RAW DATA · BM-013

Raw evidence — Qwen3.6-27B cross-machine baseline: dual 12GB vs single 20GB vs dual 32GB — and the RQ-001 answer

Until the archive is public: email lab@localaifrontier.com with the benchmark ID (bm-013) and we will send the raw package directly. We do not withhold raw data — it is simply too large for the static site bundle.

Summary of what is in the package:

• config.json — the exact runtime invocation (matches the report's Configuration block).
• runs/ — 5 measured runs as JSON (prompt tok/s, gen tok/s, TTFT, memory samples).
• warmup/ — 3 discarded warmup runs, kept for transparency.
• workload-pack-v1.json — the prompts used.
• system-info.txt — driver/runtime versions, env vars, power readings.
• checksums.sha256 — hashes for all of the above.