eng-0281
window recorded
citable URL: https://halobench.com/records/eng-0281/ — this address never moves; the anchor /records/#eng-0281 keeps resolving
105.4 Whmean 152.3 W − 10.1 W idle floor (box-idle-all-empty) = 142.2 W attributable
window 2026-10-03T13:42:22Z → 2026-10-03T14:23:52Z · provenance recorded
method wall-meter · aibeast plug energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_beast_total_energy, a non-resetting integration of the plug's 1 s power reading; raw history queried via ha_get_history at the window edges and a midpoint) · no contention
per task 4.47 Wh · 0.136 p at 30.3 p/kWh
covers run-0674
tau2 airline tasks 0-25 (run-0674, wall 2490 s). Energy counter 0.049292 -> 0.154662 kWh = 105.4 Wh total; midpoint 14:02:30Z = 0.098516 kWh cross-checks a steady draw (146.7 W first half, 157.8 W second half). Idle floor 10.1 W x 2490 s / 3600 = 6.99 Wh; active 98.4 Wh. Headline Wh-per-correct = active_wh / 22 correct = 4.47 Wh; pence-per-correct = 0.136 p @ 30.3 p/kWh; whole-run cost 3.19 p. baseline_w 10.1 W is the same box-idle-all-empty floor used for eng-0275 and eng-0277 (identical Bosgame M5 hardware). Reasoning-matched to eng-0275 (both reasoning_effort=low, same quant), so 4.47 vs 9.90 Wh-per-correct is an engine comparison on one quant; N=1 each. Production traffic shared the box during the window (interactive turns are admitted ahead of bench turns by the proxy); it was light and is included in the total.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything
Cited by — computed at build time, never stored
claims clm-0136
model pages qwen38-flash-next