The decision
I defined two human gates: the intended film-wide decisions must be visible in the pixels, and I must be willing to publish the film. Cut placement, rhythm and seams are recorded as diagnosis, not as gates.
The rubric names the film, asks for worst-instance timestamps, and samples three random parts plus the opening and closing minute. That costs review time and judgment, and a revised cut needs its own acceptance.
From the source
Human-watch rubric. Which rows gate the launch and which only diagnose:
Rows 3 and 5 GATE the launch. A FAIL on either stops it. Rows 1, 2 and 4 are recorded as diagnosis: they are defect counts that say what to fix next, and a FAIL there does not on its own hold the launch (Spec 194 Q7, answered 8 September 2026).
The recorded verdict. The export identifier is redacted:
Watched by the owner, file
[REDACTED: reviewed export A identifier]. Both gating rows PASS — the human half of Spec 191's launch gate is MET.
NASA-footage run, 10 October 2026. A concrete gap in automated caption QA:
Burned-in captions do not avoid text burned into the source (NASA's banners); the caption-collision QA only knows the editor's own elements
Keep the artifacts and verdicts distinct
| Artifact | What it supports | Boundary |
|---|---|---|
| Export A, rendered 6 September, 48:20 | The 8 September review passed both human gates | That approval covers this export only |
| Later revised cut | Its density gate passed | Needed its own export and watch |
| October NASA-footage run | Automated QA missed collisions with text burned into the source | Separate footage and run |
What this shows
Human acceptance is bound to one named export, and automated checks and human inspection contribute different evidence. In the NASA run, source comparison also kept static slides and source-native black frames from being reported as rendering defects.