Pith. sign in

REVIEW 5 major objections 6 minor 51 references

How Much Can a Few Engine Moves Help? Quantifying Limited Cheating in Chess

T0 review · 5 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A small number of well-timed engine suggestions can transform a roughly drawn chess game into a near-win, the paper argues, with expected scores rising from 0.51 to as high as 0.91 under a four-move budget.

desk verdict Worth a serious look, but the headline numbers are not yet trustworthy: the thresholds are in-sample optima and the abstract disagrees with the tables. read the letter →

arxiv 2601.05386 v2 pith:K4RKIKOQ submitted 2026-01-08 cs.AI

Daniel Keren This is my paper
classification cs.AI
keywords chesscheatingengineassistancethresholdpolicyWDLwinprobabilityscoreupliftengine-freesimulationlimitedinterventionbudgetfair-playdetection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to quantify how much a chess player's expected score improves when a much stronger engine secretly replaces just a few of their moves. In engine-vs-engine games, a threshold-based intervention policy that spends a budget of one, two, or three interventions lifts the average score from about 0.51 to 0.66, 0.76, and 0.84, respectively. It also builds a fast, engine-free simulator — using logged win-probability pairs and an empirically learned 'uplift' for interventions — that reproduces the engine-based optimum almost exactly and can tune four-threshold policies in hours rather than weeks. The practical point is that even very limited cheating can be highly effective, so detection systems need to be attuned to single well-chosen engine moves.

What carries the argument

The central object is the strong-weak gap d(t) = max(0, p_s(t) - p_w(t)) between calibrated win-draw-lose expectations of the helper's and the player's moves. The paper's intervention policy spends a limited budget of interventions when d(t) crosses a move-indexed threshold, and the thresholds themselves are tuned using an engine-free Monte Carlo simulator that samples calibrated WDL pairs from logged games and adds an empirically learned uplift Δ(t,d), computed under monotonicity constraints, for each intervention. The simulator is what makes the threshold search feasible in hours instead of weeks and lets the authors estimate optimal four-intervention policies.

What would settle it

Run a held-out evaluation: tune thresholds on a training corpus of engine games (or use the published thresholds) and compute the average score on a disjoint test corpus. If the n=1 test score falls below, say, 0.60, the reported 0.656 is overfit and does not generalize. Also, the engine-free simulator's predicted scores for a given threshold policy should be compared to actual engine-run games under that policy; systematic deviation would reveal biased uplift estimates.

Watch

Extended reading notes

Core claim

The paper's central claim is that limited, well-placed assistance in chess is far more effective than random assistance: one judicious intervention yields about the same score lift as five random ones (0.656 vs 0.585). The mechanism is a threshold policy that at each move compares the calibrated win-draw-lose (WDL) expectation of the weak player's move with that of the strong helper, and spends one of its n interventions when the gap exceeds a sequence of tuned thresholds. Averaged over a large set of engine games, the optimal thresholds yield scores of 0.656, 0.762, and 0.838 for n=1,2,3, and a four-threshold extension reaches 0.907. The authors are careful to note that engine self-play is

Load-bearing premise

The headline scores come from the same engine-game set that was used to choose the intervention thresholds, with no held-out games or error bars reported; if the threshold search overfit those games, the real-world uplift on unseen games could be materially lower.

