Does Outcome-Value Supervision Add Decision-Relevant Information to a Compact Chess Policy?

Abstract

A policy network trained on human games estimates which moves are likely to be played, whereas an outcome-value head estimates the expected game result from a position. This study asks whether outcome-value supervision contributes useful information to a compact, supervised chess network without Monte Carlo tree search (MCTS). The policy proposes four candidate moves; a value head scores the resulting child positions and reranks this fixed candidate set.

We report a prespecified, two-seed controlled ablation comparing the same 96-channel, four-block recurrent convolutional recipe trained with value-loss weight 0.5 (V+) or 0 (V0). Both primary endpoints met their prespecified positive criteria on both seeds. For endpoint A, a linear probe of the shared trunk's Stockfish-depth-8 evaluation signal produced a partial-Spearman contrast of +0.1410 for seed 123 (95% grouped-bootstrap interval [+0.1166,+0.1644]) and +0.1699 for seed 124 ([+0.1463,+0.1949]), both above the +0.05 threshold. For endpoint B, the difference in reranking log-likelihood gain after frozen-trunk value-head retraining was +0.08089 for seed 123 (95% interval [+0.06604,+0.09527]) and +0.08044 for seed 124 ([+0.06659,+0.09400]), both above the +0.03 threshold and passing the prespecified consistency checks.

The secondary policy-quality analysis did not show a consistent improvement in the policy itself: puzzle top-1 accuracy was nearly tied for seed 123 and favoured V0 for seed 124, whereas held-out training-log accuracy showed small differences in the opposite directions. These results support a protocol-specific effect of outcome-value supervision on two operational measures of representation and downstream reranking utility. They do not establish human-like strategic understanding, improved playing strength in general, or a mechanism beyond the measured endpoints. A reduced-prefix data-order replay over the first 100,000 games completed successfully in two independent CPU processes: the pre-shuffle, post-shuffle batch-sequence, and held-out fingerprints matched exactly. This validates repeatability of the current loader over the tested prefix, but cannot reconstruct the exact batch sequence used by the historical training runs.

Keywords: chess; policy-value networks; supervised imitation; auxiliary objectives; move reranking; representation probing; causal ablation; interpretability.

1. Introduction

Chess engines based on neural networks commonly combine policy and value predictions. In systems such as AlphaZero, the two predictions interact with tree search, and the network is optimized using both policy and value targets [1]. The presence of a value head in such systems does not, by itself, establish what outcome supervision contributes to a compact, supervised policy network when search is absent.

RDTChess is a from-scratch chess network trained on human games. Its inference policy does not rely on MCTS or an opening book. The experiment studied here asks a narrow question: if the policy's highest-ranked moves are held fixed as the candidate set, can a value head trained on game outcomes help select among those candidates? A deeper question is whether any such advantage is caused by the value loss changing the shared representation, rather than merely reflecting a useful score attached to each candidate.

The work proceeded in three stages. First, policy-plus-value reranking was tested behaviourally on puzzles and games. Second, a predictive analysis asked whether value outputs contained information about a Stockfish-defined target beyond flexible functions of policy outputs. Third, a controlled ablation compared models trained with and without value supervision while holding the remaining recipe fixed as far as the implementation permitted. Results from this third stage are reported below.

1.1 Contributions

This paper contributes a controlled study of one specific question: what changes when outcome-value supervision is added to a compact supervised chess network while the rest of the training recipe is held fixed as far as the implementation permits? Its contributions are:

1. A paired V+/V0 ablation over two initialization seeds, evaluated using prespecified endpoint criteria rather than puzzle performance alone.

2. A linear-probe endpoint testing how well a Stockfish-defined positional evaluation is decodable from the shared trunk, with game/position-linked groups kept together.

3. A frozen-trunk head-retraining endpoint testing whether the representations support more useful reranking of a fixed set of policy candidates under common downstream training conditions.

4. A separate descriptive policy-quality analysis showing that positive A/B outcomes do not imply an improvement in the policy head's top-1 puzzle accuracy.

5. A documented audit trail of data exclusions, checkpoint hashes, training logs, analysis outputs, and unresolved reproducibility limitations.

The intended contribution is empirical and protocol-specific. It is not a claim that outcome-value heads, policy-value learning, linear probes, or move reranking are new in themselves. The novelty of this exact combination of intervention and endpoints requires a broader literature review before any priority claim.

2. Related work

AlphaZero established a widely used policy-value architecture trained by self-play and integrated with MCTS [1]. The present work is different in setting: the policy is trained by supervised imitation of human moves, and the proposed reranking step evaluates only four policy candidates rather than searching a game tree.

Wang et al. studied alternative combinations of policy and value losses in AlphaZero-like self-play systems on smaller board games [2]. Their results show that the choice of objective can affect playing strength, but their setting differs from the supervised, no-MCTS reranking studied here. More generally, work on shared policy-value representations has argued that sharing one representation between the policy and the value function can cause overfitting and hurt generalization, which motivates decoupled policy and value networks [3].

For interpretability, McGrath et al. investigated correspondences between human chess concepts and AlphaZero's internal representations [4]. The linear-probe component proposed here is related in purpose, since it also relates internal representations to chess-relevant targets, but it asks a different comparative question: whether adding outcome-value supervision changes how well a specified evaluation target is decodable. Recent work such as Matilda combines human policy guidance with search-derived candidate information in a reranking architecture [5]; the current RDTChess intervention instead studies a learned outcome-value head and does not use search-derived candidate evaluations during training.

Schut et al. extracted machine-specific chess concepts from AlphaZero and tested whether selected concepts could be taught to grandmasters [6]. Mészáros et al. evaluated out-of-distribution rule-based and strategic reasoning in a 270-million-parameter chess transformer, reporting stronger rule generalization than strategic generalization [7]. These studies investigate concept discovery and out-of-distribution reasoning, respectively. RDTChess asks a more controlled, narrower question: whether changing only the outcome-value loss weight in a compact supervised model affects a defined evaluation probe and downstream reranking utility. Neither the existing literature nor the present experiments justify equating these operational measurements with human-like understanding.

These similarities place RDTChess within a broader literature on policy-value learning, auxiliary objectives, chess representation analysis, and move reranking. They do not establish novelty by themselves. Any claim that the particular combination studied here is novel would require a more exhaustive literature review.

3. System and methods

3.1 RDTChess and the reranking procedure

RDTChess is a compact residual convolutional chess network. The public repository describes the 128-channel, 8-block model as having 10.92 million parameters at the time of access. The controlled ablation in this study uses a separate experimental recipe with 96 channels, four blocks, recurrence setting R=2, a gate, and squeeze-and-excitation (SE) modules [10]. Its parameter count is therefore distinct from that of the public 10.92-million-parameter model, and the ablation should not be mistaken for an exact retraining of the public 128-by-8 checkpoint.

