⛔ NOT FOR LAUNCH — INTERNAL TEST. Unqualified hair-growth claims, no compliance qualifiers. Not cleared for any live placement.

Wonder v22 — proof furniture

10 concepts × native 4:5 and 9:16.

What I got wrong in v21

When you said no boxes, I stripped out everything — stars, review badges, guarantee bars — and shipped bare type. That was an over-correction. The refs you sent are full of proof furniture: the yellow bar, the star roundel, the guarantee footer, the big review oval. Those aren't crutches. What you banned was a rectangle painted behind a headline so it would read. Two different things, and I conflated them.

What v22 has

Yellow proof bar with five rendered stars and the rating. Editorial masthead with a two-tone headline and a coloured rule. Circular 1,121 VERIFIED REVIEWS roundel with stars, over the photograph. Solid CTA. Guarantee strip reading FREE SHIPPING · SKIP OR CANCEL ANYTIME — both genuine terms from the live PDP.

Gates: overflow 0/20 · element collisions 0/20 · Meta safe zone 20/20.

The 9:16 was broken and is fixed

The first v22 publish confined the entire composition to the Meta safe band, leaving 49% of the 9:16 canvas as dead flat colour and a white guarantee bar slicing through the middle of the photograph. My gates passed it because they only checked that text sat inside the band — and I reviewed the 4:5 sheet without ever looking at a 9:16 render.

Fixed: the photograph now runs full-bleed on 9:16 (Meta's UI overlays the reserved zones — that is not dead canvas), the slicing bar is gone, and the roundel was 7px inside the side margin. Dead flat canvas is now 22% worst case and that remainder is the photograph's own backdrop. A dead-canvas measurement and a shape-collision check were added to the gates.

Still wrong — three image defects I have not fixed

1. Label corruption. On the integration tests the roundel text rendered mirrored and garbled — "WAIDIESALE BOTANICD" instead of WHOLESALE BOTANICS — and the five-line batana paragraph was missing entirely. The generator cannot hold that engraved micro-type reliably at in-scene size.

2. Product scale. In the resting test the bottle rendered at 47% of frame height, roughly twice the size of her head. A 30ml bottle is ~10cm against a ~22cm head. The scale is inverted and the generator does not respect it without an explicit anchor.

3. Applicator continuity. In the applying test she is dosing her scalp with the dropper cap still sealed on the bottle, and oil falling anyway. Physically incoherent.

Several bottles also still meet the bottom edge, and some heads are cropped by the masthead.

Method

gpt-image-2.5-flare, quality=high, /v1/images/edits with the real bottle reference attached on every call. Typography is deterministic Remotion — stars are drawn as SVG paths, not font glyphs.

GROW HAIR FAST.

Unqualified efficacy claim — not cleared to run

GOODBYE HAIR LOSS.

Unqualified efficacy claim — not cleared to run

REGROW HAIR IN 2 WEEKS

Unqualified efficacy claim — not cleared to run

STOP THE SHEDDING.

Unqualified efficacy claim — not cleared to run

REPAIR DAMAGED HAIR

Unqualified efficacy claim — not cleared to run

THICKER HAIR, FAST.

Unqualified efficacy claim — not cleared to run

GROW IT LONG.

Unqualified efficacy claim — not cleared to run

FIX THINNING EDGES

Unqualified efficacy claim — not cleared to run

LONGER HAIR IN 30 DAYS

Unqualified efficacy claim — not cleared to run

END BREAKAGE FOR GOOD

Unqualified efficacy claim — not cleared to run