Does Outcome-Value Supervision Add Decision-Relevant Information to a Compact Chess Policy? Controlled Two-Seed Evidence from a Value-Supervision Ablation

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 preprint asks whether outcome-value supervision contributes useful information to a compact, supervised chess network without tree search. 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 (seeds 123 and 124) comparing the same 96-channel, four-block recurrent convolutional recipe trained with value-loss weight 0.5 or 0. Both primary endpoints met their prespecified positive criteria on both seeds: (A) a linear probe of the shared trunk decodes a Stockfish-depth-8 evaluation more accurately for the value-supervised model (partial-Spearman contrast +0.1410 and +0.1699, threshold +0.05), and (B) after frozen-trunk value-head retraining, reranking log-likelihood gain is higher (+0.08089 and +0.08044, threshold +0.03). A secondary descriptive analysis found no consistent improvement in the policy itself (puzzle top-1 accuracy). The results do not establish human-like strategic understanding, general playing strength, or a mechanism. The data-order provenance and software versions of the original training runs are only partly recoverable, and the experimental artifacts are not publicly released; these limits are documented in the paper. Files: LaTeX source and compiled PDF (21 pages).