eng-0278

The τ² 0-25 worker-config window (cfg-0183, run-0658 / run-0659). Counter differenced on the HA energy sensor: 47.96667902 kWh at 13:41:58Z -> 48.30054615 kWh at 15:27:37Z = 333.87 Wh over 1.761 hours, mean 189.6 W. wh_per_task is Wh-per-CORRECT-ANSWER: wh_total / tasks_passed (22 of 24 scored, run-0658), matching the method eng-0081 used for the earlier Q8_0 run so the two are directly comparable: 15.18 Wh per correct answer, 0.46 pence at the 30.3 p/kWh tariff. This is WHOLE-SESSION energy — the window spans simulator (user_llm) wait time and inter-task gaps, not agent-only inference — so it should not be read against a decode-only bench cell. It is far below the earlier single-slot Q8_0 run's 38.29 Wh per correct (eng-0081), for two reasons that are not a clean A/B: this run served --parallel 4, so the ~190 W draw is amortised across ~4 concurrent tasks, and it got more answers right (22 vs 15). Quant, build, reasoning and parallelism all differ at once (clm-0131). BASELINE CAVEAT: baseline_w is the canonical aihydra box-idle-all-empty floor of 10.1 W (idle-aihydra-2026-08), used so delta_w stays comparable across the whole energy series. A spot idle reading taken this session was higher (~20 W), most likely because the GPU power state had been forced high and the strix-halo fan-control daemon is now resident; a fresh idle-baseline for aihydra's post-fan-control era is an open item and would refine delta_w (it does not touch wh_per_task, which uses wh_total).