Stayfound remake · pass 2

What pass 1 got wrong and this pass fixes

The model no longer redraws the donor's captions. Every burned-in caption was erased from the input clips on a rented GPU before the swap, so there is nothing left for the model to copy. A scan of the six finished chunks finds no caption text.
The competitor's name is gone from the inputs. Their pink card has "SPOTMINDERS." printed on it, and it was plainly readable in the clips we fed the model last time, which is how their card kept coming back. The caption remover never touched it because it only masks things pinned to the screen, and a brand printed on a moving prop is invisible to that test.
The presenter is ours in every talking shot, and the padding the model used to add at the end of each chunk is trimmed off.

The QA verdict: 18 fatal defects, so neither cut is shippable

Run with tools/video_qa.py --stage final. The readable verdicts ship with this page at c001 and c002. My own eyeball check caught the pink card in 5 of 18 sampled frames and called it "several shots"; QA found it in 14 and graded every one of them fatal, which is the honest number.

Wrong product, 14 fatal. The donor films both a black card and a pink one, so wherever their card is genuinely pink the model keeps that colour straight through the swap. Our card is matte black. On c002 at 9 seconds the phone screen itself reads "Spotminders Tracking Card", so the model rebuilt the competitor's name inside the handset even though it was removed from the card in the input. The fix is to recolour the card inside the input clips and mask the phone screen, rather than ask the model again.
Unapproved faces, 3 fatal. The secondary actors who hold the product are strangers rather than our cast, because only the presenter has a reference sheet.
Malformed hands, 1 fatal and 1 major, with fused and extra fingers in shots where the hands are the action.
Two chunks still say "the car" where the line reads "the card." I added a note to the prompt telling the model how to pronounce the word, and the model read that note aloud: one chunk ends with it speaking the word "Card" as though it were a line. That note is coming out.

Bridge club hook · 35.6 s

Ours, pass 2

Spotminders, the donor

186 days hook · 36.5 s

Ours, pass 2

Spotminders, the donor

What the spoken track says, word for word

Transcribed from the finished chunks with Whisper, so you can read what the model actually said against the lines it was given.

c001-A "I'm the last person in my bridge club who'd buy a gadget, and I did. Setup runs through Apple Find My, already on every iPhone. My granddaughter held the car near mine, tapped a couple of buttons, and the map showed my handbag sitting in the kitchen."
c001-B "That was the last time I thought about it. At 1.8 millimeters thin, the card disappears inside my wallet, and the battery has held up for eight months with no reminder. Over 500,000 of us carry one."
c002-A "It's been in my handbag for 186 days, and I haven't charged it once. Setup runs through Apple Find My, already on every iPhone. My granddaughter held the card near mine, tapped a couple of buttons, and the map showed my handbag sitting in the kitchen." Followed by the model reading my pronunciation note aloud.
c002-B "That was the last time I thought about it. At 1.8 millimeters thin, the car disappears inside my wallet, and the battery has held up for eight months with no reminder. Over 500,000 of us carry one."