Two seemingly opposite research currents, autonomous AI scientists that run discovery end-to-end, and an anxious literature demanding proofs, certificates, and error bounds, are actually in dialogue, revealing that capability is outpacing verifiability across coding, clinical reasoning, and sampling. The sharpest insight is that aggregate benchmark scores are structurally blind to entire failure classes: a model's command can be correct yet get mangled in the transport layer, tanking performance without the metric ever distinguishing reasoning failures from plumbing failures. Underneath it all is a migration of the burden of trust from the human guarantor standing outside the system to certification mechanisms built inside the machinery itself.
Topics: AI verification, autonomous research agents, benchmark blindness, formal proofs, trust certification