eng-0083
window recorded
citable URL: https://halobench.com/records/eng-0083/ — this address never moves; the anchor /records/#eng-0083 keeps resolving
103.9 Whmean 135.7 W − 10.1 W idle floor (box-idle-all-empty) = 125.6 W attributable
window 2026-08-15T06:34:00Z → 2026-08-15T07:19:55Z · provenance recorded
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy) · no contention
per task 8.6562 Wh · 0.2623 p at 30.3 p/kWh
covers run-0273
aihydra run-meta.jsonl entry kv-quality-patched/q8_0 (2026-08-15T06:34:00Z, rc=0). wh_per_task is Wh-per-CORRECT-ANSWER: wh_total / tasks_passed (12 of 14, run-0273.metrics). Task-matched against eng-0082s f16 arm — identical 14 task ids, same seed. Both wall time (2755s vs 4303s) and energy (103.874 Wh vs 163.696 Wh) are well below the f16 arm despite scoring HIGHER (12/14 vs 11/14): q8_0 KV is not merely quality-neutral on the patched build (clm-0058) — it is also the cheaper arm to run, on both counts, over the identical task set. See clm-0058 for the full comparison. Retroactive join computed 2026-08-15.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything