Generation rules
- Text-to-video only, no reference media; eight seconds, 16:9, the closest native tier to 720p, audio generated with the video.
- One request is one benchmark sample. No choosing the best of several attempts.
- Prompt enhancement or rewriting switched off wherever the provider exposes the switch.
- No editing, trimming, normalising or replacing audio before submission.
- A refusal or error is recorded, not worked around.
- Raw output, the exact provider route, settings and moderation state are preserved.
How each cell's clip is chosen
A cell is one configuration in one scene. Its canonical outcome is chosen by fixed rules, never by quality:
- A provider refusal of the exact frozen prompt is the cell's outcome. A clip generated afterwards from a reworded prompt never replaces it.
- Where an earlier, frozen R00 sample already exists for the scene, that immutable sample is the canonical outcome; a newer generation of the same configuration is kept separately.
- Otherwise, the first successfully completed generation under the intended prompt and settings.
- Later generations are replicates. Set-up runs from before the settings were standardised are neither candidates nor replicates.
The result: 64 cells, of which 63 produced a clip for review and 1 is covered by a refusal. See refusals.
Provenance, at its true strength
For 9 configurations the delivered file itself names the vendor. For 7, nothing in the file does, and the exact route is attested by the person who generated it. The two are recorded separately and never merged: an attestation is published as an attestation, not as something the bytes prove. The roster shows which basis applies to each configuration.