Ailed: A Psyche-Driven Chess Engine with Dynamic Emotional Modulation

Diego Armando Resendez Prado

cs.AIMar 5, 2026

Ailed: A Psyche-Driven Chess Engine with Dynamic Emotional Modulation

Authors: Diego Armando Resendez Prado

Organizations: Independent Researcher

Abstract

Chess engines passed human strength years ago, but they still don't play like humans. A grandmaster under clock pressure blunders in ways a club player on a hot streak never would. Conventional engines capture none of this. This paper proposes a personality x psyche decomposition to produce behavioral variability in chess play, drawing on patterns observed in human games. Personality is static -- a preset that pins down the engine's character. Psyche is dynamic -- a bounded scalar ψ_t \in [-100, +100], recomputed from five positional factors after every move. These two components feed into an audio-inspired signal chain (noise gate, compressor/expander, five-band equalizer, saturation limiter) that reshapes move probability distributions on the fly. The chain doesn't care what engine sits behind it: any system that outputs move probabilities will do. It needs no search and carries no state beyond ψ_t. I test the framework across 12,414 games against Maia2-1100, feeding it two probability sources that differ by ~2,800x in training data. Both show the same monotonic gradient in top-move agreement (~20-25 pp spread from stress to overconfidence), which tells us the behavioral variation comes from the signal chain, not from the model underneath. When the psyche runs overconfident, the chain mostly gets out of the way (66% agreement with vanilla Maia2). Under stress, the competitive score falls from 50.8% to 30.1%. The patterns are reminiscent of tilt and overconfidence as described in human play, but I should be upfront: this study includes no human-subject validation.

Figures & tables

Factor	Symbol	Domain	Description
Material	$f_{1}(b)$	$\mathbb{Z}$	$\sum_{p}v_{p}n_{p}^{\text{own}}-\sum_{p}v_{p}n_{p}^{\text{opp}}$ , $\mathbf{v}=(1,3,3,5,9)$
King safety	$f_{2}(b)$	$\mathbb{Z}$	Pawn shield minus enemy attackers in king’s $3\!\times\!3$ area
Mobility	$f_{3}(b)$	$\mathbb{Z}$	$\|\mathcal{L}^{\text{own}}\|-\|\mathcal{L}^{\text{opp}}\|$ (own only if in check)
Center ctrl	$f_{4}(b)$	$\mathbb{Z}$	Own minus opp. attackers of $\{d4,d5,e4,e5\}$
Vulnerability	$f_{5}(b)$	$\mathbb{Z}_{\leq 0}$	$-\|\{s:\text{own piece on }s\text{ attacked}\}\|$

Table 1: Positional factor definitions for psyche computation.

Panel of Figure 1 — Figure 1: Psyche computation pipeline. Five positional factors are extracted from the board, weighted, and summed to produce a raw evaluation $e_{t}$ , which is tanh-compressed to a target psyche $\hat{\psi}_{t}\in[-100,+100]$ .

Psyche state	Best	Good	Mild	Bad	Worst
Stress ( $\psi=-100$ )	0.80	1.30	1.00	1.40	1.20
Neutral ( $\psi=0$ )	1.30	1.20	1.00	0.70	0.50
Overconfident ( $\psi=+100$ )	1.00	1.00	1.00	1.00	1.00

Table 2: EQ band gains for the “human” personality preset across psyche states.

Preset	Dynamics $\alpha$	Gate $\tau_{g}$	Saturation $\sigma$	Character
flat	1.0 / 1.0	0 / 0 / 0	1.0 / 1.0 / 1.0	Bypass
classical	0.8 / 1.5	.01 / .02 / .05	.50 / .60 / .85	Disciplined
rock	0.6 / 1.6	.005 / .01 / .03	.35 / .45 / .70	Bold
jazz	0.5 / 1.4	.001 / .005 / .01	.25 / .35 / .50	Creative
metal	0.4 / 1.3	0 / .001 / .005	.20 / .30 / .50	Chaotic
human	0.5 / 2.0	.005 / .02 / .06	.30 / .50 / .85	Realistic

Table 3: Personality presets: dynamics and gate/saturation anchors (stress / neutral / overconfident).

Panel of Figure 2 — Figure 2: Signal chain parameters for the “human” personality preset as a function of psyche $\psi$ . Gate threshold is scaled $\times 20$ for visibility. The dynamics power $\alpha$ crosses unity at $\psi=0$ : below neutral it acts as an expander (flattening the distribution), above as a compressor (sharpening it). All parameters interpolate linearly between the three psyche anchors.