For a position s, the policy produces a distribution πθ(a∣s) over legal moves. Let 𝒞4(s) be the four moves with the highest policy scores. Applying candidate aj gives child position sj=T(s,aj). The value head estimates the expected game outcome from its encoded player-to-move perspective; scores are converted to the root player's perspective before ranking. The reranked move is

where qθ(sj) is the value estimate expressed from the root player's perspective. This step evaluates a fixed candidate set; it is not MCTS and does not expand a search tree.

3.2 Training objective, intervention, and implementation

Let ℒπ be the supervised cross-entropy policy loss over legal moves and let ℒv be the mean-squared error between the value output and the game-result target z∈{−1,0,+1}. The objective is

The intervention changes only the value-loss weight:

With λv=0, the value loss contributes no gradient. Because the losses share a trunk, changes to shared-trunk gradients and representations are part of the intervention.

The controlled models use a 96-channel, four-block convolutional trunk, recurrence R=2, a gate, and squeeze-and-excitation modules. This is a separate experimental recipe, not a retraining of the public 128-channel/eight-block checkpoint. The training stream consists of two million selected games from the public Lichess database [8], with minimum Elo of 1800 for both players, a recorded game result (1–0, 0–1, or 1/2–1/2), and all half-moves eligible (skip_plies=0). After exclusions, 144,309,831 positions remain. The exclusion removes all positions from 13,684 games used to construct the original 20,000-position evaluation bank, plus 226 additional occurrences of puzzle-position keys outside those games; two puzzle-key occurrences inside the excluded games are already counted in the first removal. The same exclusions and data-order seeds are used in both arms. The corpus file was lichess_2026-07_5g.pgn.zst (5,368,709,120 bytes; SHA-256 af1697b2ef83831889a7440dbe3f54810587474d2c1eeb72af59422929b34ebd), but this checksum was first computed after training, not recorded at training start.

For V+ and V0, training used AdamW (fused=True), learning rate 10−3, weight decay 10−4, batch size 512, one epoch, and gradient-norm clipping with configured value 1.0. AdamW's β1=0.9, β2=0.999, and ϵ=10−8 were PyTorch defaults, not explicitly passed arguments; AMSGrad was disabled. No learning-rate scheduler was used. Training used float32 without automatic mixed precision. Seeds 123 and 124 set Python, NumPy, PyTorch, and CUDA RNGs; the data-order seeds remained fixed across arms. cuDNN deterministic mode was enabled and benchmarking disabled. The checkpoints report 282,126 optimizer steps each. This is 0.1% more than the simple quotient of retained positions by batch size; the difference is reproduced exactly by batching each 400,000-position block separately: 360 full blocks of ⌈400000/512⌉=782 batches plus a final block of 309,831 positions (⌈309831/512⌉=606 batches) give 360×782+606=282,126 (derived from the block size, the batch size and the retained-position count; the complete training-time sequence of batches was not saved for direct verification).

The corpus file's SHA-256 was calculated after training rather than recorded at training start. Its filesystem modification date predates the runs, which is supporting evidence but not proof that the exact bytes were unchanged. Software versions reported in this paper were measured during the subsequent audit (Python 3.12.4, PyTorch 2.3.1+cu121, CUDA 12.1, cuDNN 8907, NumPy 1.26.4, SciPy 1.13.1 and python-chess 1.11.2); the precise patch versions present during training were not logged. These provenance limits are not hidden by the reported positive endpoint results. Figure 1 summarizes the design.

Figure 1. Controlled two-seed design. The primary endpoints test different properties: A tests decodability of a Stockfish-defined target; B tests utility for reranking at a fixed policy/candidate set. Policy quality is reported separately and does not enter the primary decision rule.

Details of the boxes in Figure 1:

• Shared training protocol: 2,000,000 selected games; fixed exclusions and data order; one epoch; batch size 512.

• V+: λv=0.5, policy loss + value loss. V0: λv=0, policy loss only.

• Paired comparison across seeds 123 and 124: same training recipe and exclusions; compare V+ against V0 under prespecified rules.

• Endpoint A: representation probe: decode Stockfish-depth-8 evaluation from the shared trunk. Threshold: Cs≥+0.05.

• Endpoint B: downstream reranking: freeze trunk, retrain value head; fixed policy candidates. Threshold: Ds≥+0.03.

• Secondary descriptive analysis: policy quality: held-out accuracy and paired puzzle top-1 accuracy; no decision threshold.

3.3 Reduced-prefix data-order reproducibility audit

After both training seeds and primary endpoints had completed, the current data reader was replayed over the first 100,000 games (5% of the two-million-game training stream) in two independent CPU processes running concurrently, without using the GPU. Each replay used the existing reader and filtering functions, game-level exclusions, the held-out draw, 400,000-position blocks, the fixed shuffle seed, and batch size 512. Both processes exited with code 0; wall-clock times were 460.3 and 460.1 seconds under shared CPU load.

The two replays produced identical position counts and fingerprints. After filtering, both reported 6,159,402 positions, 16 blocks (15 blocks of 400,000 and a final block of 159,402), 12,042 batches, and a 20,000-position held-out set. The fingerprints were:

• Before shuffling: 59b85a4cfba1d0f199bd9ebdc152db2359d5a1992c7a35ba0ebffe504c61ebad.

• After shuffling, at batch-sequence level: 08028081de7eb700f9b8b2c0708e81e317c826c6667b8238e9d966540b65e080.

• Held-out set: c96894291e02ef0f2eba809bbf9ee05e082487dd73295db840d18443db42a44a.

The prefix replay reported 1,091,604 positions removed by the evaluation-bank game exclusion (including two also covered by puzzle-key filtering), followed by 12 additional puzzle-key-only exclusions in the prefix. Counts of positions and batches were internally consistent: the block sizes sum to the reported position count, and summing the ceiling of each block size divided by 512 gives 12,042 batches. The full two-million-game training replay had recorded 1,091,604 positions removed by bank-game exclusion and 226 additional puzzle-key occurrences outside those games.

This is a reproducibility check of the current loader on the current corpus and environment, not a direct audit of the historical training inputs: the four training runs did not record their full batch-sequence fingerprints at execution time. The match between the two independent replays is positive evidence that the present data pipeline is deterministic over the tested prefix, but does not prove that every historical run consumed byte-for-byte identical batches across the entire two-million-game stream. No training or GPU evaluation was performed for this check.

3.4 Earlier behavioural evaluation

The preliminary behavioural tests used a fixed set of 23,845 puzzle positions sampled from the Lichess puzzle database [8]. Across five tested model variants, the reported puzzle metric improved by 1.1 to 6.8 percentage points under value reranking. In separate games against a minimax opponent at depth 2, each arm comprised 500 games. The reported score metric changed as follows:

The interval reported for model E included zero, so its change was inconclusive. These games are a useful behavioural check, but they do not replace a controlled analysis of the training objective. A depth-2 minimax opponent is also a limited benchmark and these numbers should not be interpreted as general engine-strength ratings.

