eng-0081
window recorded
citable URL: https://halobench.com/records/eng-0081/ — this address never moves; the anchor /records/#eng-0081 keeps resolving
574.4 Whmean 181.4 W − 10.1 W idle floor (box-idle-all-empty) = 171.3 W attributable
window 2026-08-15T02:11:33Z → 2026-08-15T05:21:33Z · provenance recorded
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy) · no contention
per task 38.2913 Wh · 1.1602 p at 30.3 p/kWh
covers run-0271
aihydra run-meta.jsonl entry qwen38-screen/tau2-full (rc=124 — the 11400s wall bound cut the run at 26 of the requested 50 tasks; run-0271 records the 26 completed and scored). wh_per_task is Wh-per-CORRECT-ANSWER: wh_total / tasks_passed (15 of 26, run-0271.metrics). This is WHOLE-SESSION energy, not agent-only — the window includes simulator (user_llm) wait time and a ~19-minute gap between task 9 and task 10 (see run-0271s per-task log), so it should not be read as pure inference cost. Retroactive join computed 2026-08-15, the day after the bench published without its energy figure.
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
model pages qwen38-27b
docs benchmark-protocol