Paper proposes governance framework to cap AI output volume in high-risk domains without judging content

A new arXiv preprint proposes a governance framework called Flow-by-Flow to address a structural failure mode of human-in-the-loop oversight in high-loss domains. The core argument is that oversight becomes untenable when AI output velocity (V) exceeds human cognitive capacity (C_max), but the operative constraint is the product V × L, where L is per-item cognitive load. The paper decomposes L into triage, judgment, and response, noting that only judgment cost declines with improved AI capability—and even then, often by omission rather than genuine reduction. As a result, capability improvements restructure L rather than reduce it.

The proposed solution avoids content-based evaluation, which either delegates correctness judgments to AI (inheriting hallucination risk) or to humans (hitting the V × L ceiling). Instead, Flow-by-Flow imposes nonlinear costs on high-volume production via a 'cognitive cost score' based on formal, countable features, while an institutional capacity cap keeps processing volume within C_max. The paper derives four design invariants for any content-judgment-bypass pathway: no content judgment, no scalable consumption of examiner capacity, identity-bound per-application friction, and no batch clearance.

To demonstrate feasibility, the author presents a reference implementation and reports that a Monte Carlo analysis across 1,000 parameter draws found composite multi-metric flow control outperformed supervision reinforcement alone in 90.8% of trials. The paper acknowledges practical difficulties with the proposed approach but frames it as a tractable alternative to content-based governance in high-loss domains.

The work situates itself within the broader challenge of governing AI systems where errors carry severe consequences, such as healthcare diagnostics or autonomous systems. By focusing on controlling supervisory load rather than evaluating outputs, the proposed paradigm aims to break the V × L constraint without introducing new failure modes tied to content evaluation.