What question did this study set out to answer?

This research explores a chess engine that achieves high performance without traditional parameters and investigates neural network weights in a novel framework.

July 8, 2026Open Access

From One Axiom to Master-Level Chess — and the Law Inside Neural Networks: Attention, it turns out, was not all you need

Key Points

This research explores a chess engine that achieves high performance without traditional parameters and investigates neural network weights in a novel framework.
Chess engine evaluated board geometry with no tuned parameters or opening books, achieving elo-1900 performance in 3 days.
Conducted independent verification of queen-vs-rook positions and five-piece positions solved with zero errors against Syzygy.
Applied the same analytical methods to neural networks like CLIP and GPT-2 to find placement laws concentrated in specific layers.
Chess engine defeated Stockfish's Elo-1900 setting 62.5% of the time and held Elo-2100 to a draw in 50% of games.
Successfully solved 19,733,336 queen-vs-rook positions with 100% accuracy, along with over 1 billion five-piece positions.
Neural network weights revealed significant underlying placement laws, achieving 230x chance levels in token embeddings within GPT-2.

Abstract

Pre-paper v0.2 of the Smithian Fold Theory's computational program, en route to UnisonAI. Two connected results. First: a chess engine containing zero parameters — no piece values, no tuned weights, no opening book, no training — whose every evaluation is a count performed on the board's geometry and an exact rational share of the One, reached master-adjacent strength in three days of measured, refereed climbing: it beat Stockfish's Elo-1900 setting 62.5% and holds Elo-2100 to a draw in half its games. The same machinery solved queen-vs-rook exactly (19,733,336 legal positions, re-proven value-by-value in an independent clean-room implementation with zero disagreements), atop the campaign record of 1,092,871,108 five-piece positions solved at zero error against Syzygy. Solved chess value fields are dyadically smooth: 32 Walsh coefficients reconstruct 92.7–95.3% of all exact values. Second: pointing the identical pre-registered spectral instrument at trained neural networks, trained weights are found to carry placement-law in the fold's dyadic basis — unanimously (18/18 on validated released models), causally (untrained controls sit at exactly chance), located (the law concentrates in transformer expansion projections and token embeddings across three unrelated architectures — CLIP, GPT-2, Llama-3.1-8B — while contraction projections and attention sit at chance in aligned packings), reaching 230x chance in GPT-2's token embedding, and surviving 4-bit deployment quantization at every depth. Attention, it turns out, was not all you need. The paper states what is proved, what is registered as next, and how every number can be independently certified from the public repositories.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper