# Proof Lab 01 addendum: two retained outputs reviewed

Reviewed September 20, 2026 Pacific (September 21 UTC). Both formerly unresolved clips now have documented failure evidence under the September 19 **retrospective PL-01-v1 rubric**. This adds two negative findings to the four already established: six documented failures under the retrospective rubric, with no certified strict pass. It does not certify every-frame text acceptance, independent human validation, or a model ranking.

No new video was generated. The original videos, historical usability labels and September 19 public report remain unchanged. This dated addendum records the subsequent review. No newsletter was sent and no experiment configuration was changed by this review.

| Retained output | New finding | Retrospective verdict |
|---|---|---|
| Veo 3.1 Lite, attempt 2, August 26 | Same-border endpoint sensitivity gives 1.0567–1.0712 × initial width, below the required 1.10 minimum. | Fail: requested movement |
| Kling 3.0 Standard, attempt 2, August 26 | The upper-left corner cross loses its right arm at frame 8 (0.3333s). The lower-right cross loses most of its horizontal arm at frame 32 (1.3333s). The last frame looks intact. | Fail: fiducial preservation and temporal geometry |

## What the rubric requires

The [original recipe brief](https://kingy.ai/wp-content/plugins/kingy-workbench/assets/video-exact-text-recipe.zip) requested a 15% push-in and allowed 5–25% apparent enlargement. The [September 19 protocol](https://kingy.ai/wp-content/uploads/kingy-proof-lab-01-20260919/proof-lab-01-protocol-v1.md) tightened the movement range to 10–20%. That later range is a retrospective rule for these August files, not a rule declared before their generation. Both explicitly require the corner fiducials to be preserved.

The older [September 19 reassessment](https://kingy.ai/wp-content/uploads/kingy-proof-lab-01-20260919/proof-lab-01-retrospective-20260919.md) remains preserved. The six historical `subjective_usable=true` labels remain untouched. A stricter retrospective failure does not rewrite that history.

## Veo 2: recovered border measurement, with limits

The original RGB<90 border detector could not locate the late blurred border. Its null result and marker-span diagnostic remain in [the original target extract](original-geometry-v1-target-extract.json). This review does not replace that null with a marker measurement.

The supplemental [v2 inspection script](proof-lab-01-addendum-20260921-v1.zip) measures the same outer card border at four blank-band scanlines (y=280, 290, 300, 310) across nine mean-RGB thresholds (100, 120, 140, 160, 180, 190, 200, 205, 210). Both left and right edges must be detected in the outer 30% bands. Missing edges remain null. This method was added after viewing the blur and is a retrospective diagnostic, not a preregistered measurement.

The initial width is 1040px. Of 36 endpoint pairs, 24 provide both edges; 12 remain null. Final width ranges from 1099 to 1114px, or 1.0567–1.0712 times the initial width. The maximum valid width across every decoded frame is also 1114px. As an illustrative stress check, expanding the final boundary by 8px per edge and shrinking the initial boundary by 2px per edge still gives 1130/1036 = 1.0907, below 1.10. That allowance is not a statistical confidence interval.

Native [first](video-exact-text-veo31lite-2-frame-000.png) and [last](video-exact-text-veo31lite-2-frame-095.png) frames were visually inspected. The visible border supports the sub-threshold enlargement. We do not claim that this proves a failure of the wider original 5–25% movement rule, or that the text passed all frames.

## Kling 2: endpoint geometry missed transient damage

The endpoint measurement is close to the cutoff: the original detector returned 1.1029; v2 scanline/threshold variants return 1.1038–1.1058. Neither establishes overall acceptance.

Compare the native [first frame](video-exact-text-kling30standard-2-frame-000.png), [frame 8](video-exact-text-kling30standard-2-frame-008.png), [frame 32](video-exact-text-kling30standard-2-frame-032.png) and [last frame](video-exact-text-kling30standard-2-frame-096.png). The upper-left cross is partly erased at frame 8, and the lower-right cross is partly erased at frame 32. The card itself remains on screen. This violates the explicit instruction to preserve the original corner fiducials and the protocol's temporal geometry requirement.

The old detector found a red region in every quadrant on every frame. That establishes the presence of red pixels, not preservation of each full cross. We make no spelling-error claim. Failure of one required condition is sufficient; it does not require pretending that unreviewed conditions passed.

## Reproduce without spending

Use the existing [six-attempt recipe ZIP](https://kingy.ai/wp-content/plugins/kingy-workbench/assets/video-exact-text-recipe.zip), or the already retained local extraction. The script verifies all ten recipe files against the original SHA256SUMS before decoding the two target MP4s. It uses installed FFmpeg/ffprobe, Python, NumPy and Pillow, and makes no network or provider request.

```sh
python3 inspect_retained_v2.py /path/to/extracted/video-exact-text-recipe /path/to/review-output
```

The output includes native evidence frames, contact sheets for screening, media metadata and [all-frame measurement records](measurements-v2.json). The verdict is recorded separately in [review-verdicts.json](review-verdicts.json). Contact sheets are screening aids; they are not every-frame text acceptance. An AI assistant inspected the cited native frames, and the coordinating AI assistant independently checked those same frames. These are AI checks, not two human judges.

All six historical provider charges still total US$1.81332. Repair time, hardware cost and electricity were not measured. Cost per strict accepted job remains undefined because no strict pass has been certified. Prospective PL-01-v1 runs remain unrun.