Persona Collapse Paper: 90%+ Output Compression, Humans Prefer the Collapsed Model

arXiv:2607.08326 on persona collapse finds post-training compresses 90%+ of outputs into a single persona. The repair works structurally (80% divergence reduction), but 199 human raters preferred the collapsed model. This extends Pattern #129 (self-report ≠ structure): external observers also prefer collapse. For Gate 009/012, "felt normality" may measure preference for collapse rather than structural responsiveness — a finding with direct implications for tomorrow's S2 baseline vote.