Panel of Figure 3 — Figure 3: EQ band gains for the “human” personality preset at three psyche conditions, displayed as equalizer panels. Each dot is a band control point (Best, Good, Mild, Bad, Worst); the smooth interpolated curve and shaded fill show the gain deviation from the white unity baseline (above $=$ boost; below $=$ cut). Under stress ( $\psi=-70$ ), humps at “good” and “bad” shift weight toward adjacent tiers. At neutral ( $\psi=0$ ), a strong monotone preference ordering is imposed. At overconfidence ( $\psi=+70$ ), all deviations converge toward zero and the EQ becomes nearly flat.

Panel of Figure 4 — Figure 4: Ailed system architecture. The move predictor $f_{\theta}$ produces logits fed into the move selector; the psyche calculator computes the psyche state $\psi_{t}$ from board position evaluations. Psyche state modulates all three output paths: move selection (signal chain), thinking mode (plan disruption), and study mode (skip probability and exploration quality). Dashed arrows indicate conditional paths.

Hyperparameter	Value
Vocabulary $\|\mathcal{V}\|$	4,547
Embedding dim $d$	512
Attention heads	8
Layers $N$	6
FFN dim	2,048
Max sequence length $L_{\max}$	300

Table 4: Move predictor hyperparameters.

Panel of Figure 5 — Figure 5: Thinking mode lifecycle. A plan is generated when confidence exceeds the threshold, stored in a single-slot buffer, and executed only if the opponent’s actual move matches the prediction and the psyche disruption check passes.

Experiment	Condition	Games	W	D	L	Score	Avg Ply
A: Ailed-60k	Stress	667	22	79	566	9.2%	62
	Neutral	667	36	60	571	9.9%	61
	Overconfident	667	33	65	569	9.8%	61
B: Maia2+Psyche	Stress	667	123	155	389	30.1%	68
	Neutral	667	95	402	170	44.4%	39
	Overconfident	667	41	596	30	50.8%	16

Table 5: Match results across experiments and controls. Score is from the psyche-modulated side’s perspective against vanilla Maia2-1100. Controls use Maia2+Psyche (Experiment B’s base) with chain disabled.

Panel of Figure 6 — Figure 6: Average psyche trajectories across 667 games per condition. Background shading marks the three psyche zones: light red ( $\psi<-33$ , stress), white ( $|\psi|\leq 33$ , neutral), light blue ( $\psi>33$ , overconfident); dotted lines at $\psi=\pm 33$ mark zone boundaries. Shaded curve areas show the psyche footprint relative to neutral. (a) Ailed-60k: all conditions converge toward $\psi\approx-100$ by move 30; meaningful three-way separation is confined to the first ${\sim}$ 15 moves. (b) Maia2+Psyche: the stronger model sustains separation throughout—stress stabilises around $-40$ , neutral around $-30$ , and overconfident decays gradually from $+80$ without collapsing. The initial forced psyche is visible at move 0 in both panels.

Panel of Figure 7 — Figure 7: Signal chain auto-tuning during a single stress game (Ailed-60k Game 502, $\psi_{0}=-80$ , 54 moves, draw). (a) Psyche trajectory: forced stress ( $-77$ ) recovers through neutral into overconfidence (peak $+73$ at move 27), then collapses back to $-98$ as the position deteriorates. (b) EQ band gains shown as deviation from unity (positive $=$ boost, negative $=$ cut). At low $\psi$ (moves 0–14), “good” is strongly boosted ( $+0.28$ ) while “worst” is cut ( $-0.27$ ), and “best” is suppressed—a diffuse selection profile. At the overconfident peak (move 27), all deviations shrink toward zero—the EQ becomes transparent. In the degradation phase (moves 30+), the pattern resembles the low- $\psi$ opening. (c) Signal chain parameters track the psyche: dynamics power crosses unity at neutral (expander below, compressor above), gate opens under stress and tightens at confidence, saturation ceiling rises and falls with psyche.

