[Depreciated and replaced by V3] From One Axiom to Master-Level Chess — and the Law Inside Neural Networks: Attention, it turns out, was not all you need

Abstract

[Depreciated and replaced by V3] The application-specific clean rebuild has not yet been published; its authoritative theoretical boundary is now the governing V3 branch: After Turing: The Fold Machine - An Exact, Parameter-Free and Machine-Closed Derivation of Classical Computational Science from Smithian Fold Theory; From Fold to Consciousness: An Exact, Zero-Parameter and Machine-Closed Foundational Reconstruction of Consciousness and Cognitive Science from Smithian Fold Theory. The V3 source platform is https://github.com/MettaMazza/ernos-labs-sft-platform. The original DOI, concept DOI, version number and files are preserved for transparent historical provenance; this record must not be presented or cited as current V3 work. Abstract We present the science, the architecture, and the empirical validation of UnisonAI: a complete language and omni-model architecture (incorporating language, sight, hearing, speech, and video) in which every mechanism a modern AI model purchases with gradient training is replaced by a machine-verified law of the Smithian Fold Theory, utilizing zero trained parameters and zero tunable numbers end to end. Key Findings & Contributions The Spectral Law inside Weights: Using a pre-registered, self-certifying spectral instrument, we demonstrate that trained neural network weights carry a placement-law in the dyadic (Walsh) basis. Concentration is unanimous across the canonical LLM's entire knowledge-storage class (GPT-2, 13/13 tensors, 39/39 registered checks) and across image-diffusion/speech models (Stable Diffusion 1.5, SDXL, Kokoro-82M), concentrating in the transformer expansion projections (MLPs) and token embeddings while attention matrices sit at chance. DeepSeek-R1 Alignment: The placement-law is training-caused and scales up to DeepSeek-R1 at 671B, where the weights transform under the fold's transformation group exactly as solved game-theoretic value fields do. Counted Similarity Space: We show that semantic similarity is a counted object, where co-occurrence shares over held text reproduce semantic family structures with zero gradients and zero parameters. Standardized MMLU Performance: Evaluated on the canonical 128-item public MMLU test split under strict zero-parameter conditions, UnisonAI scored 9/128 (7.0%), demonstrating an active learning scaling improvement over its starting baseline of 8/128 (6.2%) via in-context Hebbian self-play and tutor consolidation loops. 57-Million-Fold FLOPs Efficiency: In execution profiling, UnisonAI generates tokens in 25.44 ms using only 86.9 FLOPs per token, representing a 57,522,124x computational efficiency increase over Google's Gemma-2B and a 73,628,319x increase over Meta's Llama-3.2-3B. Architecture & Implementation Memory is represented as deterministically-addressed exact orbits; attention is unit-capacity selection at forced locks; learning is a closure law that updates via writes rather than gradient descent. Perception (sight, hearing) is self-certified per act by integer Parseval identities. The entire architecture is verified by a 47/47 end-to-end empirical test suite including survival of process death.

Abstract of "[Depreciated and replaced by V3] From One Axiom to Master-Level Chess — and the Law Inside Neural Networks: Attention, it turns out, was not all you need" by Maria Smith, https://doi.org/10.5281/zenodo.21217278, used under its Creative Commons Attribution license (CC-BY); markup removed, otherwise unmodified.