# Proof Lab 01: usable is not the same as passing the brief

Retrospective reassessment: September 19, 2026. Original six generations: August 26, 2026. New generation spending: $0. No fresh generation, paid API, human repair or independent human judging was performed.

## What the retained files establish

All six files match their published SHA-256 checksums, decode without a reported FFmpeg error, are 1280 × 720 and contain no audio stream. The two Veo files contain 96 frames / 4.000 seconds; the other four contain 97 frames / 4.041667 seconds. All pass the protocol's file-format/duration check.

The historical ledger labels all six usable. Those labels remain unchanged. PL-01-v1 applies a narrower movement requirement: final card width must be 1.10–1.20 times initial width, with all required text and fiducials retained throughout. This is a new retrospective rubric, not the historical acceptance rule or a prospective preregistered result.

| Retained attempt | Endpoint finding | PL-01-v1 retrospective status |
|---|---|---|
| Veo 3.1 Lite 1 | Card width 1040→1011 px; ratio 0.9721 | Fail: final card shrinks instead of meeting the specified push-in |
| Veo 3.1 Lite 2 | Late blur defeats the fixed dark-border detector; marker-span ratio 1.0568 is only a diagnostic proxy | Needs review; no strict pass and no fabricated card-width measurement |
| Seedance 2.0 Fast 1 | Card width 1040→1092 px; ratio 1.0500 | Fail: below the new 1.10 minimum |
| Seedance 2.0 Fast 2 | Vertical card edges fall outside the final image; marker-span ratio 1.2642 | Fail: cropped card; no valid final full-card width |
| Kling 3.0 Standard 1 | Card width 1040→1092 px; ratio 1.0500 | Fail: below the new 1.10 minimum |
| Kling 3.0 Standard 2 | Card width 1040→1147 px; ratio 1.1029, close to the lower cutoff | Needs review: endpoint result alone is not full text/temporal acceptance |

Four established failures, two unresolved, zero certified strict passes. “Zero certified” is not evidence that every unresolved clip fails. Endpoint screenshots were visually inspected by an AI assistant. This is not independent human validation, an every-frame visual/OCR pass, a reliability estimate or a model ranking. Detected red regions on all frames do not certify intact characters or marker geometry.

## Reproduce this inspection without generating anything

1. Download the existing six-attempt recipe ZIP linked on this recipe. Extract it locally. It contains the input PNG, original prompts/settings, six MP4s and historical ledgers.
2. Verify the retained checksums before inspecting. The input SHA-256 is `1c80e4c731da1936361fa3e90898287dd7a4543a7a64eab971c6c2a38ce3973c`.
3. Run `ffprobe -v error -show_streams -show_format -of json FILE.mp4` and `ffmpeg -v error -i FILE.mp4 -f null -` with already-installed tools. These commands inspect/decode local files and make no provider request.
4. Use the supplied geometry script on the extracted directory. It decodes every frame as RGB24, measures dark border pixels on the fixed scanline `round(height × 0.42)` in the outer 30% bands, and separately records red-dominant corner regions. Read the code and limits before interpreting its output. Missing detections are null, not invented coordinates.
5. For endpoint width, compare the same outer card-border landmarks in first and last frames. Marker span is a cross-check, not a silent replacement for the required card width. Inspect any threshold-sensitive result manually. A ±2 px endpoint sensitivity puts Kling 2 near/below the cutoff; do not round it into a robust pass.
6. A full acceptance verdict still requires every-frame text/geometry review, motion-continuity review and documented reviewer judgment. Record disagreements as `needs_review`. A visual failure can establish failure without claiming the other criteria passed.

## Fixture feasibility

The input's card spans x=120–1159 (1040 px). A centered 15% enlargement makes it about 1196 px wide, leaving approximately 42 px on each side of a 1280 px frame. The required corner marks also fit. Thus the card can meet the proposed push-in without clipping. This does not preserve the separate outer fixture heading/footer; those are not among the three required text lines. No geometry correction was applied to the actual generated files.

## Cost boundary

The retained charges total $1.81332: $0.24 for the two Veo attempts, $0.73332 for Seedance and $0.84 for Kling. These are historical recorded provider charges, not current quotations or new spending. Original human repair time, hardware cost and electricity are unknown. There is no supported cost-per-strict-accepted-job figure because no strict pass has been certified. Do not divide by zero or label this $0 per accepted result.

## What to do with this result

Use the six files to calibrate judging and expose the difference between a usable-looking output and the requested motion. Keep the existing original evidence, prompts, charges and usability notes. The separate prospective PL-01-v1 protocol remains unrun; no fresh result is implied. A later blinded reviewer can resolve the two uncertain cases using these same retained files without paying for new generations.