3.5 Exploratory predictive-complementarity analysis

A separate analysis asked whether the value output provides information about which of the four policy candidates Stockfish at depth 8 ranks best. The target was the best of the four candidates according to Stockfish, requiring a margin of at least 100 centipawns; positions with mate-like scores (|score|≥9000) were excluded from the primary target. The dataset was split into discovery and confirmation groups, with positions from the same game or repeated normalized positions kept together. The confirmation subset contained 6,759 eligible positions.

The policy-only statistical model included (i) the log-probability of each candidate under the policy, centered within position and divided by a frozen global scale, (ii) its squared term centered within position, (iii) centered indicators for policy ranks 2–4, (iv) standardized confidence of the top policy move, and (v) a confidence-by-log-probability interaction. The augmented model added the candidate value feature, centered within position and scaled globally. Let LL(M) denote the mean log-likelihood assigned to the Stockfish-selected candidate by model M. The reported contrast is

Confidence intervals were estimated by grouped bootstrap.

The primary exploratory result was ΔLL=+0.12670 nat per position, with a 95% interval [+0.11455,+0.13902] and a positive value coefficient (βv=+0.917). A control comparing a flexible policy-only specification against its value-augmented counterpart gave +0.12618 nat per position, with interval [+0.11404,+0.13855]. These are predictive associations; they are not effects of the value loss and do not by themselves establish a strategic mechanism.

3.6 Important data-overlap audit

After the exploratory analysis, a full-corpus audit found that 110 distinct puzzle-position keys (out of 23,763 distinct puzzle keys) appeared in the training corpus, with 226 occurrences outside the original benchmark-game exclusions. No source game for the puzzle set was found in the two-million-game training corpus, but the repeated positions created a separate, exact-position overlap. The exploratory +0.12670 result was not recomputed on a fully position-disjoint corpus in the records summarized here. It must therefore remain labelled exploratory and potentially affected by exposure to repeated positions.

The controlled training recipe was subsequently amended to remove both (i) all positions from the 13,684 games used to construct the 20,000-entry evaluation bank, and (ii) all occurrences of the fixed puzzle-position keys. The union removes 1,091,830 positions and leaves 144,309,831 training positions, corresponding to approximately 281,855.1 batches of 512. Of the additional puzzle-key occurrences, 226 occur outside the benchmark-game exclusion; two occurrences inside benchmark games are already covered by the first exclusion. The full-corpus replay validated these counts and the exclusion logic. This cleaning rule is applied identically to V+ and V0.

3.7 Endpoint A: linear probe of the shared trunk

The reduced evaluation bank contains 15,722 positions in 11,357 connected components, split into 12,578 training and 3,144 test positions. The probe uses the flattened final shared-trunk activation (6,144 features) to predict a Stockfish depth-8 evaluation target (the binary was identified post hoc as Stockfish 18; see the provenance limitation below and the official release note [9]), transformed as tanh(cp/400), with centipawn scores clipped at ±10,000. A ridge model standardizes features using the training partition, chooses regularization by exact leave-one-out cross-validation over 17 values from 0.1 to 107, and is evaluated on component-separated held-out positions. The statistic is a partial Spearman correlation controlling for material. The same component split and grouped-bootstrap resamples are used for V+ and V0; confidence intervals use 1,000 paired group-bootstrap replicates.

For seed s, let p+,s and p0,s be the partial Spearman correlations for V+ and V0. The contrast is

The prespecified per-seed positive criterion requires Cs≥+0.05 and a confidence-interval lower bound above zero; the cross-seed stability rule must also hold. These thresholds are study-specific decision rules, not universal standards of practical importance. The primary analysis uses the reduced, position-filtered bank, not the original bank with overlap.

3.8 Endpoint B: frozen-trunk value-head retraining

Endpoint B asks whether the trunk trained with value supervision supports more useful value-based reranking after a value head is retrained under common downstream conditions. A fixed reference policy (model A) supplies the same roots, four candidates, policy logits, and child positions for paired V+/V0 evaluation. The exact training-code revision for the reference policy is not fully verified, which limits reproducibility.

The value heads are retrained on game-result labels with each trunk frozen. The dataset comprises 2,251,834 positions from later games (indices 2,000,000–2,039,999) after removing positions overlapping the training corpus and puzzle positions; 228,648 positions form validation data. Components are partitioned by game/normalized position and shared across arms. Stockfish evaluations are not used as head-training labels. The head architecture is a 1×1 convolution (96 to 16 channels, no bias), BatchNorm, fully connected layers 1024 to 128 to 1, and tanh.

Head fitting uses Adam, not AdamW, with batch size 512, at most three epochs, and learning-rate candidates 3×10−4 and 10−3. Each arm selects its own learning-rate/epoch combination using validation MSE only. There is no clipping, scheduler, or weight decay in this head-fitting code; Adam's betas (0.9,0.999) and ϵ=10−8 are defaults. The head initialization and data permutation use seed 123 for every trunk, including seed-124 trunks. The selected setting was 10−3 / epoch 1 for both arms at seed 123, and 3×10−4 / epoch 1 for both arms at seed 124.

For each arm the reranking log-likelihood change is ΔLL. The main contrast is

The prespecified positive criterion required Ds≥+0.03, a grouped-bootstrap lower bound above zero, cross-seed stability, and a consistency check: V+'s reranking ΔLL with the retrained head must be within 0.03 of the original V+ head's value. Bootstrap intervals use 10,000 grouped resamples. The threshold and grid were set before examining the controlled outcomes.

3.9 Secondary descriptive analysis: policy quality (prespecified amendment 50.13)

The primary endpoints A and B assess, respectively, whether Stockfish-evaluation information is linearly decodable from the shared trunk and whether a frozen trunk supports value-based reranking after a value head is retrained. They do not directly measure whether value supervision changes the quality of the policy head itself. Because the policy and value heads share a trunk, the value loss could act as an auxiliary task and alter policy performance. A separate policy-quality analysis was therefore specified before any controlled training runs. It is secondary and descriptive, has no decision threshold, does not enter the stopping rule, and cannot replace or change the interpretation of A or B.

For each seed separately, two quantities were reported. First, the final held-out policy accuracy recorded in the training log was extracted for V+ and V0, together with their difference in percentage points. This log value is rounded to 0.1 percentage point and has no confidence interval in the available log format. Second, both final checkpoints were evaluated at forced recurrence R=2 on the same 23,845 puzzle entries, using the same legal-move masking and top-1 policy rule as the project puzzle evaluator. The puzzle metric is agreement with the puzzle's designated first move, not a claim that every alternative move is strategically wrong.

