Home › Evidence › Records › eng-0144
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

eng-0144

window recorded
citable URL: https://halobench.com/records/eng-0144/ — this address never moves; the anchor /records/#eng-0144 keeps resolving
2055.2 Whmean 181.18 W − 10.1 W idle floor (box-idle-all-empty) = 171.08 W attributable
window 2026-08-16T16:14:57Z → 2026-08-17T03:39:04Z · provenance recorded
method wall-meter · aihydra energy sensor via Home Assistant (counter-difference on sensor.hardware_ai_hydra_energy, raw ~10s-resolution history samples, linearly interpolated to each run-meta window; two windows summed, not a single counter-diff over the outer span) · no contention
per task 128.4496 Wh · 3.892 p at 30.3 p/kWh
covers run-0358 run-0360
COMBINED energy across the full standard 26-task tau2 airline set, summed from its two constituent arm windows (eng-0141: arm1, 1446.5888 Wh over 28,800s; eng-0143: arm2/continuation, 608.6042 Wh over 12,036s) = 2055.1930 Wh total. This is a SUM of two measured windows, NOT a fresh counter-difference over the full outer span (16:14:57Z-03:39:04Z, which would also include a ~3-minute server-stop/guard/ server-restart gap between the two sessions — matching the exclude-overhead convention every other board entry's tau2 energy figure already uses, e.g. ornith-35b eng-0134 excluding its preceding 13s server-load window). mean_w_active (181.18 W) and delta_w (171.08 W) are computed over the SUM of the two active durations (28,800s + 12,036s = 40,836s), not the outer wall-clock span — a reasonable single figure since both arms ran the same config at similar load (arm1 mean 180.82 W, arm2 mean 182.04 W, within 1.2 W of each other); see eng-0141/eng-0143 for the per-arm mean/delta wattage individually. wh_per_task is Wh-per-CORRECT-ANSWER across BOTH arms: 2055.1930 / 16 (14 arm1 + 2 arm2 passed, of 25 scored) = 128.4496 Wh, 3.8920 pence at 30.3 p/kWh. THE LOWEST-RANKED Wh-per-correct-answer of any full-bench candidate's Wh-per-correct-answer figure on this board (clm-0076): trails qwen38-27b's previous-worst 38.29 Wh/correct by >3x, and ornith-35b's leading 6.48 Wh/correct by ~20x.

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-0076
model pages qwen36-27b-mtp