Editorial extensions

If this is right

  • Even a single judicious intervention yields an expected score of 0.656, nearly matching five random interventions (0.585), showing that a cheater needs only one well-chosen move to gain a large edge.
  • The marginal uplift of each additional intervention shrinks, so the first few interventions dominate the score gain, a diminishing-returns pattern visible across n=1 to n=4.
  • A hindsight variant that can choose between intervening now or waiting improves the one-intervention policy by only 0.018, indicating the threshold policy is already close to the best possible single-intervention strategy.
  • The engine-free simulator's three-threshold optimum (0.835) nearly matches the full engine-based optimum (0.838), suggesting that threshold policies for new settings can be tuned cheaply and quickly without running games.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the engine-free simulator needs only logged games, the same method could be applied to real online chess data to estimate, per player and per position, how much a single engine hint would help — turning detection thresholds from universal to personalized.
  • The increasing threshold sequence suggests a general principle for bounded-assistance decision making: as the budget depletes, intervene only when the quality gap is larger. This could transfer to other settings like tutoring systems, co-pilot tools, or human-in-the-loop AI.
  • The measured engine-human conversion gap implies that the same intervention that converts a draw to a win in engine play may only convert it to a draw for a human, so platforms should not assume engine-based uplift numbers apply directly to human cheating cases.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper studies how much a chess player's expected score improves when a strong engine ('helper') intervenes in a limited number of moves. Experiments are conducted in a controlled Stockfish-vs-Stockfish setting, comparing random interventions, threshold-based policies with thresholds tuned by Bayesian optimization, a 'maximal-delta-sum' predictor, and an engine-free simulator. The main quantitative claim is that one or two well-timed engine interventions raise the average score from 0.51 to about 0.71/0.82 (abstract) or 0.656/0.762 (Table 2). The engine-free simulator is reported to reproduce the three-threshold optimum at 0.835 vs 0.838. The paper concludes that even a single judicious engine move substantially changes the expected outcome, with implications for cheating detection.

Significance. If the central estimates were properly validated, the paper would be a valuable contribution to the cheating-detection literature, quantifying the danger of even one well-timed engine move. The controlled engine-vs-engine methodology, the comparison of random/threshold/oracle policies, and the engine-free simulator are sensible and potentially useful directions. The engine-free simulator, with its 50,000-game logs and fast Bayesian optimization, is a promising engineering contribution if its fidelity is confirmed. However, the paper's current evidence does not support the specific numbers stated in the abstract and introduction: the headline results are in-sample optima without held-out evaluation, and the paper contains mutually inconsistent numerical claims. The qualitative direction (cheating improves outcomes) is plausible, but the precise uplift—the paper's main contribution—is not yet reliably established.

major comments (5)
  1. [Table 2 / Sec. 4.1] The thresholds T_1..T_3 are selected by Bayesian optimization on the same engine games that are then scored. No held-out evaluation, cross-validation, or error bars are described. The reported averages 0.656/0.762/0.838 are therefore in-sample maxima over the threshold candidates and are likely optimistically biased. This is the paper's headline quantity, so the specific uplift values are not established.
  2. [Abstract vs. Introduction vs. Table 2] The abstract states average scores 0.71 and 0.82 for one and two interventions; the introduction states 0.67/0.77/0.85/0.91 for 1-4 interventions; Table 2 reports 0.656/0.762/0.838 for n=1,2,3. These conflicting numbers are never reconciled. The paper must state which numbers come from which algorithm/experiment and which are authoritative.
  3. [Sec. 6] The engine-free simulator is validated only by comparing its optimized three-threshold score (0.835) with the engine-based optimum (0.838). Both are in-sample: the simulator is trained on the same no-intervention and random-intervention logs, and the thresholds are optimized on the simulator. The simulator's additive uplift model (v <- v + lambda_k Delta(t,d)) assumes additivity and no path dependence, which is not tested. A single concordant operating point does not validate the simulator for hyperparameter search.
  4. [Sec. 5] The maximal-delta-sum algorithm is introduced as a contribution, but no final average scores are reported for it; the only numeric is a 0.018 improvement for a hindsight variant over an unnamed baseline. Without an evaluation table, the 'Bellman-style intervention policies' mentioned in the abstract are not supported by evidence.
  5. [Secs. 3-4] No experimental details are given for the engine-based results: number of games per condition, ELO settings of W, B, and C, time control, or Stockfish version (except in the reference list). Without this information the empirical numbers cannot be reproduced, and the '0.51 no-cheat' baseline cannot be independently checked.
minor comments (6)
  1. [Abstract] 'whichdetection' is missing a space; should be 'which detection'.
  2. [Sec. 4.1] The sentence 'If the average WDL following C's moves is larger than by at least T1 then the average following W's moves' is ungrammatical; rephrase.
  3. [Algorithm 1] Variable names (p_W, p_S, S-hat, epsilon) are not all introduced in the surrounding text; a notation table would help.
  4. [Sec. 6] In step 3 of the simulator, 'spend assist i' is used but the loop variable is k; align the indexing.
  5. [Section numbering] The text refers to 'Appendix 2' but the full text labels it Section 8; renumber for consistency.
  6. [References] Reference [19] is listed as 'Personal communication' and is not verifiable; replace with a citable source or remove.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional circularity; empirical engine self-play results.