For the paired puzzle comparison, the report includes each model's top-1 accuracy, the V+ minus V0 difference in percentage points, and the two discordant counts: entries solved by V+ but not V0, and entries solved by V0 but not V+. An exact McNemar p-value is shown as descriptive only; the classical test does not account for dependence among repeated instances of the same position key and has no decision-making role. A 95% interval for the accuracy difference was estimated by a paired bootstrap clustered by distinct normalized puzzle-position key, using 10,000 resamples and the prespecified random seed 123 (23,763 distinct keys among the 23,845 puzzle entries).

The held-out training-log accuracy and the puzzle accuracy concern different evaluation distributions and targets. Their absolute values are not compared to each other; only V+ versus V0 differences within each evaluation are described. Concordant differences favouring V+ across both evaluation types and both seeds would have been compatible with a policy-quality improvement associated with value supervision. The held-out log differences had opposite signs across the two seeds, whereas both puzzle differences were non-positive (the seed-123 interval includes zero); no metric was selected post hoc as preferred. Small or uncertain differences are not interpreted as evidence that no effect exists. Both seeds were analyzed, but the results do not establish cross-evaluation concordance. This analysis is not a confirmatory causal endpoint and does not establish a mechanism.

3.10 Seeds and decision rule

The paired ablation was executed for seeds 123 and 124. For each endpoint, the per-seed effect threshold, interval criterion and cross-seed stability rule were applied as specified in the experiment record (protocol entry 50.5). The protocol reported the following stability statistics: SA=0.0342<0.0777 and SB=0.0025<0.0403. These are treated as passing under the preregistered rule. Each endpoint is judged independently; secondary policy-quality results cannot substitute for A or B or alter their criteria.

4. Results of the controlled two-seed ablation

All three analysis jobs for seed 124 exited successfully; no additional training was run for this analysis. The A and B primary endpoints were evaluated once per seed under the committed protocol. Values below are taken from the saved analysis JSON outputs and project record. The commits containing these outputs were pushed on 18 October 2026 to a private remote repository (head 4fe7ae96), so they are not publicly retrievable.

4.1 Endpoint A: evaluation information decodability

V+ had higher partial Spearman correlation with the Stockfish evaluation target than V0 on both seeds. The per-seed contrasts exceeded the prespecified threshold of +0.05, and their paired grouped-bootstrap confidence intervals were entirely above zero.

Table 1. Endpoint A. Partial Spearman correlation controlling for material and V+ minus V0 contrast Cs.

The mean of the two per-seed contrasts is +0.1554. Both per-seed contrasts meet the threshold and the preregistered stability criterion passed. The chosen ridge penalty was 104 for all four arms except V0/124, where it was approximately 3.16×103; no selected penalty was at the edge of the search grid. The result indicates that, in this probe setup, the shared representation from V+ makes the specified Stockfish evaluation target more linearly recoverable than the corresponding V0 representation. It does not establish that the representation is a faithful human-like strategic explanation.

4.2 Endpoint B: downstream reranking with retrained heads

The V+ trunk also supported a larger reranking log-likelihood improvement than the V0 trunk after each trunk's value head was retrained on the same type of outcome labels. The selected hyperparameter differed between seeds as permitted by the prespecified validation-only selection rule. The consistency check passed for both seeds.

Table 2. Endpoint B. Reranking log-likelihood gains and contrast Ds after frozen-trunk value-head retraining.

The mean contrast is +0.0807. Both contrasts exceed the prespecified +0.03 threshold; both lower confidence limits exceed zero; and the cross-seed stability rule passed. The original-head versus retrained-head consistency differences for V+ were 0.00176 at seed 123 and 0.01566 at seed 124, both below 0.03. Across the six prespecified learning-rate/epoch combinations, all seed-124 point estimates were positive, but one (10−3 / epoch 3: D=+0.02739) was below +0.03; this robustness table is descriptive and does not replace the primary validation-selected comparison.

The result is evidence that, under this evaluation procedure, the V+ trunk supports more useful downstream value-based reranking among a fixed set of policy candidates than the V0 trunk. Because the reference policy and candidate positions are fixed in this endpoint, the contrast is not explained by different policy candidate sets within this comparison. The result does not show that every resulting move is objectively better, nor does it establish general playing-strength improvement.

4.3 Secondary endpoint: policy quality (descriptive only)

The secondary policy-quality results were mixed and did not show a consistent advantage for V+.

Table 3. Prespecified descriptive policy-quality comparisons. Puzzle uncertainty is grouped by distinct position key; held-out log accuracies are rounded to 0.1 percentage point.

At seed 124, the paired discordant counts were 1,494 puzzle entries solved only by V+ and 1,635 solved only by V0; the exact McNemar p-value was 0.0123, shown descriptively only and not used for any decision. At seed 123, the puzzle interval includes zero. The held-out log differences have opposite signs across seeds and are rounded without confidence intervals. These measures concern different evaluation distributions; the appropriate conclusion is that the secondary analysis does not support a consistent policy-quality improvement for V+ in these data. This does not negate the A/B results, which measure different properties. Figure 2 plots both comparisons.

Figure 2. Secondary policy-quality results. The puzzle comparison uses paired bootstrap intervals clustered by distinct position key. The training-log differences are rounded to 0.1 percentage point and have no available confidence intervals; the two evaluation types are shown separately because they measure different distributions.

The plot itself is only in the PDF; every plotted value is in Table 3 above. In the plot, the seed-123 puzzle interval includes zero, the seed-124 puzzle interval is below zero, and the two held-out log differences are −0.2 (seed 123) and +0.1 (seed 124). This secondary metric does not show a consistent V+ improvement.

4.4 Combined decision

Table 4. Primary endpoint decision under protocol entry 50.5.

The two primary endpoint decisions are positive under the prespecified rules. This supports two operational findings: outcome-value supervision increased the decodability of the specified Stockfish evaluation target from the shared trunk, and it increased the utility of the trunk for downstream reranking with retrained value heads at a fixed policy/candidate set. The secondary policy-quality findings are not positive and remain descriptive. Figure 3 plots the per-seed contrasts of both primary endpoints.

Figure 3. Primary endpoint contrasts for both seeds. Points are observed contrasts and horizontal whiskers are the prespecified grouped-bootstrap 95% intervals. Dashed orange lines mark the per-seed thresholds. Both endpoints passed the predefined positive and stability criteria; the plot summarizes the per-seed evidence without implying general chess strength or human-like understanding.

The plot itself is only in the PDF; every plotted value is in Tables 1 and 2 above. In panel A, the contrasts are +0.1410 (seed 123) and +0.1699 (seed 124) against the +0.05 threshold. In panel B, they are +0.0809 (seed 123) and +0.0804 (seed 124) against the +0.03 threshold.

5. Interpretation

The primary results support a bounded conclusion. Under the training and evaluation recipe studied here, adding outcome-value supervision (V+, λv=0.5) rather than setting its weight to zero (V0) was associated with, and under the controlled two-seed intervention provides evidence for, two specific changes: (i) the Stockfish-depth-8 evaluation target was more decodable from the shared trunk by the prespecified linear probe, and (ii) the trunk supported greater reranking log-likelihood gain after value-head retraining while policy proposals and candidate positions were held fixed in that downstream test. Both effects passed their respective prespecified per-seed and stability rules.

