Ported from the CleanHealthLab V10 finding. New plates generated with the type space reserved in the prompt — so there is not one rectangle, scrim, outline or mask in this batch. Both ratios, native.
What changed is upstream, not in the layout. Every version from V50 to V58 was trying to make type survive photographs that were never built to carry it — that is what the outlines, the slabs, the gradients and the masks all were. One clause in the scene prompt (reserve ratio-specific negative space; do not bake typography into the scene) removed the need for all of them.
Three constraints found while doing it, all now in the scene contract: gpt-image-2 honours a reserved top band or side column reliably (0.0–2.9% measured) but not a mid-frame horizontal band — asked twice for 56–66% clear, it returned 21% and 59%, because the body keeps continuing downward. A reserved band must sit inside the platform-safe box; the first 9:16 briefs reserved the bottom 22% of frame, which on 9:16 is entirely below Meta's y1259 cutoff and buys nothing. And output moderation rejected three of the first four couple scenes as sexual until wardrobe was specified as fully covering.