# ARM 500 W — completion notes (2026-08-27, 16:19:30Z → 16:28:40Z)

175 submissions, 139 scored, **0 failures**, 1,068 power samples. Cap 500.00 W at open
and close; **all 1,068 samples carried `power_limit = 500.00`** → **no VOID cells**.
No foreign prompt ids on the shared card. No suspect cache hits; no repeated output
filenames. Seats undisturbed (worst `load_duration` 0.002 s / 0.003 s = no reload).

## The drift gate tripped once, and the reason is measured

**flux1-dev sentinels: +5.3% (n=2 pairs) — over the registered 5% REDO gate.**

It is a thermal/power settling ramp, not instability. The card's own trace (power CSV,
20 s buckets, busy samples only):

| minute | lane | temp | draw | clocks.sm |
|---|---|---|---|---|
| 16:19 | klein | 47 C | 358.9 W | 2625 MHz |
| 16:20 | flux1-dev opens | 60 C | **499.4 W (at the cap)** | 2392 MHz |
| 16:21 | flux1-dev | 72 C | 499.9 W | **1957 MHz** |
| 16:22 | flux1-dev | 76 C | 499.9 W | 1927 MHz |
| 16:23 | flux1-dev closes | 77 C | 499.7 W | 2010 MHz |
| 16:24–16:28 | schnell, SDXL | 72→59 C | 376→275 W | 2498→2677 MHz |

flux1-dev's 1024 px cell walks 9.951 s (open, 16:20:11, 60 C) → ~10.41 s (the whole
scored block, flat) → 10.635 s (close, 16:23:54, 77 C). The sentinel-OPEN is
structurally the coldest, least-power-limited render of that lane's block, so the
open→close bracket spans the settling transient by construction — and will do so
identically at every rung.

**The scored window itself is stable.** First half vs second half of every
(lane, width, steps) group with n>=4:

| lane | median within-block shift | max |
|---|---|---|
| klein-4b | +1.5% | +1.7% |
| flux1-dev | **+0.7%** | +1.8% |
| flux1-schnell | -0.2% | -6.7% |
| sdxl-albedobase | -0.2% | -2.1% |
| sdxl-base | -3.0% | -4.8% |

## A correction to the analysis, not to the data

The gate was first computed as **ratio-of-medians** and reported a phantom **-13.8%** on
klein's arm-close re-check. The instrument is PAIRED and the cell set is bimodal
(512 px ~0.29 s vs 768 px ~0.53-0.68 s), so one changed value flips which mode the
median lands in. Nine of the ten pairs moved +0.0% to +2.9%; the tenth was the arm's very
first scored render carrying weight-load overhead (0.42 s → 0.291 s on re-run) — the
first-of-group effect the sealed bench documents. **Median of PAIRED ratios = +1.4%.**
`analyse.py` now reports the paired statistic as the verdict and prints ratio-of-medians
beside it as the cross-check. Both are shown; neither is hidden.

## Per-lane medians (scored only, n=139)

| lane | tier | n | median s | min | max | sealed 600 W | paired vs sealed |
|---|---|---|---|---|---|---|---|
| klein-4b | A | 31 | 0.531 | 0.286 | 1.174 | 0.480 | 1.104x |
| flux1-dev | A | 30 | 5.829 | 3.120 | 10.509 | 5.295 | 1.102x |
| flux1-schnell | B | 24 | 0.935 | 0.526 | 1.798 | — | — |
| sdxl-albedobase | A | 27 | 4.011 | 3.780 | 4.333 | 3.810 | 1.047x |
| sdxl-base | B | 27 | 1.720 | 1.574 | 1.983 | — | — |

Sealed comparisons are the pre-registered CROSS-ENVIRONMENT observation, never a
denominator. The ladder's own denominator is the fresh 600 W arm, which has not run.

## The finding that shapes the rest of the ladder

**Only flux1-dev is power-limited at 500 W.** It sits pegged at 499.9 W with its SM clock
pulled from 2392 down to 1927 MHz. Every other lane runs well under the cap while busy:
albedobase 353 W, schnell 376 W, sdxl-base 275-289 W, klein 359 W (klein's figure is at
lane granularity only — the pre-registered sub-2-second sampling caveat).

Mid-load `nvidia-smi -q -d PERFORMANCE` probes read **`SW Power Cap: Active`, P1**, during
flux1-dev; the open/close snapshots read Not Active because the card was idle then. The
flag is the evidence; the counter is not (the armP lesson).

## LLM probe

gemma4:26b median **207.14 tok/s** decode (n=6, 206.89-207.41).
mistral-small3.2:24b median **94.64 tok/s** decode (n=6, 94.61-94.73).
Both at num_ctx 32768, keep_alive -1; worst `load_duration` 0.002 s / 0.003 s.