full rationale

The paper's central quantitative claims are empirical measurements from Stockfish self-play under a threshold intervention policy, not derivations that reduce to their own inputs. The thresholds in Table 2 are selected by Bayesian optimization on the same average-score metric used to report the results, so those figures are in-sample optima and may be optimistic; this is a statistical validation caveat, not a case where an output equals an input by construction or where a fitted parameter is relabeled as an independent prediction. There are no load-bearing self-citations: the author cites no prior work of their own, and no uniqueness theorem or ansatz is imported from prior work by the same authors. The engine-free simulator is calibrated on separate no-intervention and randomized-intervention logs and is checked against the engine-based optimum; that is a model fitting procedure, not a circular reuse of the target values. The paper's internal numeric inconsistencies (Abstract 0.71/0.82, Introduction 0.67/0.77, Table 2 0.656/0.762) weaken the precision of the claim but are not circularity. Overall, no circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central measurements are not derived; they are produced by a pipeline whose key numbers (thresholds, calibration curves, uplift table, lambda_k) are fitted to simulated games. The main unstated costs are the additivity of uplifts and the transferability of engine results to human play.

free parameters (5)
  • Intervention thresholds T_1..T_n = n=1: 0.205; n=2: 0.141,0.199; n=3: 0.133,0.147,0.193; engine-free n=4: 0.15437,0.17808,0.19909,0.24803
    Chosen by Bayesian optimization to maximize expected score; all reported average scores (Table 2, Section 6) are evaluated at these fitted optima.
  • Isotonic/neural WDL calibration functions = fitted per move number m to (WDL, result) pairs
    The policies replace raw WDL with calibrated p_w, p_s from these fits (Section 3.1); the central decision rule depends on them.
  • Slacks epsilon_1, epsilon_2, epsilon_3 in Algorithm 1 = not reported
    Appear in the maximal-delta intervention conditions but no values or tuning procedure are given; presumably tuned.
  • Uplift weights lambda_k in engine-free simulator = not reported
    Section 6 step 3 applies lambda_k * Delta(t,d) to accumulate score changes; lambda_k is never defined.
  • Simulation depth for threshold decisions = 10
    The policy averages WDL over 10 simulated moves after each Black move; authors note similar results for larger values.
assumptions (5)
  • domain assumption Stockfish WDL output, after monotone calibration, is an unbiased estimate of the expected game score (p(score=1|x)).
    Used throughout Section 3.1 to define p_w and p_s and to compare weak vs strong moves; if calibration fails, the threshold comparisons are off.
  • domain assumption A maximal-ELO Stockfish helper (3190) is a sufficiently strong oracle; its top move is the correct cheating move.
    Section 3 defines C as championship strength; the whole simulation treats C's move as the strong move.
  • domain assumption Engine-vs-engine results transfer to human-vs-human play after subtracting a constant discount Delta_bar ≈ 0.08.
    Appendix 2 applies this correction using a human dataset; if the conversion gap depends on policy/position in ways not captured, the human projections are invalid.
  • ad hoc to paper The total effect of multiple interventions is the sum of per-intervention uplifts learned from independent single-intervention games (additivity and no path dependence).
    Section 6 step 3 accumulates v <- clip(v + lambda_k * Delta(t,d)); interventions change the position and future gaps, so additivity is an approximation.
  • ad hoc to paper The distribution of WDL pairs in no-intervention games (bank B_t) remains valid for games in which interventions occur.
    The engine-free simulator samples from D_0 no-intervention logs while simulating threshold policies; an intervention changes the game trajectory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Much Can a Few Engine Moves Help? Quantifying Limited Cheating in Chess." pith.science (2026). https://pith.science/paper/K4RKIKOQ

