This paper presents a zero-gradient compilation method that embeds exact arithmetic circuits directly into pretrained LLM weights without training data or runtime hooks. The approach uses three analytical mechanisms: (1) two-layer Mellin-GLU quadrature over bias-free SwiGLU/GeGLU layers, converting unknown per-token RMSNorm scaling into translation along a log-spaced gate grid with >250× dynamic range and IEEE 754 error bounds; (2) cross-token scale cancellation via signed Chebyshev Indefinite-RoPE self-sink attention and memoryless SiLU-antisymmetric GatedDeltaNet shifts; (3) axis-aligned width extension (d → d+K) with √(d/(d+K)) norm compensation, Base-1000 canonical BPE emission, and differential (+254, -256) tied unembeddings. An automated 5-stage compiler synthesizes 4-8 digit signed {+, -, ×} circuits in constant TC⁰ depth, plus 3-literal Boolean circuits (100%, N=1,728; validated on 42-layer Gemma 4 E4B) and 5-bit Majority/Parity circuits (100%). On 6-digit arithmetic (N=1,200, BF16), IBM Granite 4.2 3B (1.07M-1.33M active non-zeros, ≤0.036%) and Alibaba Qwen 3.5 4B (2.25M-2.39M non-zeros, ≤0.057%) achieve 99.17% and 100.00% ALU-prefix accuracy (100% with FP32 sidecar) and 98%-100% mid-sentence resumption, outperforming 12.8K-sample LoRA (42.2%/61.7%, 0% inline continuation). A factorized block-sparse sidecar fits A100 memory at 7.89/10.65 GiB (4.60/8.14 GiB under INT8/INT4 + FP32 sidecar) and passes all 14/14 pre-registered >2K-token Tango splits (N=4,392; 0/11,134 OOD triggers).

Read original