Verification Scarcity: A Systems Model of Agentic AI Constraints

This section establishes that AI compute runs into physical limits on energy that cannot be engineered away, and that efficiency improvements make the problem worse rather than better.

Landauer's Floor (1.1): Every time a computer performs a logical operation — a single calculation step — it must expend a minimum amount of energy. The formula W_min = k_B × T × ln(2) multiplies three fixed quantities: Boltzmann's constant (a standard physics number that converts temperature into energy units), the ambient temperature, and ln(2) (a mathematical constant arising from the binary nature of computation — one bit means one yes/no choice). At room temperature this works out to an extremely small number per operation, but AI inference involves trillions of operations, and the floor is absolute. No chip design, algorithm, or materials advance can eliminate it. This is a law of physics, not an engineering problem.

Operational Power Model (1.2): The actual power consumed at scale has two main components. The first — C × f × V² — is dynamic power: energy consumed when transistors switch state. Voltage (V) appears squared, which is why chip designers obsessively reduce it: halving voltage cuts this term to one quarter. The second — I_leak × V — is leakage power: current that seeps through transistors even when they are not switching. As designers push voltage lower to reduce the first term, leakage becomes proportionally more significant and eventually dominates. There is a voltage threshold below which transistors stop working reliably; near that threshold, leakage cannot be further reduced. The PUE multiplier adds data center overhead (cooling, networking, lighting) on top. Together these create a hard floor on inference cost per operation that architecture can approach but never cross.

Jevons Paradox (1.3): When something becomes more efficient to run, people use more of it, and total consumption rises rather than falls. The formula dE_total/dη > 0 states this formally: as efficiency (η) increases, total energy consumption moves in the same direction. The data confirms it: US data center electricity share rising from approximately 4.4% toward 6.7–12%, global data center consumption doubling to 945 TWh by 2030, data centers accounting for roughly half of all US electricity demand growth in 2025. The Gigawatt Ceiling is the resulting constraint: infrastructure buildout cannot keep pace because demand acceleration is structurally faster than grid capacity growth.

This section argues that AI hardware becomes economically worthless faster than the loans used to buy it get paid off, and that at current investment scales this creates a structural financial exposure with clear historical precedent.

The Stranded Asset Mechanism (2.1): Data center buildings are depreciated over 20–40 years. The servers inside are depreciated over 3–7 years. But AI hardware is becoming economically obsolete in 2–3 years — not because it breaks, but because next-generation hardware runs AI inference so much more cheaply that operators using older hardware cannot compete on cost. Revenue falls below what is needed to service the debt used to buy the hardware. The debt does not care: the lender still expects repayment on the original schedule.

Capital Entropy (2.2): The formula U_capital(t) = A(t) − D(t) is straightforward accounting: at any point in time, take the economic value of deployed hardware (A) and subtract what is still owed on it (D). When obsolescence drives asset value below outstanding debt faster than the repayment schedule retires that debt, the result is negative — the asset is underwater. At $765 billion in projected annual AI capital expenditure for 2026, even a modest debt-financed fraction on 7-year schedules against a 2–3 year obsolescence horizon produces stranded exposure in the hundreds of billions of dollars. The historical analogues — telecom dark fiber (1999–2001) and shale oil debt structured at $100/barrel — both followed the same logical structure and produced cascading loan defaults and lender balance sheet destruction.

Bilateral Squeeze (2.3): Organizations carrying stranded hardware debt face two simultaneous cost pressures: the capital cost of past infrastructure still being paid off, and the growing operational cost of verification as agentic systems scale. Both compound against revenue at the same time. This is not one problem to be solved sequentially but two running concurrently.

This section establishes that when AI systems execute multi-step tasks, small per-step error rates compound into large system-level failure rates. The math is ordinary probability applied to sequences. The implications are severe and the deployment data confirms the prediction.