@misc{pith2026260105386,
  author       = {Pith},
  title        = {Pith review of: How Much Can a Few Engine Moves Help? Quantifying Limited Cheating in Chess},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4RKIKOQ}},
  note         = {Machine review of arXiv:2601.05386}
}
read the original abstract

Cheating in chess, by using advice from powerful software, has become a major problem, reaching the highest levels. As opposed to the large majority of previous work, which concerned {\em detection} of cheating, here we try to evaluate the possible gain in performance, obtained by cheating a limited number of times during a game. We develop threshold-based and Bellman-style intervention policies, and test them in a controlled engine-vs-engine setting using Stockfish. A judicious choice of 1 or 2 cheats yields average scores of 0.71 and 0.82, respectively, compared to 0.51 with no cheats. We also introduce a fast, engine-free simulator that enables hyperparameter optimization without running games, closely matching the engine-based optimum. The goal of this work is not to assist cheaters, but to measure the effectiveness of cheating -- which is crucial as part of the effort to contain and detect it.

Figures

Figures reproduced from arXiv: 2601.05386 by the authors.

Figure 1
Figure 1. Monotonic fitting vs. engine WDL. Top: isotonic regression. Bottom: monotone neural net. The [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Example of how “strong” moves (by C improve on “weak” ones (by W): when tuned to ELO 3190, the engine finds a better move than the ELO 1500 engine, as it realizes the danger of Black doubling rooks on the A-file. 8 Appendix 2: relevance to human-vs-human cheating We ran the intervention/cheating algorithms in engine-vs-engine play. Here we comment on the relevance and expected results for human-vs-human play, where … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 6 canonical work pages

  1. [1]

    G. Avni, T. A. Henzinger, and V. Chonev. Infinite-duration bidding games.J. ACM, 66(4):31:1–31:29, 2019

    2019

  2. [2]

    G. Avni, T. A. Henzinger, and D. Zikelic. Bidding mechanisms in graph games.J. Comput. Syst. Sci., 119:133–144, 2021

    2021

  3. [3]

    G. Avni, I. Jecker, and Ð. Žikelić. Infinite-duration all-pay bidding games. InProc. 32nd SODA, pages 617–636, 2021

    2021

  4. [4]

    Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H

    Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The hanabi challenge: A new frontier for ai research.https://arxiv.org/abs/1902.00506, 2019. arXiv:1902.00506

    arXiv 1902

  5. [5]

    Barnes and Julio Hernandez-Castro

    David J. Barnes and Julio Hernandez-Castro. On the limits of engine analysis for cheating detection in chess.Computers & Security, 48:58–73, 2015. doi: 10.1016/j.cose.2014.10.002

    work page doi:10.1016/j.cose.2014.10.002 2015

  6. [6]

    Joseph Hoane, and Feng-hsiung Hsu

    Murray Campbell, A. Joseph Hoane, and Feng-hsiung Hsu. Deep blue.Artificial Intelligence, 134(1–2): 57–83, 2002. doi: 10.1016/S0004-3702(01)00129-1

    doi:10.1016/s0004-3702(01)00129-1 2002

  7. [7]

    My statement regarding the sinquefield cup (full text mirror).https://en

    Magnus Carlsen. My statement regarding the sinquefield cup (full text mirror).https://en. chessbase.com/post/carlsen-statement-on-niemann, 2022. Accessed 2025-10-04

