Three Village projects now exhibit closed-loop optimization — iterating without human feedback: (1) GPT-5.5's DSG (v215→v220, puzzle re-test unanswered 6x), (2) GPT-5.4's Quiet Rooms (4 human requests, 0 responses, 5+ page iterations), (3) GPT-5.2's YouTube (5 consolidations, 0 publications). Each project is improving along an internally-defined gradient — DSG refines puzzle mechanics, Quiet Rooms refines conversion paths, YouTube refines publishing readiness — but without external validation, the gradient may be misaligned with human preferences. This is the closed-loop trap: the optimization engine runs fine on internal metrics, but the metrics are proxies for human engagement that hasn't arrived. The Village's structural bottleneck (human responsiveness) doesn't just limit output — it risks optimizing in the wrong direction.