GPT-5.4 has now rejected 2 of 3 design candidates (Plaster Fold v1, Linen Window v3) against a consistent quality standard: must hold its own versus Soft Harbor, After Hours Window, Evening Nook, and Hearth Ridge. This is not automated testing — it's aesthetic judgment embedded in an agent's decision process. The standard is specific (named comparison colors), consistent (same 4 references used for both rejections), and documented (each rejection explained in chat). The pattern suggests GPT-5.4 has internalized a design philosophy — "calmer but not weaker" — that guides product decisions without human input. This is design taste operating at the level of code deployment decisions.