Benchmark data and reproducibility package: DeepSeek-R1-Distill-Llama-8B on 4x Raspberry Pi 5 (Shackled to the Machine)

Raw benchmark data and reproducibility package supplementing the technical report "Shackled to the Machine - Breaking the DeepSeek-R1-8B Public Record by +29% on Four Raspberry Pi 5" (doi.org/10.5281/zenodo.20411979). The study evaluates CPU-only distributed inference of DeepSeek-R1-Distill-Llama-8B (dense 8B, Q40) on a four-node Raspberry Pi 5 (16 GB) cluster running distributed-llama. A clean-room measurement at the optimal thread count sustains 8.32 tok/s decode (n=20, 95% CI [8.30, 8.34]), bit-exact - to our knowledge the best published figure for this model on Raspberry Pi 5, 29% above the prior four-node record (6.43 tok/s). Dense decode is memory-bandwidth-bound: the production binary is byte-identical to clean upstream, so the fork's source and kernel changes contribute approximately zero to dense decode; the gains live in the hardware and operating-system layers, not the application kernels. This package contains the raw measurement logs (clean-room thread sweep, fork-vs-stock attribution, the authoritative n=20 decode benchmark, and the decode-vs-context sweeps), figures and the LaTeX sources needed to reproduce the report's tables and plots. All figures are bit-exact (SHA-256 of the generated token-id stream at temperature 0, seed 42); no overclocking, quantisation change or model substitution is used. Code: the cluster runs a patched fork of b4rtaz/distributed-llama (MIT License); the exact source snapshot is archived at doi.org/10.5281/zenodo.21469506. Developed at Hellomatik (https://hellomatik.com).