This is the shipped A-final.
Every clip below uses the same clean join as the shipped A-final. On-screen words come from the locked script, the transcript, or Doe's live page. Each render ran on the Vast cut box in about 33 seconds.
The left video is what shipped. The right video is the same plan put through the new HyperFrames stage, which took 82 s on the cut box. It has the style1 sticker, her before/after as a wipe on "didn't", and the house captions. Every element sits clear of her eyes and mouth, and there is no review card.
This is the shipped A-final.
This is the HyperFrames stage (hf_layer.py), built from the same plan.
Shipped A-final (PIL stickers, ffmpeg overlays with a 0.12 s fade and a hard cut out).
HyperFrames, same inputs. The sticker pops in and lifts out, and the overlays sit below TikTok's header band with rounded corners. The before/after lands before first, then after. The captions are the house look, unchanged.
Her first two words, large, behind her head (HyperFrames mattes her out on the box in 46 s). Her own before/after as one card that wipes from before to after.
Two review cards from Doe's page: the stars fill one by one and the quote lights up as it is read. Known fault: in this stretch her face sits high, so the card covers her eyes. Placement would have to come from our face detection.
Five caption treatments on the same 4.5 s: the house look, house with a pop, the spoken word in yellow, a white box building word by word, and words rising as spoken. Variant 4's box shows at full width before its words arrive.
Same words as each reference. HTML draws colour emoji and the comment-reply bubble (style3), which render_sticker.py cannot.
One command on the box (88 s) returns full-page and scroll screenshots, 85 images, the fonts, colours, spacing and shadows, and the visible text. It captures a desktop viewport only.
