LocalScore methodology (llamafile, 9-task geometric mean) + HX 370 / Radeon 890M submitted result

Methodology page states the published Prompt and Generation tok/s are a blend across NINE tasks (prompt lengths 16/64/384/1024/1280/2048/4096 and generation 16..3072), scored as a geometric mean via llamafile on a single device. The HX 370 numbers live at https://www.localscore.ai/accelerator/721 (AMD Ryzen AI 9 HX 370 w/ Radeon 890M, 26.5 GB): Llama 3.2 1B Q4_K_M 551 pp / 67.0 gen / TTFT 2.53 s; Llama 3.1 8B Q4_K_M 99 pp / 12.9 gen / TTFT 13.61 s; Qwen2.5 14B Q4_K_M 51 pp / 7.1 gen / TTFT 26.30 s. Because the headline rate is a blend and not one (prompt, output) pair, records taken from it are status=draft and must not be quoted as a single-context llama-bench figure. Community-submitted, not vendor.