eng-0145

aihydra deep-thought-posttrain throughput matrix (run-0364..run-0371, 8 cells x 2 backends) + house 4-item guard (run-0363), combined window 06:31:48Z-06:33:51Z (123s) -- matches the matrix+guard runs exactly (excludes the preceding sanity-check server starts and the earlier two-point kvprobe, none of which are referenced by a run id, per the deepseek-v4-flash/laguna-s-21/ornith-35b convention of excluding unreferenced windows). Counter interpolated from raw ~10s-resolution HA history samples pulled fresh this session: window-start bracket 24.8110571354627676/24.8111601322889376 kWh (06:31:40.557Z/06:31:50.518Z) -> interpolated 24.8111341 kWh at 06:31:48Z; window-end bracket 24.8124376982450476/24.8124698847532276 kWh (06:33:50.538Z/06:34:00.528Z) -> interpolated 24.8124392 kWh at 06:33:51Z. Delta 0.0013051 kWh = 1.3051 Wh over 123s = 38.20 W mean, 28.10 W over the 10.1 W box-idle baseline -- a small, genuinely measured delta consistent with a 361.8M model doing twelve few-second llama-bench cells and one 8-second guard call, not an artefact: the ROCm cells (06:31:48-06:32:09, slower prefill but the two most compute-dense cells in the matrix) visibly carry the counter's steepest climb in the raw samples above (24.8111601 -> 24.8120493 kWh across 06:31:50Z-06:32:30Z, the matrix's busiest 40s), settling to a shallower slope during the Vulkan cells and the guard call. wh_per_task is null: this window covers throughput + a capability guard, not a scored task set (the tau2 window is a separate energy record, joined once that run completes). tariff at the standing 30.3 p/kWh.