Your Quantization Format Is Not Free: Same-Size, Same-Label GGUF Files Decode Very Differently on Edge CPUs

Public artifact for the HotMobile 2027 submission. Controlled measurements on three CPU microarchitectures (Cortex-A76, Gracemont, Golden Cove) showing that a GGUF quantization format label is a poor predictor of decode cost: provenance changes per-tensor layout (up to 38% throughput difference at the same label), the Q4_K_M default sits below the streaming envelope on every core, and format rankings do not transfer across core types. Includes raw measurement records, per-format quantization logs, runtime repack-coverage logs, PMIC energy, perplexity, control runs, and the compiled manuscript.