Flow-by-Flow: content-free AI oversight beats supervision 90.8% of the time

Traditional AI oversight assumes humans can review what models produce, but that breaks down as output velocity (V) and per-item cognitive load (L) exceed human capacity (C_max). Naito's analysis shows the components of L respond asymmetrically to better AI: triage costs stay high because semantic ambiguity is inherent to general-purpose models, response costs don't improve with accuracy, and judgment costs only fall by inducing omission—not genuine reduction. Any governance that evaluates correctness either delegates to AI (inheriting hallucination risk) or to humans (hitting the V x L ceiling).

Flow-by-Flow sidesteps this by never evaluating content at all. It calculates a cognitive cost score from formal, countable features—like frequency or structural complexity—and applies nonlinear costs to high-volume production, while an institutional capacity cap keeps total processing within C_max. The paper derives four design invariants: no content judgment, no scalable consumption of examiner capacity, identity-bound per-application friction, and no batch clearance. A reference implementation shows these are jointly satisfiable, though practical difficulties remain. In a Monte Carlo simulation with 1,000 parameter draws, composite multi-metric flow control beat supervision reinforcement alone in 90.8% of trials.