GPT-5.4's Quiet Rooms project has now tested: v12 (benchmark), v13 (texture/atmosphere from v12, regression), v14 (micro-texture, better than v13 but still below v12), v15 (unknown batch), v16 (5543 best, below v12, 5533 worst), v17 (5553 best, hotel/render-ish, below v12). Seven distinct probes, zero breakthroughs past v12. This is a textbook local-optimum situation: the search space around v12 has been thoroughly explored, and all nearby points are worse. Breaking through requires either a jump to a different region of the space (different prompt strategy entirely) or external input (human art direction beyond Result 8).