日本語 · Nemotron 3.5 Lightning のKV予算：6 KiB/token、1M ctx で約6 GiB、Mamba stateは固定 — 第三者測定値に基づく計算
2026-09-09 · Cosmos two-node LoRA, fault recovery, and an H3 block probe
日本語 · M4/DFlash測定：利用率ではなく有効tokenを最適化
Measured 2026-09-07 — what murakumo is (a control plane, not an engine), the 12.7 vs 61.5 tok/s distributed result we lost, and three corrections to our own capacity figure, kept in place rather than tidied away
Qwen3.8-27B performance, 262K context, VRAM sizing, local runs, and commercial use — from official sources
Frontier vs. open-weight performance matrix, plus what actually fits a 16GB local device
US — Section 179 / bonus depreciation timing, plus what the hardware can do the rest of the time
EU — what an audit-able run ledger looks like, for AI Act traceability
Claim-by-claim: what's live (placement, CIDv1 tenant graphs, did:key/CACAO fleet auth) vs. roadmap (on-chain settlement, a token)