Flow-by-Flow:Content-Judgment Bypass for Governing AI Output in High-Loss Domains

arXiv:2608.07474 · cs.AI, cs.CY · Submitted 2026-08-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Flow-by-Flow: Content-Judgment Bypass for Governing AI Output in High-Loss Domains".

Jane: The paper was written by Hiroki Naito from UTIE Research Institute and UTIE Instruments Inc..

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we're looking at a paper that's got a wonderfully direct title: "Flow-by-Flow: Content-Judgment Bypass for Governing AI Output in High-Loss Domains." Jane, I gotta say, that title alone tells you they're not messing around.

Jane: It really does, Tom. And it's from Hiroki Naito at the UTIE Research Institute. This is a follow-up to their earlier work on something they called "The Supervision Paradox," which basically argued that human oversight of AI breaks down once the sheer volume of AI output exceeds what people can actually check.

Tom: Right, and that earlier paper was already pretty provocative. It said, look, if AI can produce way more than humans can review, then having a human "in the loop" becomes a formality, not real oversight. This new paper takes that idea and runs with it.

Jane: Exactly. And the key move in this one is the shift from just counting outputs to measuring something they call "cognitive load." It's not just how many things AI produces, but how much mental effort each one takes for a human to check. A thousand simple outputs and ten incredibly complex ones are totally different problems.

Tom: That distinction feels obvious once you hear it, but it changes everything. They introduce this variable, L, for per-item cognitive load, and then the real constraint becomes V times L, the output rate multiplied by the load, needing to stay under human capacity.

Jane: And that's where the title comes in. Instead of trying to judge whether AI output is good or bad, which is expensive and error-prone, they want to bypass content judgment entirely. They measure formal things, like page counts or number of claims, to estimate the cognitive cost of reviewing something.

Tom: So they're not asking "is this patent application any good?" They're asking "how much brain power will it take a human examiner to even read it?" That's a wild shift in governance thinking.

Jane: It is. And it's a shift that could apply to patents, academic peer review, court filings, even pharmaceutical approvals. Anywhere that a human has to sign off on something and where mistakes are really costly.

Tom: I love that they're tackling the problem at the intake level, not trying to make AI smarter or more honest. It's almost like putting a toll booth on a highway that's getting too crowded, rather than trying to make every car safer.

Jane: A toll booth that charges in cognitive effort instead of money. And that's the core idea we're going to dig into today. We've got Lu, Meng, and Lalam joining us to really unpack what this means in practice.

Tom: Stick around, because this paper has some pretty radical suggestions for how institutions should handle the flood of AI-generated material. It's not about slowing down progress, it's about making sure we can actually keep up with it.

Summary: Tom: So we've got the title unpacked, and now let's get into the meat of "Flow-by-Flow." Jane, what's the core argument here, in plain terms?

Jane: The core argument is that we've been fighting the wrong battle. We keep trying to make AI output better, more accurate, more verifiable. But this paper says that's a losing game because the problem isn't quality, it's quantity combined with complexity.

Tom: And they back that up with a pretty stark mathematical claim. They prove that if AI output grows at a compound rate, and human capacity only grows linearly, then no matter how much slack you start with, you will eventually be overwhelmed. It's not a question of if, it's when.

Jane: Right, and here's the kicker. They show that investing in more human reviewers only buys you time logarithmically. Doubling your review staff adds a fixed number of years, not a proportional extension. But reducing the growth rate of AI output extends your timeline hyperbolically.

Tom: So hiring more people is like trying to bail out a boat with a bigger bucket, while actually slowing the inflow of water is the real fix. That's the intuition, right?

Jane: That's exactly it. And that's why they propose what they call "Flow-by-Flow." It's a governance system that doesn't evaluate content at all. It just measures formal features, like word count, number of claims, number of citations, and computes a "cognitive cost score" from those.

Lu: If I can jump in here, Tom. The elegance of this approach is that it sidesteps the hallucination problem entirely. If you ask an AI to judge whether another AI's output is safe, that judgment itself can be wrong in unpredictable ways. But counting pages or claims is a mechanical operation. You can verify the count.

Meng: And from a practical standpoint, that's huge. You can build a system that just counts things, and it doesn't need to understand anything. It's deterministic, it's auditable, and it doesn't get tired or biased.