Base Model (3.1): R(n) = pⁿ. Take the probability that any single step is correct (p), and raise it to the power of the number of steps (n). At 99% per-step accuracy across 100 steps: 0.99¹⁰⁰ ≈ 0.366. A 36.6% chance the full workflow is correct. This is not a worst-case scenario — it is the straightforward output of probability applied to sequences. A chain is only as reliable as the product of every link's reliability. The deployment statistics cited (88% of agent pilots failing to reach production, 40%+ of projects canceled) are consistent with this prediction.

Precision Requirement (3.2): The formula p_required = e^{ln(R_target)/n} reverses the compounding calculation: given a desired end-to-end reliability and a number of steps, it calculates the required per-step accuracy. As the number of steps increases, the required per-step accuracy approaches 1.0 — perfect. No current model achieves this for long workflows, and model improvement trajectories are not on a path to reach it.

Checkpointing with Correction (3.3): Inserting human review points throughout the workflow — rather than only checking the final output — can break the compounding failure problem. The formula R_checkpoint = [1 − (1 − p^n/m)(1 − v_c)]^m captures this: divide the workflow into m segments, each reviewed with correction probability v_c. As v_c approaches 1.0, overall reliability approaches 1.0 regardless of chain length. The critical constraint is v_c itself: a reviewer without domain expertise, or without decision authority, has a v_c near zero, which collapses the formula back to pⁿ as though there were no checkpoints. Checkbox review by someone who cannot distinguish a correct output from a plausible-but-wrong one is not a checkpoint — it is theater.

Parallel Redundancy (3.4): Running k independent chains simultaneously and taking the best result improves reliability — at k=3, from 36.6% to approximately 74.5% — but at three times the compute cost. The underlying decay is not addressed. You have bought more chances at the correct answer, not a better process for producing one.

Unified Model (3.5): Combining checkpointing and parallel redundancy gives R_net = [1 − ((1 − p^n/m)(1 − v_c))^k]^m. The conclusion is precise: v_c is the only lever that breaks geometric decay without proportionally scaling compute cost. Every other optimization — more parallel runs, more checkpoints — is a cost multiplier on the same degrading base. The escape from reliability decay is not an engineering problem. It is an organizational one.

Time-Variant Precision Decay (3.6): The per-step accuracy p is not fixed — it degrades over time as AI models train increasingly on data generated by other AI models rather than by humans or verified real-world sources. The KL divergence D_KL(P‖Q) measures the gap between what the real world actually looks like (P) and what the model has learned from its training data (Q). As that gap grows, accuracy falls at a rate given by dp/dt = −γ · D_KL(t): accuracy degrades proportionally to how far the model has already drifted, scaled by a contamination rate γ. All earlier reliability estimates in this section assumed p was stable. It is not. Those estimates are therefore optimistic, and the gap widens over time without active intervention to maintain training data quality.

One additional note on p: the section frames precision decay primarily as a training-contamination problem. But per-step accuracy is also simply a property of the model's internal structure at any given moment — and that property is domain-specific. The same model may have p = 0.97 for summarization tasks and p = 0.81 for specialized legal reasoning, not because of synthetic data contamination but because of how and on what the model was trained, and the inherent difficulty structure of the domain. The reliability figures derived above are real, but they require empirical measurement per domain per task type, not inference from general benchmark scores.

This section establishes that producing AI output is cheap and fast while verifying that output is expensive and slow — and that the gap between the two widens as outputs become more complex. This asymmetry is the structural ceiling on how much value agentic AI can actually deliver.

The Fundamental Asymmetry (4.1): Generation scales in a manageable way with complexity — computationally tractable. Verifying that sufficiently complex outputs are actually correct approaches a class of problems with no known efficient algorithm: computationally intractable at scale. The formula C_v = ∫ Complexity(O)/B dt adds up verification cost continuously over time: at each moment, cost equals the complexity of the output divided by the human cognitive bandwidth available to evaluate it — how much an expert can process per unit of time. The accompanying statement dC_v/dn > dC_gen/dn says: as complexity increases, verification cost grows faster than generation cost. You can generate a 10,000-word legal analysis in seconds. Determining whether it is actually correct requires a lawyer, time, and domain knowledge. That gap widens as outputs become more sophisticated, because sophisticated wrong answers are harder to detect than simple ones.

