eng-0088
window recorded
citable URL: https://halobench.com/records/eng-0088/ — this address never moves; the anchor /records/#eng-0088 keeps resolving
22.01 Whmean 184.3 W − 10.1 W idle floor (box-idle-all-empty) = 174.2 W attributable
window 2026-08-15T12:17:08Z → 2026-08-15T12:24:18Z · provenance recorded
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy) · no contention
covers run-0278
EOS-cliff cause-isolation matrix, cell B / backend-flip-only (run-0278, cfg-0076) — a cliff-detector guard run (15 generation requests, 0 degenerate), not a reward-scored capability run; wh_per_task/ pence_per_correct_answer left null for the same reason as eng-0087. ROCm draws visibly more power than the Vulkan baseline cell over the identical workload shape (184.3 W mean vs eng-0087's 160.0 W) — consistent with clm-0050's per-model/per-phase backend note, not itself evidence about the cliff mechanism. Counter samples: start 19.3329050987958876 kWh (shared boundary with eng-0087's end), end 19.3549172133207276 kWh (13:24:19 local, 1s after window close). 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