChessAssistLens addition)	36	1.07	1.06	−0.05

Shaded band: |z| < 2. Points beyond ±4 are drawn as arrows at the edge; the exact z is in the column. Filled crimson marks |z| ≥ 2.

Explanation

Why the models score it this way

Why it matters: every score can be traced back to specific, inspectable evidence.

Across 36 eligible decisions, the player matched Stockfish's first choice 39% of the time against a projected 37% for this rating and these positions (z = +0.20). In the 15 hardest positions, top-1 agreement was 33% versus 23% projected (z = +0.94). Maia-3 found the chosen moves less surprising than its own expectation for this rating (1.06 vs 1.07 nats per move, z = -0.05). Model scores (similarity to controlled synthetic assistance, not a probability of cheating): baseline 0.41, human-aware 0.24, sequence-aware 0.22. The synthetic-equivalent assistance estimate is 6% of eligible decisions. Largest contributions toward the assisted pattern: share of near-optimal moves (<= 25 cp) (0.889). Largest contributions toward the human pattern: engine-best choices where the likely human move differs (0.105); mean human surprisal (Maia-3) (1.06).

Text is generated deterministically from the structured model outputs above; it does not create or modify scores.

Game-level SHAP contributions

Contributions in log-odds of the full game-level model (D) relative to its base value; the largest twelve features by magnitude. These decompose the game-level score. The per-move SHAP values in the inspector come from a separate decision-level model and do not add up to this score.

← toward human patterntoward assisted pattern →

engine-best choices where Maia-3's most probable move differsvalue 0.105−0.63
mean human surprisal (Maia-3)value 1.056−0.17
share of moves with Maia-3 predicted probability < 5%value 0.083−0.09
top-1 match in hard positions relative to projection (z)value 0.942−0.09
share of near-optimal moves (<= 25 cp)value 0.889+0.06
human-model surprisal relative to projection (z)value −0.052−0.06
top-1 engine match relative to projection (z)value 0.203−0.05
top-1 engine match ratevalue 0.389−0.03
top-3 engine match relative to projection (z)value 0.693+0.03
count of engine-best, hard, low-Maia-probability movesvalue 1+0.03
correlation of difficulty with engine agreementvalue −0.496−0.03
rate of engine-best, hard, low-Maia-probability movesvalue 0.028−0.03