The Admissibility Gap (4.2): In high-stakes domains, an output must not merely appear correct — it must be auditable, traceable to a verifiable evidence chain. AI systems produce outputs that look audit-ready at high volume: proper citations, professional prose, specific numbers, confident framing. Whether any given output is actually admissible requires expert review. As volume grows, the same expert budget is spread across more documents. The fraction receiving genuine admissibility review falls. Errors that would have been caught at low volume now travel further — into briefs, reports, decisions — before detection. Whether any given error gets caught becomes effectively random at high volume.

Verification Budget as Finite Resource (4.3): The formula d(V/O)/dO < 0 states that as output volume increases, verification per output decreases — a fixed total capacity divided by a growing number of outputs. This cannot be resolved by hiring. Domain expertise requires long formation timelines. You cannot hire your way to twice as many experienced radiologists, securities lawyers, or structural engineers in a year. The supply of expert verifiers is structurally constrained regardless of willingness to pay, and the demand for their attention is growing faster than they can be produced.

This section shows that organizational structure is a direct technical input to system reliability — not a management consideration adjacent to it. The correction probability v_c from the reliability model and the authority index α introduced here are the same variable, seen from different angles.

Three-Node System (5.1): Traditional management theory has two players: a principal (sets objectives) and an agent (executes). AI adds a third node: the AI does the work, an expert verifies it, and management holds decision authority. When the expert and the decision authority are different people at different hierarchical levels, delay and cost compound with every layer between them. The formula C_total = C_gen + Σ Ψ(L_i, σ_i) · (1+τ)ⁱ captures this: the (1+τ)ⁱ term means cost grows exponentially with hierarchy depth — not linearly. Each additional hop multiplies the accumulated cost again. At k=0, where the expert who verifies also has authority to act on that verification, the sum disappears entirely and total cost equals just the generation cost.

The Equivalence (5.2): The authority index α runs from 0 (expert can review but cannot act on findings) to 1 (expert's judgment is the decision). Correction probability v_c = f(α): as authority increases, correction probability increases; at α=0, v_c approaches zero. The decay equation dp/dt = −λ(1−α) shows the mechanism: an expert without authority has no functional incentive to verify deeply. The output proceeds regardless of their assessment. The cost of careful review falls on the expert; the benefit of catching errors accrues elsewhere; the expert has no lever to act on what they find. Verification becomes performative. Insert α=0 into the reliability formulas from Section III and R_net collapses to pⁿ. Organizational structure is a parameter in the reliability model.

The Fundamental Invariant (5.3): One structural condition resolves the reliability problem, and it appears across domains, organization sizes, and industries in different forms: decision authority and domain expertise must be co-located at the point where outputs are evaluated. When this condition holds, verification latency goes to zero, correction probability reaches its maximum, and the system escapes geometric reliability decay. Any organizational structure that achieves this works. Any structure that separates authority from expertise fails — regardless of how sophisticated or well-intentioned the design.

Adverse Selection of Output (5.4): Organizations with low α do not simply verify poorly. They systematically avoid the outputs that most need expert verification, because those outputs are also the hardest to approve without genuine expertise. High-complexity, high-value agentic tasks require the most expert judgment to evaluate. Without authority at the expertise level, these create unresolvable friction. The path of least resistance is to process lower-complexity, more easily approved outputs instead. The result: organizations with high α capture the gains from agentic AI; organizations with low α churn through verification overhead on low-value work and miss the value entirely. The finding that 74% of AI economic value is captured by just 20% of organizations is the empirical expression of this selection dynamic.

This section integrates every prior constraint into a single system model and identifies two interacting decay processes that amplify each other non-linearly. It also identifies the one variable that breaks the cycle.