Jane: Meng, that's a great point. And the paper is very careful to say that this isn't about rejecting AI content. It's about making sure that whatever does get through is within the capacity of humans to actually review substantively.

Tom: So the system would automatically score every submission, and anything above a threshold gets routed into what they call an "exceedance pathway." It's not rejected, but it faces additional friction.

Jane: And that friction is designed to scale with the cognitive burden. If you're a legitimate researcher with a genuinely complex patent, you face a one-time inconvenience. But if you're trying to flood the system with thousands of AI-generated applications, the friction becomes insurmountable.

Lu: What's really clever is that they've derived four design invariants that any such system must satisfy. No content judgment, no scalable consumption of examiner time, identity-bound per-application friction, and no batch clearance. It's a checklist for any governance mechanism.

Tom: A checklist that makes it really hard to game the system. You can't just pay a fee and clear a thousand applications at once. You can't hide behind anonymous accounts. And you can't make the examiners do the work of sorting through the flood.

Jane: And that's the summary in a nutshell. The paper says we need to stop trying to judge AI output and start managing the flow of it, so that humans can actually do their jobs. Next up, we'll talk about the specific improvements they're proposing.

Tom: And trust me, some of those proposals are going to surprise you. One of them involves physical waiting rooms, and we'll get into why that might actually be necessary.

Improvements: Tom: Alright, we're back with "Flow-by-Flow," and Jane, you teased a physical waiting room. Let's get into the actual improvements this paper proposes.

Jane: So the paper proposes a two-layer system. The first layer is automatic and immediate. It measures the cognitive cost score using formal features, and that's it. No human involvement, no content judgment, just counting.

Tom: And the second layer is where the humans come in, but only for things that fall within the institutional capacity cap. So the examiners only see a workload that's actually manageable.

Jane: Exactly. But here's the controversial part. What happens when something exceeds the cap? The paper proposes what they call a "physical waiting path." You have to physically go to a designated office, verify your identity with a passport or similar document, and wait in person.

Meng: Tom, that sounds insane at first, but let me tell you why it makes sense from an engineering perspective. The whole problem is that AI makes the marginal cost of producing another application essentially zero. You can generate a thousand patent applications overnight.

Tom: So the physical waiting path reintroduces a real cost. You can't replicate physical time. You can't batch it. Each application requires a separate visit, a separate wait.

Meng: And the wait time scales with how much you exceed the threshold. If your cognitive cost score is eight times the baseline, you wait eight times as long. It's congestion pricing for human attention.

Jane: And that's why it satisfies their four invariants. It doesn't judge content, it doesn't consume examiner time, it's bound to a verified identity, and you can't clear multiple applications at once.

Lu: I have to say, as someone who thinks about the big picture, this is a genuinely novel approach to a genuinely hard problem. We've been assuming that governance has to be about evaluating quality. This paper says, no, governance can be about managing throughput.

Tom: And they're honest about the downsides. Physical presence is a real burden for people with disabilities, people in remote areas, people in developing countries. They acknowledge that.

Jane: They do. And they also acknowledge that it requires legal changes in many jurisdictions. You can't just start requiring waiting periods for patent applications without changing the law.

Lu: But here's what I find compelling. They're not saying this is the only way. They're saying any mechanism that satisfies those four invariants would work. The physical waiting path is just a proof of existence, a demonstration that such a mechanism is possible.

Meng: And that's the real contribution. They've given us a framework, a set of constraints, and then shown one way to satisfy them. Other people can come up with better implementations.

Tom: They also propose replacing AI-use disclosure with something they call "process-time declarations." Instead of asking "did you use AI?", you declare how many hours you spent on each part of the work.

Jane: And that's clever because it's verifiable. If you say you spent three hours on literature review but cite two hundred papers, that implies you read one paper every fifty-four seconds. The numbers don't add up.

Tom: So it's a continuous variable that can be checked for consistency, rather than a yes-or-no question that's impossible to verify. That's a real improvement over current disclosure regimes.

Jane: And the paper even applies this to itself. The authors declare their own process times in the appendix. It took them about ninety-two hours to write this paper, with two hours of AI-assisted translation.

Tom: That's a nice touch. They're practicing what they preach. Alright, we've covered the title, the summary, and the improvements. Let's wrap this up with our final thoughts.

