Finite-Field Edge Inference with GF(137) Kernels and Compact Virtual Gate Emulation

Preprint. This work reports an experimental Python/C++ finite-field edge-inference stack in which dense neural-network operations are represented over the residue ring Z/137Z and executed by fused native kernels. The benchmark stores model weights as byte-valued residues, fuses matrix multiplication, modular reduction, and threshold activation, and compares the result against float32 and JIT baselines. The release also describes a compact virtual gate emulator as a deterministic finite-field systems benchmark, not as a physical quantum simulation result. The accompanying software and reproducibility package is archived separately as Zenodo DOI 10.5281/zenodo.20436771. This record includes the English manuscript PDF and a Russian translation PDF.