idle-aihydra-2026-08

First idle_baseline record for aihydra, covering its first days of service. The box-idle floor of 10.1 W is remarkably low for a 128 GB machine and is the figure every delta_w in the energy records subtracts from. igpu-resident-quiet MEASURED 2026-08-11 (eng-0035: 10 minutes, 73 GB model loaded, zero requests): **13.6 W mean, peak 35.1 W** — just +3.5 W over the empty-box floor. The keep_alive question is answered: holding the 122B resident costs ~2.5 kWh/month, ~76p at 30.3 p/kWh. This also anchors energy attribution for cloud-simulator arms: with resident-quiet this low, a high mean during such arms means genuine computation, not wait-state draw. `npu-resident-quiet` remains unmeasured — still the question an always-on NPU triage loop turns on. Duty cycle is 0.79 — this box has been working, not idling, across the period. That is unrepresentative of steady state and will fall once the benchmark programme ends; the standing-cost figures should be re-derived then rather than quoted from this window. Standing cost assumes 30.3 p/kWh grid import. With solar and battery the realised cost is lower and sometimes zero, which is what the energy records' tariff_window field exists to capture.