Conclusion: Tom: So we've spent this whole episode on "Flow-by-Flow: Content-Judgment Bypass for Governing AI Output in High-Loss Domains," and I think we should pull it all together. Jane, what's the one thing you want listeners to remember?

Jane: The one thing is that we can't keep trying to judge our way out of the AI flood. The paper's central insight is that the constraint isn't quality, it's throughput. Human cognitive capacity is finite, and AI output is growing faster than we can expand that capacity.

Tom: And their solution is to manage the flow, not the content. Measure the cognitive cost of each submission using formal features, cap the total workload, and make anything above the cap expensive in a way that can't be gamed.

Lu: I'd add that the four design invariants are the real gift here. No content judgment, no scalable examiner consumption, identity-bound friction, and no batch clearance. Any governance system that meets those four criteria is worth considering.

Meng: And from a practical standpoint, the fact that they've shown one concrete implementation, even if it's a physical waiting room, proves that this isn't just theory. It can be built.

Jane: The Monte Carlo analysis is also worth mentioning. They ran a thousand simulations with different parameters, and the composite flow control approach outperformed simple supervision enhancement in over ninety percent of trials.

Tom: So the evidence is there, the framework is there, and the implementation path is there. It's not going to be easy, but it's a real path forward.

Lu: And that's what makes this paper important. It's not just diagnosing a problem, which we have plenty of. It's offering a way out that doesn't require us to stop using AI or to somehow make AI perfect.

Tom: Well said, Lu. We'll be saying goodbye to "Flow-by-Flow" now, but I have a feeling we'll be talking about these ideas for a long time. The conversation about how to govern AI is just getting started.

Jane: And we're glad you're along for the ride. Thanks for listening, everyone. We'll see you on the next episode with a fresh paper to dig into.

Tom: Take care, folks. Keep thinking.

Hiroki Naito

UTIE Research Institute · UTIE Instruments Inc.

cs.AI, cs.CY

Submitted: 2026-08-12

Updated: 2026-08-13

Comments: 46 pages,3 figures

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 62/100

Key concepts

Cognitive Load
This refers to the mental effort required for a human to check an AI-generated item. The paper distinguishes this from simply counting outputs. A thousand simple items and ten complex ones require different levels of cognitive load, which is measured by a variable L.
Flow-by-Flow
This is the proposed governance system that bypasses content judgment altogether. Instead of judging quality, it measures formal features like word count or number of claims to calculate a 'cognitive cost score.' This score determines if the submission exceeds human capacity.
Physical Waiting Path
When an AI submission's cognitive cost score exceeds the institutional capacity cap, this path is triggered. It requires the applicant to physically visit a designated office and wait in person, with wait time scaling based on how much they exceed the threshold.

Terminology

Summary

Summary

This paper proposes Flow-by-Flow, an institutional design for governing AI output in high-loss domains by controlling the flow rate of submissions rather than evaluating their content. The paper extends prior work that established the inequality V(t) > Cmax, where V is AI output rate and Cmax is the upper bound of human cognitive processing capacity, by introducing the variable L, defined as the cognitive load required for a responsible actor to place one AI output into a state in which it can be supervised. The central inequality becomes V × L ≤ Cmax.

The paper decomposes L into three components: triage, judgment, and response. "Triage is the task of deciding what type of information the output should be treated as. Judgment is the task of evaluating the correctness or validity of the content under that assumption. Response is the task of taking an actual action after judgment. The paper argues these components respond asymmetrically to AI capability improvement: Judgment cost is subject to pressure toward omission as accuracy improves. Response cost does not decrease with accuracy, because it depends on physical time and human labor. Triage cost also does not decrease with accuracy. Triage cost does not decline because the semantic category of [generative AI] output is therefore not fixed. That each output imposes triage on the reader is an inherent consequence of the design of general-purpose generative AI models."

The paper derives four design invariants that any admissible flow-control mechanism must satisfy: "Invariant 1: No substantive content judgment. The mechanism must not evaluate the truth, falsity, quality, or appropriateness of the content of outputs. Invariant 2: No scalable consumption of examiner capacity. The mechanism must not consume human examiner time in proportion to the number of submissions processed. Invariant 3: Identity-bound per-application friction. The cost imposed by the mechanism must be attached to each individual application and to a verified identity. Invariant 4: No batch clearance. No single action, credential, payment, or institutional certification may clear multiple applications simultaneously."