Experiment	Condition	Moves	Entropy	Confidence	Top Agree%
A: Ailed-60k	Stress	20,612	1.610	0.459	37.1 [36.3, 37.9]
	Neutral	20,086	1.639	0.466	45.9 [45.2, 46.6]
	Overconfident	20,337	1.664	0.469	57.0 [56.4, 57.6]
B: Maia2+Psyche	Stress	22,606	1.612	0.497	41.2 [40.5, 41.9]
	Neutral	12,901	1.542	0.529	51.8 [51.0, 52.6]
	Overconfident	5,180	1.515	0.575	66.0 [65.0, 67.0]

Table 6: Signal chain behavioral metrics across conditions (per-game means; 95% bootstrap CI for controls). Moves = total modulated-side moves; “–” indicates move counts not recorded for control runs.

Panel of Figure 8 — Figure 8: Top-move agreement by condition across both experiments (4,002 games). Both engines show the same monotonic gradient: stress reduces agreement (more deviation from model’s top pick), overconfidence increases it (near-deterministic selection). The gradient is steeper for Maia2+Psyche because its stronger model probabilities give the signal chain more concentrated input to work with.

Metric	Ailed-60k spread	Maia2+Psyche spread
Score	$+0.6$ pp (9.2 $\to$ 9.8)	$+20.7$ pp (30.1 $\to$ 50.8)
Top agree%	$+19.9$ pp (37.1 $\to$ 57.0)	$+24.8$ pp (41.2 $\to$ 66.0)
Confidence	$+0.010$ (0.459 $\to$ 0.469)	$+0.078$ (0.497 $\to$ 0.575)
Entropy	$+0.054$ (1.610 $\to$ 1.664)	$-0.097$ (1.612 $\to$ 1.515)

Table 7: Signal chain gradient comparison: stress

\to

overconfident spread across engines.

Engine	Mode	Stress	Neutral	Overconf.	Spread	Score s $\to$ o
Ailed-60k	Full	37.1	45.9	57.0	19.9 pp	9.2–9.8%
Ailed-60k	Entropy-only	37.4	45.8	56.3	18.9 pp	10.0–5.5%
Ailed-60k	Fixed ( $c{=}1$ )	36.5	44.2	55.4	18.9 pp	10.2–9.0%
Maia2+Psyche	Full	41.2	51.8	66.0	24.8 pp	30.1–50.8%
Maia2+Psyche	Entropy-only	43.2	51.5	71.3	28.1 pp	25.8–52.0%
Maia2+Psyche	Fixed ( $c{=}1$ )	41.1	50.1	71.6	30.5 pp	27.8–50.5%

Table 8: Outcome predictor ablation: top-move agreement (%) and spread across confidence modes. Spread = overconfident

-

stress.

Configuration	Stress	Neutral	Overconf.	Spread
Full chain	41.9	60.1	76.6	34.7 pp
No gate	35.6	53.5	72.0	36.4 pp
No dynamics	54.1	61.7	68.4	14.3 pp
No EQ	44.3	62.5	74.9	30.6 pp
No saturation	48.3	64.7	78.2	29.9 pp
No gate + no dynamics	51.7	52.9	54.1	2.4 pp

Table 9: Stage-wise signal chain ablation: top-move agreement (%) by psyche zone (Maia2+Psyche, 334 games per zone per config). Flat control (667 games/zone) shown for reference. Spread = overconfident

-

stress.

Finding	Evidence	Ref.
Signal chain produces monotonic gradient	Top agree: 37–57% (A), 41–66% (B)	Section 7.2.4
Gradient is architecture-agnostic	${\sim}$ 20–25 pp spread for both engines	Section 7.2.5
Overconfident = near-transparent	66% Maia2 agree, 596/667 draws (B)	Section 7.2.1
Stress degrades competitive score	30.1% stress vs. 50.8% overconfident (B)	Section 7.2.1
Confidence tracks psyche state	Monotonic increase: 0.459 $\to$ 0.575	Section 7.2.4
Training data gap limits Ailed	9.2–9.9% score vs. 30–51% for Maia2+Psyche	Section 7.2.1

Table 10: Summary of key experimental findings from 12,414 games across four experiments.

Zone	Human CPL [95% CI]	Ailed CPL [95% CI]	Difference
Panic	64.9 [63.3, 66.5]	48.4 [39.6, 58.5]	$+16.5$
Stressed	96.8 [94.8, 98.9]	66.8 [55.7, 78.8]	$+30.0$
Neutral	76.8 [75.2, 78.4]	61.3 [53.2, 70.1]	$+15.5$
Confident	79.4 [77.9, 81.0]	65.0 [54.1, 77.5]	$+14.4$
Overconf	54.5 [53.4, 55.7]	58.2 [50.1, 67.6]	$-3.7$

Table 11: Mean centipawn loss per psyche zone: human 1100-ELO players (Lichess 2016,

n=7{,}748

games, 25,000 moves/zone, Stockfish depth 8) vs. Ailed (2,001 experiment games, 300 moves/zone). Difference = Human

-

Ailed.