The policy-quality analysis helps delimit the claim. V+ did not improve top-1 puzzle accuracy consistently: the difference was near zero at seed 123 and favoured V0 at seed 124, while the held-out log accuracy differences had opposite signs. Thus the main finding is not that value supervision generally makes the policy choose better first moves. It is that, in this experiment, value supervision changed what could be decoded from the trunk and how useful that trunk was for a separate reranking task.

This remains distinct from the question of whether the network understands why a chess move is good in a human-like, causal or strategic sense. A probe demonstrates recoverability of a specified target, not that the network explicitly represents a human explanation. A reranking score can exploit learned correlations without robustly reasoning about a sacrifice or a multi-move plan. The next test for such a stronger claim would manipulate positions in controlled pairs—for example, remove the tactical condition that makes a sacrifice sound and test whether the model changes its preference appropriately—on positions not used for training or hyperparameter selection.

The Stockfish-depth-8 target is an operational reference, not ground truth. The engine binary identified as Stockfish 18 during the audit was not explicitly version-logged at the time the evaluation caches were produced. File creation dates make it likely that the caches came from that binary, but do not prove it. This uncertainty, as well as incomplete training-time data-order provenance, constrains the strength of the causal and reproducibility claims.

6. Limitations and provenance

• The study has only two training seeds. Per-seed bootstrap intervals quantify uncertainty over grouped evaluation positions under the implemented procedure; they do not estimate the full distribution of independent training runs.

• The training corpus hash was calculated after the runs rather than captured before training. The file modification time is consistent with the reported corpus being present earlier, but is not conclusive evidence that its bytes never changed. The file size is exactly 5 GiB (5×10243 bytes), which suggests a size-limited extract of the monthly archive; how the file was produced was not recorded.

• The complete training-time sequence of batches was not saved. A reduced 100,000-game prefix replay completed successfully in two independent CPU processes and matched exactly on the reported fingerprints and counts. Even this match validates only the current reader on the current file and environment; it does not prove the exact batches seen during all four historical training runs.

• Each checkpoint records 282,126 optimizer steps, about 0.1% more than the simple position-count/batch-size quotient. The difference is reproduced exactly by per-block batching (360×782+606=282,126), but the original training-time batch sequence is unavailable for direct verification.

• The exact policy-training code revision for the fixed reference model in endpoint B is not fully verified. The reference checkpoint hash and training record are available, but full lineage remains incomplete.

• The Stockfish binary returned id name Stockfish 18 during a post-hoc UCI check. The binary file was created on disk on 24 August 2026, before both evaluation caches were created (14 October 2026), so the caches are likely, but not proven, to have been produced by this same binary; the engine version was not stored in the original cache-generation logs.

• Training-time patch versions of Python and PyTorch were not logged. Versions reported in Methods were measured during the post-hoc audit, not during each training run. The effective gradient-clipping value of 1.0 is read from configuration code rather than a per-run log line.

• In V0, the value loss contributes a zero gradient to the value head, while AdamW can still apply decoupled weight decay to parameters whose gradient tensor is zero. This effect on the unused V0 value head was not separately quantified; endpoint A concerns the shared trunk and endpoint B retrains its value heads under common downstream conditions.

• The head fitting for endpoint B uses seed 123 for head initialization and data order even for seed-124 trunks. This is shared across V+ and V0 and follows the committed script, but it means endpoint B's two seed rows are not independently reseeded in every downstream component.

• The original exploratory +0.12670 predictive-complementarity analysis has exact-position overlap with 110 puzzle keys (226 occurrences) in the historical training stream. It remains exploratory and is not used as confirmatory evidence for the controlled results.

• The secondary policy-quality measure uses a fixed puzzle set with 23,845 entries and 23,763 distinct position keys. Its paired intervals cluster by key; the McNemar result is descriptive only. The puzzle target names a designated first move and does not exhaust all strategically acceptable choices.

• Exact normalized-position exclusion does not rule out broader strategic or near-position similarity leakage. Stockfish depth 8 can also misevaluate positions, especially tactically.

• The experimental artifacts are not publicly released. Commits containing the study logs/JSON and notebook updates were pushed on 18 October 2026 to a private remote repository that is not intended to become public; training checkpoints are not included in the repository. Readers therefore cannot retrieve the artifacts, and the preregistered protocol (protocol entry 50.5) and the notebook, whose commit timestamps are the only evidence of prespecification, cannot be checked independently.

7. Conclusion

In a compact supervised chess network without MCTS, a prespecified two-seed intervention on the value-loss weight produced positive results on both primary endpoints. Relative to V0, V+ yielded higher decodability of a Stockfish-depth-8 evaluation target from the shared trunk and greater downstream reranking log-likelihood gain after frozen-trunk value-head retraining, with both seed-specific contrasts meeting the decision thresholds and the predefined stability rules. These findings make a specific empirical contribution: they connect a controlled change in the training objective to both a representation-probing result and a downstream evaluation result, rather than relying only on playing strength or policy accuracy.

The secondary policy-quality analysis did not show a consistent improvement in the policy head. The results also do not demonstrate human-like strategic understanding, a specific internal mechanism, or general engine-strength gains. Those claims require further work, especially counterfactual tests that modify the causal conditions underlying a sacrifice or strategic plan. Reproducibility is also limited by incomplete training-time data-order and software provenance. The reduced-prefix replay is positive evidence for repeatability of the current reader, but the historical training-time batch sequence was not stored and cannot be reconstructed with certainty from this test alone. The study should therefore be read as controlled but bounded evidence about outcome-value supervision under one compact-network recipe, not as a general claim about machine understanding of chess.

Appendix A. Full experimental and reproducibility record

This appendix preserves the implementation details, raw values, model identifiers, audit records, and unresolved inconsistencies available at manuscript preparation. The labels code, checkpoint, journal, and derived distinguish the evidence source where relevant. Values described as post-hoc measurements were not recorded at the time of training.

A.1 Training corpus, selection, and exclusion ledger

The input archive was lichess_2026-07_5g.pgn.zst, with reported size 5,368,709,120 bytes and post-hoc SHA-256 af1697b2ef83831889a7440dbe3f54810587474d2c1eeb72af59422929b34ebd. Its modification date was 10 August 2026, before the controlled runs, but no checksum was recorded at training start; the unchanged-file claim is therefore supported by a timestamp, not proven. The file size is exactly 5 GiB, which suggests a size-limited extract of the monthly archive (derived; the creation procedure was not recorded). The selection used 2,000,000 games, minimum Elo of both White and Black at least 1800, a recorded result (1–0, 0–1, or 1/2–1/2), and all half-moves (skip_plies=0).

