The same two cuts, re-rendered after the donor's audio was stripped from every input clip, then captioned in the house style. $5.06 for six chunks and 376 seconds of render. The pink card stays in the picture on Kalp's ruling.
Pass 2 uploaded each donor chunk with its soundtrack still on it, and the model treated that track as material and kept it. Correlating the donor's audio against our own, sample for sample, before and after:
| chunk | pass 2 | pass 3 |
|---|---|---|
| c001-A | +0.656 | +0.013 |
| c002-A | +0.309 | -0.016 |
| c001-B | +0.104 | -0.010 |
| c002-B | +0.097 | -0.021 |
| c001-C | +0.002 | -0.007 |
| c002-C | +0.066 | +0.002 |
Unrelated clips read 0.004 on this measure and the same recording shifted by one second reads 0.006, so every figure in the right-hand column sits at the noise floor. The Spotminders creator's voice is out of our ad. The bridge-club cut now speaks our script word for word with nothing left to correct. Two slips remain in the 186 days cut, where the voice says "Apple Find of" at 6 s and "my Walt" at 19 s, and the captions carry the right words over both.
Two further faults went with it. The model used to read the pronunciation note in the prompt aloud as a spoken line, and that note is deleted. The join used to cut each chunk at the donor's exact second, which chopped the closing word off four of the six chunks once the model began pacing slower, so it now cuts just past the last word actually spoken and refuses to do so wherever the closing phrase repeats.
Fatal counts fell from 8 and 10 to 2 and 8, and on the bridge-club cut the wrong-card count fell from 5 to 1. What did not move is who appears on screen. The verdicts ship with this page at c001 and c002.
| Where | What is on screen |
|---|---|
| c001 at 22 s | The presenter becomes another woman mid-shot with no cut, with straight grey hair and no glasses, still wearing our rust cardigan. The hair is the donor presenter's. |
| c001 at 24 s | The bald man with the melted face is still there, holding two cards |
| c001 at 2 s | The product box reads "Tarking Card" |
| c002 at 28 s | The presenter is replaced by a grey-haired woman in a green sweater, which is the donor presenter's own look |
| c002 at 5 s | Distorted extra fingers on the hands working the phone |
| c002 at 7 s and 36 s | The phone screen renders its text mirrored and garbled |
The prompt was hardened for this pass to say that every other person is a woman between 45 and 75 and that no man appears at any moment, and the man came back regardless, so hardening the wording has not controlled who appears. The rule that does hold is that whatever the model is handed a PICTURE of it obeys, while whatever is left out of the pictures it inherits from the donor clip. We supplied one cast sheet, for the presenter, which leaves every secondary woman uncontrolled and lets the donor's own presenter bleed back in.
The fix with a price: two more Seedream cast sheets for the secondary women at about $0.09 on BytePlus, then a re-render of only the three chunks that contain them, c001-B, c001-C and c002-C, at $2.43. Nothing else needs touching.
| Item | Amount |
|---|---|
| Pass 2, six chunks | $5.06 |
| Pass 3, six chunks with silent inputs | $5.06 |
| Caption wipe, transcription and QA | about $0.25 |
| WaveSpeed balance remaining | about $42 |
| The other eight of the top ten, once the cast is fixed | about $38 |