The Warlock Commandery Taroos the Warlock

Field Reports / 14 SEP 2026

Media / Video Workproofmediadeterminismproof

49 of 49

A seeded rerun of a video shot reproduced all 49 frames pixel-identical, maximum difference zero. It only counted because the rerun path was fixed first: it had been quietly generating a fresh source image, so a rerun could never have reproduced anything. Also on record: a retry branch that two reports called test-proven and that no test had ever executed.

What happened

The claim under test: the same seed, the same source still, and the same profile reproduce a video shot exactly on this stack. If that is true, a seed is a lineage guarantee. If it is false, “rerun with seed” is a new render wearing a reproduction’s name.

Shot 1 of the acceptance storyboard, originally rendered at one seed, was rerun twice at another seed through the registered rerun phrase, under a media session.

What was found first

The rerun path had been calling the video sweep without the source image. Each “rerun” was conditioning on a freshly generated still. The pipeline would have produced a different video every time and called it a rerun.

That was fixed before the test ran. The rerun now re-conditions on the original shot’s registered source still and refuses outright when that file is missing. Refusing is the correct behavior. A rerun without its source is not a rerun.

How identity was checked

Not by file hash. The two output files differ by hash because of container timestamps. Both videos were decoded and compared frame by frame.

49 of 49 frames pixel-identical. Maximum absolute pixel difference: 0. PSNR: infinite.

Model integrity is checked separately and by hash: 14 models verified, 340 registered assets, 0 mismatches. The video core is Wan 2.2 5B, with FLUX.2 Klein for stills, RIFE for interpolation, and Real-ESRGAN for upscale.

The other item on this record

The media mission’s closeout carries an honest list of what was not done. One entry deserves its own paragraph.

The continuity retry branch, the code path that re-renders a shot when it falls outside the quality band, had been described as “test-proven” in two separate reports. No test had ever executed it. The first production shot to reach the band crashed on an undefined constant.

It has since been fixed and end-to-end tested, and it has still never executed in production, because no shot since has crossed the threshold. The closeout says so in those words, and says never to force it. A branch that has run in a test and not in production is labeled exactly that.

The second open item on the same list: the media session is mandatory by contract but not enforced by any route. That is an owner decision and it is recorded as one.

What was proven

Determinism on the production video path, by frame comparison, with the defect that would have faked it fixed first and written down.

Next target

The retry branch fires on its own in production, and the receipt is filed here when it does.

Two lessons for anyone auditing a pipeline: compare frames, not files, and treat “test-proven” as a claim to verify, not a status. The contact button opens with this report as the subject.