The training-position stream before controlled exclusions contained 145,401,661 positions. The audit's total of 145,421,661 counts this stream plus the 20,000-position evaluation bank. Of the training positions, 1,091,604 positions were removed by excluding all positions from 13,684 games represented in the original evaluation bank. A further 226 puzzle-position-key occurrences outside those games were removed; two such occurrences inside excluded games were already counted. The union of removals is 1,091,830 positions, leaving 144,309,831 training positions. The quotient by batch size 512 is 281,855.1, while the checkpoints record 282,126 optimizer steps. The implementation batches each 400,000-position block separately, so partial final batches per block can increase the step count; however, the historical batch sequences were not saved for a direct full-stream reconstruction.

The original 20,000-entry evaluation bank was reduced to 15,722 entries after 4,278 entries (21.39%) were found to have at least one occurrence in training. The report records 3,118 affected distinct keys out of 18,839. The reduced bank contains 11,357 connected components and is split into 12,578 training and 3,144 test positions. Its opening-position proportion changed from 44.6% to 30.4% after filtering. The exclusion procedure was replayed on the full stream during the original audit, with no bank keys or retained head-data keys found among the final retained training positions.

The puzzle benchmark contains 23,845 entries and 23,763 distinct normalized position keys. The historical stream contained 110 of those keys (0.463% of distinct keys); the audit reports 226 occurrences outside the bank-game exclusions and two more occurrences inside excluded bank games. All occurrences matched by the controlled puzzle-key exclusion are removed from the new V+/V0 training stream. The older exploratory predictive-complementarity result (+0.12670 nat/position) was produced before this cleaned recipe and has exact-position overlap; it remains exploratory and is not used as confirmatory evidence here.

A.2 Exact main-model architecture and optimizer configuration

All four controlled models use 96 channels, four convolutional blocks, recurrence setting R=2, a gate, squeeze-and-excitation (SE), 19 input planes, and 4,096 policy actions. This controlled recipe is separate from the public 128-channel/eight-block model described in the repository. The common code snapshot recorded for the controlled recipe was commit e6b5dabb3955b40386033e87d732192e049341fb (reported clean at the audit). The training logs of the four runs record other commit identifiers (b60e3aea, 21119405, e71778ed, 05f6af60), which differ from e6b5dabb only by documentation and notebook commits: git diff of chess_rl/ and main.py between e6b5dabb and each of these four commits is empty.

The loss was the masked legal-move cross-entropy without label smoothing plus λv times mean-squared error between value prediction and the game outcome from the side-to-move perspective, z∈{−1,0,+1}. V+ used λv=0.5; V0 used λv=0. There was no moves-left head. Training used torch.optim.AdamW with fused=True, learning rate 10−3, weight decay 10−4, batch size 512, one epoch, float32, and no automatic mixed precision. Gradient-norm clipping was applied through clip_grad_norm_; the configured value was 1.0, inferred from configuration code rather than a per-run log line. No learning-rate scheduler was used. Betas (0.9,0.999), epsilon 10−8, and amsgrad=False were PyTorch defaults read from the checkpoint rather than explicitly supplied arguments.

Seed 123 or 124 set Python, NumPy, PyTorch, and CUDA RNGs, principally controlling weight initialization. Data-order RNGs remained fixed across arms: default_rng(0) for held-out selection and default_rng(1) for shuffle. The 20,000 held-out positions were sampled at 1%; the shuffle operated in 400,000-position blocks. cuDNN deterministic mode was enabled and benchmarking disabled. The historical reader used four worker processes; its full ordering during the original four runs was not fingerprinted. A zero value-loss weight means no gradient contribution from value MSE; AdamW can nevertheless apply decoupled weight decay to the V0 value-head parameters when their gradient tensor is zero. That effect was not quantified separately.

Table 5. Controlled checkpoint identifiers. Hashes identify the exact final files reported by the project logs.

A.3 Endpoint A: complete probe specification

The probe input is the final shared-trunk state at R=2, flattened to 6,144 dimensions, in evaluation mode and float32, processed in batches of 500. The target is Stockfish centipawn evaluation from the side-to-move perspective, clipped to ±10,000 cp and transformed as tanh(cp/400). Ridge features are standardized using the training split; standard deviations below 10−6 are replaced by 1. The Gram matrix is formed in float32 and converted to double for eigendecomposition. The penalty grid has 17 values 10k/2 for integer k=−2,…,14, spanning 0.1 to 107; exact leave-one-out is used on training components only. The test statistic is Spearman correlation partialling out material, with formula

The 80/20 split is by connected component using seed 123. Paired 95% intervals use 1,000 group-bootstrap resamples with seed 123 and common resamples across arms. The penalty was 104 for V+/123, V0/123, and V+/124; it was approximately 3.16×103 for V0/124. No selected penalty was at a grid edge.

Table 6. Endpoint A exact reported values and decision inputs.

The mean contrast is +0.155418. The protocol's stability statistic was SA=0.0342<0.0777, treated as passing by rule 50.5. Each per-seed contrast exceeds +0.05 and each lower interval bound is positive. A single seed-124 execution took about 1 min 45 s and returned code 0.

A.4 Endpoint B: head-retraining settings, fit values, and sensitivity grid

The frozen-trunk head dataset was drawn from games indexed 2,000,000–2,039,999 (40,000 indexed games, of which 39,961 yielded positions; 2,865,412 candidate positions). The audit reports 613,576 exact duplicate positions already seen in main training and four puzzle positions removed, leaving 2,251,834 retained positions; the two removal counts overlap by two positions (the union is 2,865,412−2,251,834=613,578, derived), so the final retained count from the journal is the authoritative total. The validation subset contains 228,648 positions. No puzzle-source games overlapped this head-data slice. Positions were partitioned 9/1 by connected component, using SHA-1 of the smallest game identifier modulo 10; groups remain shared across arms. The fixed policy reference (model A) supplies the same roots, top-four candidates, logits, and resulting child positions for V+ and V0. The exact policy-training source revision for that reference remains unverified.

Each trunk was frozen in evaluation mode at the final checkpoint and R=2. The retrained value head consists of a 1x1 convolution (96 to 16 channels, no bias), BatchNorm, linear layers 1024 to 128 to 1, and tanh. Heads were fitted to game-outcome labels only (not Stockfish scores), with Adam (not AdamW), batch size 512, at most three epochs, learning-rate candidates 3×10−4 and 10−3, float32, and seed 123 for both head initialization and lot permutation, including the seed-124 trunks. There was no clipping, scheduler, or weight decay in this code path; Adam's betas (0.9,0.999) and epsilon 10−8 were defaults. Each arm selected its setting independently using validation MSE only.

Table 7. Selected endpoint-B heads. The validation-only selection rule was specified in advance; different seeds selected different learning rates.

Table 8. Endpoint B: original-head versus retrained-head log-likelihood gains and coherence check.

