[Depreciated and replaced by V3] Defeating Deep Learning: A Zero-Parameter Geometric Engine Beats KataGo at 19x19 Go
Year: 2026
Type: article
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. 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. Go rebuilt as exact counted law. Fold Go is a zero-trained-parameter computational proof in which legality is connected-component counting, board states are finite three-colour fields, and solved values are produced by exhaustive execution rather than learned policy or value networks. The engine exactly reproduces the legal-position inventories 1, 57, 12,675, and 24,318,165 for boards 1×1 through 4×4, plus 5 and 489 for 1×2 and 2×3, with zero disagreements from an independent referee. It freshly secures exact empty-board values 1×1 = 0, 1×2 = 0, and 2×2 = +1, the latter through 17,038,501 visited nodes. The same programme has produced two independent, replay-verified 2–0 batches against GNU Go 3.8 on 9×9: every move was legal, every game ended pass–pass, and every internal score reproduced. Together, the exact census, solved games, and competitive receipts show a continuous route from theorem-forced rule structure to functioning play without trained weights.
Author: Smith, Maria
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7167808894