# ARM 450 W — completion notes (2026-08-27, 16:42:37Z → 16:54:39Z)

215 submissions (widened n=6 sentinel brackets), 139 scored, **0 failures**, 1,412 power
samples. Cap 450.00 W at open and close; **all 1,412 samples carried
`power_limit = 450.00`** → **no VOID cells**. No foreign prompt ids. No suspect cache
hits, no repeated output filenames. Seats undisturbed (worst `load_duration` 0.004 s).

## Gate: PASSED

n=10 paired arm-close re-check (the Amendment 3/4 gate): **+2.5%** — inside the 5% gate.
(Ratio-of-medians printed beside it reads -7.0%, the same bimodal artefact Amendment 4
documents; the paired figure is the verdict.)

## Diagnostic: the widening worked

| lane | sentinel bracket, paired (n=6) | was at 500 W (n=2) |
|---|---|---|
| klein-4b | +0.7% | +2.3% |
| flux1-dev | **+3.2%** | +5.3% |
| flux1-schnell | -0.3% | +3.7% |
| sdxl-albedobase | -0.2% | +0.3% |
| sdxl-base | -0.4% | +2.6% |

At n=2 the bracket was dominated by single-render noise; at n=6 every lane reads inside
±3.2% and the residual on flux1-dev is the settling ramp Amendment 4 identified. No lane
would have tripped a 5% threshold at this rung even had the bracket still been the gate.

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

| lane | tier | n | median s | min | max | sealed 600 W | paired vs sealed |
|---|---|---|---|---|---|---|---|
| klein-4b | A | 31 | 0.566 | 0.286 | 1.107 | 0.480 | 1.149x |
| flux1-dev | A | 30 | 6.286 | 3.342 | 11.325 | 5.295 | 1.189x |
| flux1-schnell | B | 24 | 0.990 | 0.545 | 1.849 | — | — |
| sdxl-albedobase | A | 27 | 4.045 | 3.824 | 4.410 | 3.810 | 1.063x |
| sdxl-base | B | 27 | 1.705 | 1.585 | 2.078 | — | — |

Sealed columns remain the cross-environment observation, never a denominator.

## Interim 450 vs 500 (paired, same cell + seed) — NOT the headline

The registered denominator is the fresh 600 W arm, which has not run. Reported here only
because the two arms are complete and directly comparable.

| lane | n | median 500 W | median 450 W | paired 450/500 |
|---|---|---|---|---|
| flux1-dev | 30 | 5.829 | 6.286 | **1.072x** |
| flux1-schnell | 24 | 0.935 | 0.990 | 1.052x |
| klein-4b | 31 | 0.531 | 0.566 | 1.040x |
| sdxl-albedobase | 27 | 4.011 | 4.045 | 1.015x |
| sdxl-base | 27 | 1.720 | 1.705 | 1.003x |

flux1-dev by resolution: 512 px 1.067x · 768 px 1.072x · 1024 px 1.081x.

## Which lanes the cap actually binds (power CSV, per lane's own measurement block)

| lane | 500 W arm: p50 / p90 draw, SM | binds | 450 W arm: p50 / p90 draw, SM | binds |
|---|---|---|---|---|
| klein-4b | 364.8 / 407.4 W, 2617 MHz | no | 359.2 / 397.8 W, 2520 MHz | no |
| flux1-dev | 499.9 / 500.8 W, 2010 MHz | **YES** | 450.0 / 450.7 W, **1796 MHz** | **YES** |
| flux1-schnell | 395.3 / 487.9 W, 2092 MHz | edge | 381.9 / 452.0 W, 1894 MHz | **YES** |
| sdxl-albedobase | 296.9 / 435.5 W, 2677 MHz | no | 298.8 / 429.4 W, 2632 MHz | edge |
| sdxl-base | 249.5 / 422.4 W, 2677 MHz | no | 274.7 / 421.7 W, 2662 MHz | edge |

The ordering of the slowdown follows the ordering of the binding, which is the result
holding together: flux1-dev is pegged at both caps and pays the most; sdxl-base never
approaches either and pays nothing measurable.

**Two honest caveats.** (1) klein slows 4.0% while its sampled p90 sits ~50 W under the
cap — at 0.29–1.1 s a render, a 500 ms sampler cannot see transient peaks, which is the
sub-2-second caveat pre-registered in §4. klein's "no" means *not sustained at the cap*,
not *never touches it*. (2) The in-block `PERFORMANCE` flag proved a sampling-luck
instrument: a single instantaneous ssh read lands wherever it lands between renders, and
at 450 W it caught `SW Power Cap: Active` only during klein. **The power CSV, not the
flag, is the binding evidence at this granularity**; the flag stands only as corroboration
where it fires.

## LLM probe

gemma4:26b **207.16 tok/s** decode (n=6, 207.00–207.37) — 500 W was 207.14. Flat.
mistral-small3.2:24b **93.50 tok/s** (n=6, 93.05–93.56) — 500 W was 94.64. -1.2%.
Both at num_ctx 32768, keep_alive -1.