The selected-head contrasts were D123=+0.080895 (95% interval [+0.066040, +0.095275]) and D124=+0.080440 ([+0.066588, +0.093996]); their mean is +0.0806675. The protocol reports SB=0.0025<0.0403, and the original-V+ versus retrained-V+ consistency differences were 0.00176 and 0.01566, both below 0.03. All main endpoint B fits passed the implementation checks. The full robustness grid is given in Table 9. A single seed-124 B analysis took 10 min 25 s and returned code 0.

Table 9. Endpoint B robustness grid: all 6 learning-rate/epoch settings for each seed. The prespecified validation-selected result above remains primary; this table is descriptive.

For seed 123, every point estimate in the grid exceeds +0.03, although two lower confidence limits are below +0.03. For seed 124, all six lower confidence limits are above zero, but the 10−3/epoch-3 point estimate (+0.02739) is below +0.03. These descriptive robustness results do not change the primary validation-selected rule or decision.

A.5 Secondary policy-quality results: full paired counts

The policy-quality metric is top-1 agreement with the puzzle database's designated first move at forced R=2. There are 23,845 entries and 23,763 distinct keys; uncertainty is a paired bootstrap clustered by key over 10,000 resamples. The exact McNemar test is descriptive and was not adjusted for dependence among repeated keys or used in the decision rule.

Table 10. Full secondary policy-quality values, including the paired discordance counts.

The held-out training-log accuracies were V+/123 50.0% versus V0/123 50.2% (difference -0.2 percentage points), and V+/124 50.4% versus V0/124 50.3% (difference +0.1 points). They are rounded to 0.1 percentage point and have no confidence intervals. These differences have opposite directions across the two evaluation types/seeds. The seed-124 policy evaluation took approximately 40 seconds and returned code 0. This secondary result does not support a consistent policy-quality improvement.

A.6 Stockfish configuration, cache lineage, and post-hoc software audit

Stockfish evaluations used depth 8, one thread, the engine's default hash-table size, and passed the position index through the engine's game parameter on each call. Mate scores were mapped to ±10,000 centipawns. The recorded executable path was stockfish-windows-x86-64-sse41-popcnt.exe, SHA-256 5004608CB62B40C83B18DDD1C58ABA4C0226FEDDD70CD3B135A6733E73530AE7. The source-tree NNUE network identifiers were nn-c288c895ea92 and nn-37f18f62d772. The post-hoc binary handshake returned id name Stockfish 18 and uciok; the first UCI attempt hung and was terminated, and a second attempt under a 20-second timeout succeeded. This version check only queried UCI identity and evaluated no positions. The binary was created on disk on 24 August 2026 (archive extraction) and reports compilation modification date 31 January 2026. Cache logs/value_vs_sf_evals.csv was created 14 October 2026 at 12:27:52 (208,144 bytes); logs/rerank_sf_evals.csv at 16:58:04 that day (2,680,022 bytes). The Stockfish 18 lineage for these caches is likely from dates and local binary presence, but not proven because the original cache-generation logs did not store the UCI identity. Neither cache is tracked by Git.

The hardware values below were measured during the post-hoc audit on 18 October 2026, not captured in each training run: NVIDIA GTX 1080 Ti, compute capability 6.1, 11,264 MiB VRAM; NVIDIA driver 581.29; AMD Ryzen 5 5600G (6 cores / 12 threads); 43.9 GB RAM as reported by Windows; Windows 10 build 10.0.19045. The environment measured during that audit was Python 3.12.4, PyTorch 2.3.1+cu121, CUDA 12.1, cuDNN 8907, NumPy 1.26.4, SciPy 1.13.1, and python-chess 1.11.2. Exact Python/PyTorch patch versions at the historical training time are unknown.

A.7 Execution and monitoring notes

All endpoint jobs completed with exit code 0 and were run once per seed. One earlier launch of endpoint A for the four models (seeds 123 and 124 in a single invocation) was stopped during data loading, before any value was produced, and was replaced by the two-model invocation described here; it yielded no result. For seed 124, the A job took about 1 min 45 s, B took 10 min 25 s, and the policy-quality evaluation about 40 s. The V0/124 training log records a 23,166-second run (6 h 26 min), from 09:44:20 to 16:10:26, ending after checkpoint serialization with 282,126 steps and no logged error lines. The average throughput recalculated from the checkpoint step count is about 12.18 steps/s. The reported maximum GPU temperature was 72 C (mean 70.6 C over 3,935 five-second samples) and maximum GPU memory use 3,407 MiB. Video playback was recorded from 14:32 to 15:21; throughput was 11.67 steps/s over that window versus about 12.1–12.5 otherwise, while maximum GPU memory was higher. This co-occurrence is recorded descriptively only and is not evidence that video playback caused the throughput difference. The other seed/run details should be taken from their training logs rather than inferred from this one run.

A.8 Analysis label defect and other known provenance issues

The analysis scripts contain hard-coded display labels high_A_123, high_B_123, and C_123 / D_123 in a rule-of-stopping output line. In the seed-124 JSON outputs, some of those labels carry seed-124 numeric values (for example, high_A_123 contains +0.1949, the upper confidence bound for C124). The labels are erroneous; the numerical result fields are read by mesure_regles by arm/contrast names from the etape1 and etape2 JSONs, not via high_A_123. An internal audit, performed with an AI coding assistant, found that the label defect did not affect the stage-2 decision, because the decision inputs were read by arm and contrast name (verified by recomputing both verdicts from the four JSON files). Scripts were left unchanged to avoid altering the preregistered analysis during outcome review.

Other unresolved items: (i) exact training-time software versions were not logged; (ii) complete full-stream batch fingerprints were not collected during the original runs; (iii) bit-for-bit determinism of a complete training run was not tested (only the first three batches were previously reproduced identically); (iv) the effective gradient clipping value is read from code/configuration, not each run log; (v) the V0 value head's AdamW weight-decay effect was not quantified; (vi) endpoint-B head initial weights were PyTorch defaults and were not separately archived; (vii) each endpoint measurement was executed once per seed; and (viii) the corpus checksum was only generated post hoc.

A.9 Reduced-prefix data-order replay: complete record

The replay script controles/controle_ordre_donnees.py was executed twice in separate CPU processes on the first 100,000 games (5% of the two-million-game stream), concurrently and without GPU use. The two executions took 460.3 and 460.1 seconds (exit code 0) under shared CPU load. Both reported 6,159,402 retained positions, 16 blocks (15 of 400,000 and a final block of 159,402), 12,042 batches of size 512, and a 20,000-position held-out set. The batch count equals the sum of ⌈block size/512⌉ over the recorded blocks.

Table 11. Data-order replay fingerprints: both independent processes returned identical values.

The two JSON outputs also have the same fingerprint values as notebook entry 50.20. Their file hashes and archive commits are:

• Script SHA-256: 42a9adb7623260e071419e210fbf5e1154e7d90a277234aef5e73aebba5f765c.

• JSON 1 SHA-256: f12fcf66892e24d1389f1da6320775158b18ed93a3ebf76632ac87e86421a297.