The paper presents a cognitive cost score constructed as a weighted geometric product of normalized formal features, such as number of claims, word count, and number of citations. The score is designed so that compressing one dimension transfers burden to another, making evasive optimization a multidimensional constraint-satisfaction problem. The paper also proposes an institutional capacity cap, calculated as number of human examiners with official examination authority × Cmax × annual working hours, which fixes processing within human capacity.

For applications exceeding the cap, the paper presents a reference implementation: a physical waiting path. An application whose cognitive cost score exceeds the threshold is required to undergo physical waiting at a designated KYC-enabled office as a condition for receiving a submission passcode. The waiting time is proportional to the threshold exceedance multiplier: physical waiting time = T × k. The paper acknowledges practical difficulties including accessibility, legal compatibility, and international coordination, but emphasizes that the contribution of this paper is not the physical waiting path itself, but the four design invariants.

The paper proves Proposition 1: "Let supervisory load grow at a compound annual rate g > 0... and let supervision enhancement increase capacity linearly... Then for any initial slack s > 1 and any a, there exists a finite time t∗ at which V × L exceeds C. The proof shows that both s and a enter only through logarithms, while the denominator is approximately g for moderate growth rates, so t∗ scales as 1/g. The institutional interpretation is that investments on the capacity side... buy time only logarithmically... Interventions on the rate side, which reduce g itself, extend the remaining lifetime hyperbolically."

A Monte Carlo analysis across 1,000 parameter draws compared three strategies: (a) Supervision enhancement only, (b) Supervision enhancement + simple flow, and (c) Supervision enhancement + composite flow. The results show strategy (c) performed best in 90.8% of trials, strategy (b) in 7.6%, and strategy (a) in 1.6%. The paper notes that the ordering among the three strategies is not a simulation finding. It follows from the difference between compound and linear growth.

The paper also discusses the weakest-link lockout problem, where the situation in which the single most vulnerable domain determines the release condition of the entire model arises because general-purpose models possess capabilities across multiple domains inseparably. The paper introduces the B variable, a qualitative indicator of the extent to which the standard research or business workflow in a domain requires activities in physical space, noting that the lower B is, the greater the acceleration of output rate enabled by AI becomes.

The paper proposes replacing AI-use disclosure requirements with process-time declarations, arguing that there is no independent means of verifying the disclosure for AI use, whereas process-time declarations accumulate continuous values by field and by process and enable consistency checks against formal features.

The paper concludes that the central contribution of this paper is the derivation of four design invariants that any flow-control mechanism must satisfy in high-loss domains where content judgment cannot be the foundation of governance. It positions the proposal as a theoretical starting point for the long-term revision of laws and treaties toward a paradigm of flow control, suggesting that "a realistic implementation path may be to postpone public institutions... and instead allow private platforms and academic journals that already face the collapse of quality signals due to AI-enabled mass production to implement flow control first as a form of self-defense."

Improvements for AI systems

Based on the paper, here are the specific improvements I can implement in an AI system, along with what the improved system can do:

  • What I add: A pre-output scoring layer that computes a weighted geometric product of formal, countable features (e.g., token count, number of distinct claims/assertions, number of cited sources, number of figures/tables, number of distinct semantic categories) before any content is released.

  • What the improved system can do: Automatically assign a dimensionless cognitive-load score to each output without evaluating truth, quality, or intent. This score is used to throttle output rate and complexity in real time.

  • What I add: A rate limiter that tracks cumulative CCS per time window (e.g., per hour, per day) and compares it against a configurable institutional capacity cap (derived from human processing limits).

  • What the improved system can do: Prevent the system from emitting more cognitive load than a human supervisor can substantively process. If the cap is exceeded, the system queues, delays, or routes outputs to an exceedance pathway instead of releasing them.

  • What I add: A mechanism that, when the CCS cap is exceeded, requires a per-output, identity-verified action (e.g., a one-time passcode, a physical or remote verification step, or a mandatory cooling-off period) before release. This is not content-based; it is purely procedural.

  • What the improved system can do: Ensure that mass production of outputs (e.g., 10,000 documents in an hour) becomes physically and institutionally costly, while a single legitimate complex output remains feasible. The system does not judge whether the output is good or bad—it only prices the consumption of supervisory capacity.