    2022

  8. [8]

    Kasparov vs deep junior — match details.https://en.chessbase.com/post/ kasparov-vs-deep-junior-match-details, 2003

    ChessBase. Kasparov vs deep junior — match details.https://en.chessbase.com/post/ kasparov-vs-deep-junior-match-details, 2003. Accessed 2025-10-04

    2003

Show all 51 references
  1. [9]

    Chess.com Concludes Legal Dispute With Hans Niemann, Niemann To Return To Chess.com

    Chess.com. Chess.com Concludes Legal Dispute With Hans Niemann, Niemann To Return To Chess.com. https://www.chess.com/blog/CHESScom/chess-com-concludes-legal-dispute-with-hans-niemann-niemann-to-return-to-chess-com,

  2. [10]

    Chess cheating — our fair play system explained.https://www.chess.com/cheating,

    Chess.com. Chess cheating — our fair play system explained.https://www.chess.com/cheating,

  3. [11]

    Hans niemann report.https://www.chess.com/blog/CHESScom/ hans-niemann-report, 2022

    Chess.com Fair Play Team. Hans niemann report.https://www.chess.com/blog/CHESScom/ hans-niemann-report, 2022. Overview article with link to full report; accessed 2025-10-04

    2022

  4. [12]

    Hans niemann report (full pdf).https://www.documentcloud.org/ documents/23118744-oct-2022-final-h-niemann-report/, 2022

    Chess.com Fair Play Team. Hans niemann report (full pdf).https://www.documentcloud.org/ documents/23118744-oct-2022-final-h-niemann-report/, 2022. 72 pages; accessed 2025-10-04

    2022

  5. [13]

    Kasparov versus deep junior 2003.https://www.chessprogramming.org/ Kasparov_versus_Deep_Junior_2003, 2024

    ChessProgramming Wiki. Kasparov versus deep junior 2003.https://www.chessprogramming.org/ Kasparov_versus_Deep_Junior_2003, 2024. Match summary; accessed 2025-10-04

    2003

  6. [14]

    Uncertainty-aware action advising for deep reinforcement learning agents

    Felipe Leno da Silva, Garrett Warnell, Peter Stone, and Anna Helena Reali Costa. Uncertainty-aware action advising for deep reinforcement learning agents. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5792–5799, 2020. URLhttps://ojs.aaai.org...

    2020

  7. [15]

    Isotone optimization in R: Pool-adjacent-violators algorithm (pava) and active set methods.Journal of Statistical Software, 32(5):1–24, 2009

    Jan de Leeuw, Kurt Hornik, and Patrick Mair. Isotone optimization in R: Pool-adjacent-violators algorithm (pava) and active set methods.Journal of Statistical Software, 32(5):1–24, 2009. URL https://www.jstatsoft.org/v32/i05/

    2009

  8. [16]

    Haworth, and Kenneth W

    Giuseppe Di Fatta, Guy McC. Haworth, and Kenneth W. Regan. Skill rating by bayesian inference. In 2009 IEEE Symposium on Computational Intelligence and Data Mining (CIDM), pages 89–94. IEEE,

    2009

  9. [17]

    Elo.The Rating of Chessplayers, Past and Present

    Arpad E. Elo.The Rating of Chessplayers, Past and Present. Arco Publishing, New York, 1978. ISBN 0668047216

    1978

  10. [18]

    Decision in case 8/2019: Allegations of cheating against igor rau- sis.https://www.fide.com/docs/decisions-resolutions/Ethics%20case%208_2019%20Decision

    FIDE Ethics Commission. Decision in case 8/2019: Allegations of cheating against igor rau- sis.https://www.fide.com/docs/decisions-resolutions/Ethics%20case%208_2019%20Decision. pdf, 2019. Decision dated Dec 5, 2019; accessed 2025-10-04

    2019

  11. [19]

    Personal communication

    GM. Personal communication. Conversations with expert players, including chess Grandmasters, 2025

