eng-0085
window recorded
citable URL: https://halobench.com/records/eng-0085/ — this address never moves; the anchor /records/#eng-0085 keeps resolving
19.21 Whmean 193.15 W − 10.1 W idle floor (box-idle-all-empty) = 183.05 W attributable
window 2026-08-15T11:19:21Z → 2026-08-15T11:25:19Z · provenance recorded
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy) · no contention
covers run-0275
EOS-cliff x q8_0-KV probe, n_max=3 (run-0275, cfg-0073) — a cliff-detector guard run (15 generation requests, 0 degenerate), not a reward-scored capability run, so wh_per_task/pence_per_correct_answer do not apply here (left null rather than forced into the wrong denominator); Wh/request = 19.2072/15 = 1.28. Counter samples: start 19.2282059639692276 kWh (2026-08-15T11:19:19Z, 2s before window open — nearest recorded tick), end 19.2474132031202276 kWh (2026-08-15T11:25:19Z, exact). Mean active draw 193.2 W is the highest of the three joins in this batch — draft-mtp speculative decode at n_max=3 draws more power than plain serving, consistent with clm-0055's energy section (speculation's power draw rises, not flat). Retroactive join computed 2026-08-15, same day as the run.
Wh, never joules — read as a cumulative-counter difference across the window; delta over the idle baseline is the only figure that means anything