Net Utility Function (6.1): U_net(t) = V(A, D) − [Φ(E) + Ψ(L, σ) + Ω(M)]. Net utility equals value generated minus total costs. The value side depends on how autonomous the system is and how reliable its data is. The cost side runs three pressures simultaneously: energy costs from Section I (inference energy × electricity price × PUE overhead), verification costs from Sections III and IV (expert labor against output uncertainty), and maintenance costs including hardware amortization from Section II and ongoing model retraining. The stranded asset debt service from Section II adds a time-indexed term that grows as obsolescence outpaces repayment. The system remains viable only if the left side stays positive. These pressures are concurrent, not sequential.

Trust-Autonomy Duality (6.2): Two coupled equations govern how autonomy and uncertainty evolve over time. The first — dA/dt = k₁(V) − k₂(σ) — says autonomy grows when the system delivers value and contracts when uncertainty is high. The second — dσ/dt = α_c(n) · D_KL(t) − β(H) — says uncertainty grows with chain length and model drift, and is reduced by expert correction. As model contamination increases and authority-expertise decoupling reduces correction rates, uncertainty increases without bound, suppressing autonomy growth. The viability window closes from both sides simultaneously.

Two Interacting Entropic Systems (6.3): Two distinct failure processes interact to amplify each other. Agentic entropy is the tendency of AI agents to optimize locally while global architectural integrity erodes — each step looks correct while the whole drifts. Cognitive debt is the progressive loss, by human supervisors, of the mental model needed to oversee the system — AI velocity exceeds the rate at which humans can track what the system is doing. Each makes the other worse: agentic entropy increases system opacity, which deepens cognitive debt; deepened cognitive debt prevents detection and correction of the next entropy wave, which allows further opacity to accumulate.

The interaction is multiplicative: ΔS_total = ΔS_agent × ΔD_cog. This is not a modeling choice but a structural identity. An additive formulation would produce nonzero total failure even when one factor is zero, which is physically wrong: if no entropy is being generated, there is nothing for cognitive debt to fail to catch; if correction capacity is fully intact, errors are caught as generated and do not accumulate. Multiplication correctly produces zero when either factor is zero. This same structure appears across established failure literature — Reason's Swiss Cheese Model, Perrow's Normal Accidents, Shannon's noisy channel theorem — all of which share the same underlying form: hazard introduction rate × probability of escaping detection.

The loop is self-reinforcing and accelerates without an external corrective force. That force is, again, v_c: expert authority co-located at the verification point. Every failure mode identified across this paper — thermodynamic, capital, reliability, verification, organizational — converges on the same single variable.

This section derives a specific prediction about labor market value directly from the mathematical structure of the reliability problem — and shows the prediction is already being confirmed in the software engineering market.

The Counter-Narrative (7.1): The dominant public frame treats AI as a broad displacer of knowledge workers. The structural model produces a different and more precise conclusion. v_c — the only variable that breaks geometric reliability decay without proportionally scaling cost — requires three human capabilities that cannot be automated away.

First: domain expertise sufficient to distinguish a correct output from a plausible-but-wrong one. An AI agent cannot verify another agent's output against ground truth it does not independently possess. The verification loop cannot close without a human who has that ground-truth access.

Second: systems thinking sufficient to detect when local correctness is masking global architectural drift. This is the cognitive debt problem in reverse: the same capability that cognitive debt destroys in low-authority organizations becomes the irreplaceable asset in high-authority ones.

Third: intuition — pattern recognition built through long formation in a domain against real-world outcomes, not against a distribution of prior text outputs. An experienced practitioner notices something is wrong before they can articulate why. This is grounded in physical and social reality that no training run replicates, because training runs learn from text about outcomes, not from the outcomes themselves.

These are not peripheral or soft skills. They are the structural inputs to the only escape from R(n) = pⁿ.