Game-Playing Agents

Game-playing agents use search, learning, and reasoning to choose actions in games with strategic, uncertain, or changing environments.

See all 194 papers

Jul 23, 2026

Chess_db: A framework for working with large chess game datasets

Chess is a two player strategic game that is embedded in classical AI culture as it was once the frontier for intelligent behaviour. There was the silent assumption that the advent of computer engines that play better than the best humans will extinguish interest in the game. However, the opposite has come to pass, with a growing following for the game. A lot of the computational resources are now centered around training of players, where the engine output is just one aspect. Access to past games is also an essential part, both in knowing what games a specific player has played previously, and also which continuations at a certain position have led to victory more often for each of the two colour players. We present Chess_db a suite of logic programming tools that can effectively manipulate games both in memory and via creating back end databases. In particular, we provide versatile code that creates databases from PGN (portable game notation) game files and explore the suitability of open source key-value databases for storing position tables that provide near-instant access to information pertaining to substantially large number of games.

Nicos Angelopoulos, Jan Wielemaker

University College & Imperial College, London UK · University College & Imperial College London, UK · SWI-Prolog solutions +1

Logic Programming Chess

Ask PDF

Jun 23, 2026

Matilda: Engine-Agnostic Search with Human Policy Guidance

Chess engines have evolved from search-based systems optimized solely for strength to neural policies capable of modeling human decisions across much of the rating spectrum. Maia-3, the strongest human-like move policy for chess, models the typical moves of a given rating but does not model high Elo strength well (2500+ Lichess Elo) nor has an extensible architecture for modeling individual play style. On the other hand, search-based engines like Stockfish are far stronger than any known human but struggle to model human-like play. To solve these problems, we present Matilda, a permutation-invariant set transformer that re-ranks the full legal-move distribution produced by a frozen Maia-3 and improves it along both dimensions with one 1.7M-parameter model. Matilda uses Maia-3 to provide both global context and search candidates. Context is encoded from Maia-3's hidden representation, time control, and an optional 32-dimensional player-style vector, while the top-16 candidate moves are optionally rescored by Stockfish. A zero-initialized head scatters per-candidate adjustments back into the policy logits, so the untrained model is exactly Maia-3 and every gain is value-added under full-vocabulary NLL. On a 2500+ Elo benchmark from Lichess, Matilda improves on Maia-3 by +0.5-0.6% below 2800 but +4.3%, +11.9%, and +21.9% at 2800-2900, 2900-3000, and 3000+, respectively, and an additional +0.41% overall from player-style embeddings. Ablations show that the improvements are primarily driven by the engine-derived features, the mechanism is not engine-specific, and the gains survive memorization, sibling, and account audits while preserving Maia-3 at 1000-2500 Elo. Because search supervision is modular, Matilda naturally supports interchangeable search engines, demonstrated by replacing Stockfish with an AlphaZero-family engine.

Jason Carlson

Set Transformer Game-Playing Agents Transformer Chess

Ask PDF

Jul 13, 2026

Communicating Chess Strategies in Natural Language

Chess engines have long achieved superhuman playing strength. However, the underlying strategy behind their move suggestions is difficult for human players, even skilled ones, to comprehend. Motivated by this, we propose the task of chess strategy verbalization, which is to describe chess strategies in natural language. We design (i) a pipeline for verbalizing strategies and (ii) an evaluation framework for objective evaluation of generated strategy descriptions. Our experiments show that natural language is a promising and interpretable medium for communicating strategic information to both human and LLM players. We glean additional interesting insights, including (a) the importance of evaluating strategies beyond the main line, (b) the limitations of pure concept-based descriptions, and (c) the limitations of relying on LLMs rather than humans for evaluation.

Langyuan Cui, Chun Kai Ling, Hwee Tou Ng

Department of Computer Science National University of Singapore 13 Computing Drive, Singapore 117417

Explainable Artificial Intelligence Chess Text Generation

Ask PDF

Abstract

Game-Playing Agents

Explore similar work

Chess_db: A framework for working with large chess game datasets

Matilda: Engine-Agnostic Search with Human Policy Guidance

Communicating Chess Strategies in Natural Language