Script 7 is a finished ad. Scripts 6 and 2 have every chunk after the first, and each of those six transcribes word for word against the locked script with no competitor name anywhere. The opening chunk of each is regenerating as you read this. The last row shows why: feeding the model the donor's own voice made it recite her words instead of ours, and stripping the audio fixed it.
This runs 6.2 seconds and is the whole ad, showing only hands and the product, and its caption has been restored.
Chunk 1 is being regenerated silent right now. Chunks 2 and 3 below both transcribe exactly to the locked script.
Chunk 1 is being regenerated silent right now. Chunks 2 to 5 below all transcribe exactly to the locked script.
Identical prompt and references each time. The only difference is whether the donor's own voice was left on the input video.