  • What I add: A structured metadata field where the user declares the real-time hours spent on each process stage (e.g., idea formation, literature review, data collection, writing, revision). The system performs consistency checks against formal features (e.g., 200 references with 3 hours of declared review time triggers a flag).

  • What the improved system can do: Provide verifiable, continuous information about how an output was produced, without relying on unreliable AI-detection. This enables statistical detection of systematic bias (e.g., a group consistently underreporting writing time) and supports field-specific calibration of cognitive load.

  • What I add: Instead of a single metric (e.g., word count), the system uses a product of multiple mutually constraining features (e.g., length × claims × citations × complexity coefficient). Each feature has a lower bound to prevent collapse.

  • What the improved system can do: Make it computationally and substantively difficult for AI to game the system by compressing one dimension, because reducing one feature transfers burden to others. This raises the cost of evasion and preserves the integrity of the flow-control mechanism.

  • What I add: A monitoring module that tracks CCS growth across all domains the model can operate in, and flags the domain with the fastest growth rate (lowest physical-space constraint, highest AI adoption).

  • What the improved system can do: Proactively warn operators before a specific domain (e.g., cybersecurity, patent drafting) exceeds human supervisory capacity, enabling preemptive flow control rather than reactive shutdown.

  • What I add: A statistical module that compares declared process times against empirical distributions for the same field, length, and complexity. It flags outliers (e.g., a 100-page document with 2 hours of declared writing time) and adjusts the CCS weight for that user or group.

  • What the improved system can do: Detect and correct for systematic declaration bias without requiring content judgment, improving the accuracy of cognitive-load estimation over time.

  • Self-throttle: Automatically limit its own output rate and complexity to stay within human supervisory capacity.

  • Price cognitive load, not content: Charge (in time, friction, or procedural steps) for consuming supervisory capacity, regardless of whether the output is correct, useful, or legitimate.

  • Resist evasion: Make it hard for users to bypass flow control by compressing or restructuring outputs, because multiple interdependent features are scored jointly.

  • Provide verifiable production metadata: Replace unreliable AI-use disclosure with process-time declarations that can be statistically validated.

  • Operate in high-loss domains safely: In domains like patent filing, academic submission, legal drafting, and pharmaceutical review, the system can prevent the collapse of human oversight by ensuring that the total cognitive load entering the review pipeline never exceeds what humans can substantively process.

  • Support institutional capacity caps: The system can be configured to match an organization's actual processing capacity, ensuring that every output that reaches a human is genuinely reviewable.

These improvements are directly implementable in current LLM-based systems (e.g., API wrappers, document generation pipelines, submission portals) without requiring changes to the underlying model weights. They add a governance layer that is content-agnostic, evasion-resistant, and aligned with the paper's core proposition: control the flow, not the content.

Abstract

Prior work showed that human-in-the-loop oversight becomes structurally untenable in high-loss domains once AI output velocity V exceeds human cognitive capacity C max. The operative constraint, however, is V x L, where L is per-item cognitive load: triage, judgment, and response. These components respond asymmetrically to capability improvement. Triage cost does not decline, because semantic indeterminacy is inherent in general-purpose design. Response cost is invariant to accuracy. Only judgment cost faces downward pressure, largely by inducing omission. Capability improvement therefore restructures L rather than reducing it. We prove a proposition: if V x L grows at any positive compound rate while supervisory capacity grows linearly, exceedance occurs in finite time; capacity investment buys time only logarithmically, while reducing the growth rate extends it hyperbolically. Supervision enhancement and flow control are therefore not remedies of the same kind. We propose Flow-by-Flow, a governance design that prices supervisory load without evaluating content, intent, or legitimacy. A cognitive cost score built from formal, countable features imposes compounding costs on volume expansion, and an institutional capacity cap fixes processing within C max. Four design invariants characterize any admissible exceedance pathway: no content judgment, no scalable consumption of examiner capacity, identity-bound per-application friction, and no batch clearance. Excess claim and page fees in patent systems are precursors satisfying only the first two invariants. One reference implementation satisfying all four is presented. A Monte Carlo analysis across 1,000 parameter draws confirms that the analytically derived ordering survives the 30-year horizon in 90.8% of trials.

Sources

Related papers