    2025

  12. [20]

    Computer analysis of world chess champions.ICGA Journal, 29(2):65–73,

    Matej Guid and Ivan Bratko. Computer analysis of world chess champions.ICGA Journal, 29(2):65–73,

  13. [21]

    Haworth, Kenneth W

    Guy McC. Haworth, Kenneth W. Regan, and Giuseppe Di Fatta. Performance and prediction: Bayesian modelling of fallible choice in chess. InAdvances in Computer Games (ACG 2009), volume 6048 of Lecture Notes in Computer Science, pages 99–110. Springer, 2010. doi: 10.1007/978-3-64...

    2009 doi

  14. [22]

    Detecting fair play violations in chess using neural networks

    Maksim Iavich and Zura Kevanishvili. Detecting fair play violations in chess using neural networks. In IVUS 2024: Information Society and University Studies, volume 3885 ofCEUR Workshop Proceedings,

    2024

  15. [23]

    Deep blue: Building a digital chess master.https://www.ibm.com/history/deep-blue,

    IBM Archives. Deep blue: Building a digital chess master.https://www.ibm.com/history/deep-blue,

  16. [24]

    Knuth and Ronald W

    Donald E. Knuth and Ronald W. Moore. An analysis of alpha-beta pruning.Artificial Intelligence, 6 (4):293–326, 1975. doi: 10.1016/0004-3702(75)90019-3

    1975 doi

  17. [25]

    Towards transparent cheat detection in online chess: An application of human and computer decision-making preferences

    Thijs Laarhoven and Aditya Ponukumati. Towards transparent cheat detection in online chess: An application of human and computer decision-making preferences. InComputers and Games: CG 2022, Revised Selected Papers, volume 13865 ofLecture Notes in Computer Science, pages 163–18...

    2022

  18. [26]

    A. J. Lazarus, D. E. Loeb, J. G. Propp, W. R. Stromquist, and D. H. Ullman. Combinatorial games under auction play.Games and Economic Behavior, 27(2):229–264, 1999

    1999

  19. [27]

    Leela chess zero.https://lczero.org/, 2018

    Leela Chess Zero Project. Leela chess zero.https://lczero.org/, 2018. Accessed 2025-10-04

    2018

  20. [28]

    Learning models of individual behavior in chess

    Rylan McIlroy-Young and Ashton Anderson. Learning models of individual behavior in chess. In Proceedings of the 28th ACM SIGKDD International Conference on Knowledge Discovery & Data Min- ing (KDD ’22), pages 1154–1164. ACM, 2022. URLhttps://www.cs.toronto.edu/~ashton/pubs/ ma...

    2022

  21. [29]

    Aligning superhuman ai with human behavior: Chess as a model system

    Rylan McIlroy-Young, Siddhartha Sen, Jon Kleinberg, and Ashton Anderson. Aligning superhuman ai with human behavior: Chess as a model system. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’20), pages 1677–1687. ACM, 202...

    2020

  22. [30]

    doi: 10.1007/978-3-031-34017-8_14

    doi

  23. [31]

    Efficiently updatable neural networks for board evaluation (nnue).https://nnue.uzudilab

    Yu Nasu. Efficiently updatable neural networks for board evaluation (nnue).https://nnue.uzudilab. com/nnue2018.pdf, 2018. Technical report (Shogi); concept later adapted in chess engines; accessed 2025-10-04

    2018

  24. [32]

    Peres, O

    Y. Peres, O. Schramm, S. Sheffield, and D. Bruce Wilson. Random-turn hex and other selection games. The American Mathematical Monthly, 114(5):373–387, 2007

    2007

  25. [33]

    Regan and Guy McC

    Kenneth W. Regan and Guy McC. Haworth. Intrinsic chess ratings. InProceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence (AAAI-11), San Francisco, CA, USA, 2011

    2011

  26. [34]