• JSON 2 SHA-256: 1ff82f88dcc4a546a96fde3bac716eb37601169076c6ee09c485e2524616b934.

• Script and JSON archive commit: 4fe7ae96c5100aa95d22d830a6e9df7479b19111.

• Notebook record commit: 253bd50a086f4334b597f993ef5c1839b58362ca.

Both JSONs contain all three matching BLAKE2b fingerprints. This replay demonstrates reproducibility of the current reader on the current first-100,000-game prefix and environment. It does not prove that the historical four training runs consumed exactly those sequences across all two million games.

A.10 Artifact history at the manuscript cutoff

All commits described below were pushed on 18 October 2026 to a private remote repository (origin, head 4fe7ae96); they are therefore not publicly retrievable. Commit identifiers are recorded for provenance, not as public retrieval links. The model checkpoints (.pt), complete training corpus, and Stockfish caches are not part of this source file.

• Controlled result/protocol notebook: 081535ac80d4dd6e3f8cecda199accff3d89deb1.

• Training logs for V+/123 and V0/123: 6feacc5b61de7301c42310e7a6b2b6648fa4a9b0.

• Training logs for V+/124 and V0/124: bb8a2c4c09a793e8095eebe1b5257714fe2a7199.

• Seed-124 endpoint JSON outputs: f7753cb3c48e5ba034d42a738dbd7ca8431e9989.

• Seed-123 and seed-124 endpoint logs / remaining result outputs: bd1c29a1b0bd8d2d8c652c641038f0ae3c4e5b00.

• Reduced-prefix replay results: 253bd50a086f4334b597f993ef5c1839b58362ca.

• Replay script and two JSONs: 4fe7ae96c5100aa95d22d830a6e9df7479b19111.

The result JSONs were read numerically by field / arm; the known hard-coded label defect above affects display names rather than the numeric fields used by the rule script. The corpus, exclusion lists, caches, and complete experiment artifacts are not publicly released.

Data and code availability

The RDTChess inference project is publicly described at github.com/RDTvlokip/RDTChess. The experimental artifacts are in a private remote repository (pushed on 18 October 2026) that is not intended to become public.

• Training and analysis records. The analysis JSONs and logs for endpoints A, B and the secondary policy-quality analysis were committed in three commits: 081535ac80d4dd6e3f8cecda199accff3d89deb1, f7753cb3c48e5ba034d42a738dbd7ca8431e9989 and bd1c29a1b0bd8d2d8c652c641038f0ae3c4e5b00.

• Data-order replay. The notebook entry is recorded in commit 253bd50a086f4334b597f993ef5c1839b58362ca. The control script and two JSON outputs are archived under controles/ in commit 4fe7ae96c5100aa95d22d830a6e9df7479b19111. Script: controles/controle_ordre_donnees.py, SHA-256 42a9adb7623260e071419e210fbf5e1154e7d90a277234aef5e73aebba5f765c. JSON 1 SHA-256 f12fcf66892e24d1389f1da6320775158b18ed93a3ebf76632ac87e86421a297. JSON 2 SHA-256 1ff82f88dcc4a546a96fde3bac716eb37601169076c6ee09c485e2524616b934.

• Remaining availability limits. Model checkpoints are not committed. Although the two independent current-reader replays matched, no historical batch fingerprints were saved during training, so the precise batch sequence consumed by the original runs cannot be proven retrospectively. The training corpus, exclusion lists, evaluation caches and full environment provenance are not publicly released.

The reduced-prefix replay script and its two result JSONs are also archived under controles/; their exact content and commit identifiers are documented in Appendix A. These commits were pushed to the private remote repository on 18 October 2026. The data-order replay covers only the first 100,000 games and cannot recreate the historical batch order over the full stream.

Acknowledgements

This work was conducted as an independent research project. No external funding or institutional affiliation is asserted.

AI assistance. AI assistants (Claude Code and ChatGPT) were used for code development, analysis scripts, audit checks, and the drafting and review of the manuscript. The author set the research question, approved the protocol and its decision rules, authorized each experimental run, and is responsible for the content.

Citation

If you use this work, please cite it as follows:

@misc{charlet_2026_rdtchess_value_ablation,
  author = {Charlet, Théo},
  title  = {Does Outcome-Value Supervision Add Decision-Relevant
            Information to a Compact Chess Policy? Controlled
            Two-Seed Evidence from a Value-Supervision Ablation},
  month  = oct,
  year   = 2026,
  doi    = {10.5281/zenodo.23454492},
  url    = {https://doi.org/10.5281/zenodo.23454492}
}

DOI: 10.5281/zenodo.23454492

github.com/RDTvlokip

References

1. D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419):1140–1144, 2018. doi.org/10.1126/science.aar6404

2. H. Wang, M. Emmerich, M. Preuss, and A. Plaat. Alternative Loss Functions in AlphaZero-like Self-play. In 2019 IEEE Symposium Series on Computational Intelligence (SSCI), pp. 155–162, 2019. doi.org/10.1109/SSCI44817.2019.9002814

3. R. Raileanu and R. Fergus. Decoupling Value and Policy for Generalization in Reinforcement Learning. In Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research 139:8787–8798, 2021. proceedings.mlr.press/v139/raileanu21a.html

4. T. McGrath, A. Kapishnikov, N. Tomašev, A. Pearce, M. Wattenberg, D. Hassabis, B. Kim, U. Paquet, and V. Kramnik. Acquisition of chess knowledge in AlphaZero. Proceedings of the National Academy of Sciences, 119(47):e2206625119, 2022. doi.org/10.1073/pnas.2206625119

5. J. Carlson. Matilda: Engine-Agnostic Search with Human Policy Guidance. arXiv preprint arXiv:2606.25176, 2026. arxiv.org/abs/2606.25176

6. L. Schut, N. Tomašev, T. McGrath, D. Hassabis, U. Paquet, and B. Kim. Bridging the human–AI knowledge gap through concept discovery and transfer in AlphaZero. Proceedings of the National Academy of Sciences, 122(13):e2406675122, 2025. doi.org/10.1073/pnas.2406675122

7. A. Mészáros, P. Reizinger, and F. Huszár. Out-of-Distribution Evaluation of Rule-Based and Strategic Reasoning in Chess Transformers. In Proceedings of the 43rd International Conference on Machine Learning, Proceedings of Machine Learning Research 306:88388–88406, 2026. proceedings.mlr.press/v306/meszaros26a.html

8. Lichess.org. Lichess open database: standard rated games and chess puzzles. Dataset portal, accessed 18 October 2026. database.lichess.org

9. The Stockfish developers. Stockfish 18. Official release announcement, 31 January 2026. stockfishchess.org/blog/2026/stockfish-18

10. RDTvlokip. RDTChess: Compact residual chess network and inference code. Software repository, accessed 18 October 2026. github.com/RDTvlokip/RDTChess