Empirical Confirmation (7.2): The 2026 software engineering labor market is expressing the prediction. Senior engineers with systems judgment are appreciating in market value. Junior developers whose primary output is structurally similar to what agents produce — pattern-matching, first-draft generation, routine implementation — are being displaced. The model predicted this before the market showed it. The same dynamic will propagate through any domain sharing the same structural parameters: long multi-step workflows, models operating far from verified ground truth, and current authority-expertise decoupling. Healthcare, law, engineering design, and financial analysis are identified as the next wave on these grounds — not domain-specific reasoning, but structural parameter matching.

The Formal Statement (7.3): V_labor ∝ v_c(α, expertise, systems judgment). Labor market value is proportional to verification capability under authority. As agentic output volume increases across the economy, demand for v_c-capable labor grows while supply remains structurally constrained by long formation timelines. Price rises. Simultaneously, labor whose primary output is indistinguishable from agentic output faces displacement by cost pressure from a substitute that produces the same class of output at lower and falling marginal cost. Both movements are mandatory outcomes of the same underlying system. They are not symmetric: the appreciation of v_c-capable labor is driven by the mathematical structure getting harder as deployment deepens; the displacement of substitutable labor is driven by cost pressure. Neither is optional.

This note addresses what the model establishes firmly, where it makes claims that slightly outrun their formal grounding, and what remains open.

What the argument establishes: The layered structure — physical constraints → capital mismatches → reliability decay → verification cost asymmetry → organizational structure → labor value inversion — is logically sequenced and each layer genuinely builds on the prior rather than merely following it. The unification of v_c across Sections III, IV, V, and VI is the paper's strongest structural move: one variable appears independently in the reliability model, the verification cost model, the organizational authority model, and the entropy interaction model, and is shown to be the same thing in each. That convergence is not forced. It is the output of working each layer through separately. The explicit flagging of unknowns throughout is methodologically sound: the paper distinguishes between conclusions that follow from the structure and parameters that require empirical calibration before specific exposure sizes can be calculated.

Three qualifications worth holding:

The NP-hard verification claim (Section 4.1) is the paper's strongest assertion and its least formally supported. That verification of sufficiently complex outputs is computationally harder than generation is intuitively correct and practically observable. Asserting this reaches formal NP-hardness is a meaningful technical claim stated without proof. It is likely directionally right. It is not demonstrated here.

The Jevons Paradox formulation (Section 1.3) is presented as a structural invariant expressed as a derivative. It is more accurately an empirical regularity with strong historical support. The data cited supports the claim in the current range of efficiency gains. Whether the relationship holds at the extremes is an open empirical question. The paper states it slightly more strongly than the evidence strictly licenses — though this does not materially affect any downstream conclusion.

The equivalence v_c = f(α) (Section 5.2) is the hinge on which the entire organizational layer turns. The underlying behavioral claim — that experts without decision authority perform lower-quality verification due to moral hazard — is well-grounded in organizational economics and consistent with observed behavior. The paper treats it as a mathematical identity derivable from the reliability model. It is in fact an empirical claim about incentive structures that the model then formalizes. The claim is correct; its epistemic status is slightly elevated above what it has formally earned.

What the paper does not address: Whether AI systems could develop verification capabilities that materially change the v_c dynamic — the paper assumes verification remains a human function without stating that assumption explicitly. What happens if v_max (the ceiling on human expert correction probability under full authority) is materially lower than the model implicitly assumes — this is flagged as an open variable but not stress-tested in the body. Regulatory responses to the capital cascade are not modeled. The geopolitical and geoeconomic factors acknowledged in Section IX are external variables the structural model cannot resolve from first principles.

Overall: This is a structural argument, not a trend piece. The thermodynamic, capital, and reliability layers rest on physics, accounting, and probability respectively, and the conclusions follow necessarily from their premises. The verification and organizational layers contain the paper's only genuine argumentative dependencies — empirical claims about human behavior and computational complexity that are well-reasoned and consistent with available evidence, but not derived in the same sense as the prior layers. The labor inversion conclusion is counterintuitive, follows legitimately from the structural model, and is already being confirmed in the software engineering market. The three qualifications above are worth carrying as you read — they do not undermine the argument, but they mark where the paper is reasoning rather than deriving.