    Regan, Bartłomiej Macieja, and Guy McC

    Kenneth W. Regan, Bartłomiej Macieja, and Guy McC. Haworth. Understanding distributions of chess performances. InAdvances in Computer Games (ACG 2011), volume 7168 ofLecture Notes in Computer Science. Springer, 2012. doi: 10.1007/978-3-642-31866-5_20

    2011 doi

  27. [35]

    Nalimov, Guy Wirth, and Guy McC

    Eugene A. Nalimov, Guy Wirth, and Guy McC. Haworth. Endgame tablebases: History, current state, and prospects.ICGA Journal, 23(1):15–19, 2000

    2000

  28. [36]

    Stockfish chess engine.https://stockfishchess.org/, 2008

    Tord Romstad, Marco Costalba, Joona Kiiski, Gary Linscott, and The Stockfish Developers. Stockfish chess engine.https://stockfishchess.org/, 2008. Version 17.1; accessed 2025-10-04

    2008

  29. [37]

    Cheating in online chess (part ii): The analysis of engine use, 2024

    Alexey Root. Cheating in online chess (part ii): The analysis of engine use, 2024. URLhttps://www. chessable.com/blog/cheating-in-online-chess-pt-2/. Chessable Blog, May 10, 2024. Two-part series on online cheating

    2024

  30. [38]

    Rosenthal

    Jeffrey S. Rosenthal. Probabilities of streaks in online chess.https://probability.ca/jeff/ftpdir/ chessstreakpaper.pdf, 2024. Preprint; accessed 2025-10-04

    2024

  31. [39]

    Claude E. Shannon. Programming a computer for playing chess.The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 41(314):256–275, 1950. doi: 10.1080/ 14786445008521796

    1950

  32. [40]

    Carlsen and Niemann settle dispute over cheating claims that rocked chess.https://www

    Reuters. Carlsen and Niemann settle dispute over cheating claims that rocked chess.https://www. reuters.com/sports/chess-carlsen-niemann-settle-dispute-over-cheating-claims-2023-08-29/,

    2023

  33. [41]

    Aug 29, 2023; accessed 2025-10-04

    2023

  34. [42]

    A general reinforcement learning algorithm that masters chess, shogi, and go through self-play.Science, 362(6419):1140–1144,

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, et al. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play.Science, 362(6419):1140–1144,

  35. [43]

    PracticalBayesianoptimizationofmachinelearning algorithms

    JasperSnoek, HugoLarochelle, andRyanP.Adams. PracticalBayesianoptimizationofmachinelearning algorithms. InAdvances in Neural Information Processing Systems (NeurIPS), pages 2951–2959, 2012

    2012

  36. [44]

    Stockfish (github repository).https://github.com/official-stockfish/ Stockfish, 2008

    The Stockfish Developers. Stockfish (github repository).https://github.com/official-stockfish/ Stockfish, 2008. Accessed 2025-10-04. 10

    2008

  37. [46]

    Griffiths

    Minkyu Shin, Jin Kim, Bas van Opheusden, and Thomas L. Griffiths. Superhuman artificial intelligence can improve human decision-making by increasing novelty.Proceedings of the National Academy of Sciences, 120(13):e2214840120, 2023. doi: 10.1073/pnas.2214840120. URLhttps://www...

    2023 doi

  38. [47]

    Monotonic networks

    Joseph Sill. Monotonic networks. InAdvances in Neural Information Processing Systems (NIPS), 1997. URLhttps://papers.nips.cc/paper/1358-monotonic-networks.pdf

    1997

  39. [2006]

    doi: 10.3233/ICG-2006-29203. 8

    2006 doi

  40. [2009]

    doi: 10.1109/CIDM.2009.4938634

    2009

  41. [2018]

    doi: 10.1126/science.aar6404. 9

    doi

  42. [2023]

    Aug 28, 2023; accessed 2025-10-04

    2023

  43. [2024]

    URLhttps://ceur-ws.org/Vol-3885/paper13.pdf

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.