Part V - The Material Constitution of the AI Era

Compute, Capacity, and the Constitutional Limits of Governance

Part IV mapped how rules travel and who shapes their movement. Part V shifts from the linguistics of governance to its physics — from the frameworks through which rules spread to the material foundations that determine what can be built, who can build it, and how far their authority can reach.

Traditional authority rests on rule primacy: whoever controls the rules controls the system. Frontier AI introduces a different structural logic — capacity primacy — in which control of computational capacity conditions which rules can be made effective and, increasingly, who can set the technical defaults that become embedded throughout the stack. A state without sufficient compute cannot train frontier models; a state that cannot train models cannot shape
the defaults that propagate through the mechanisms described in Part IV.^[1]

This shift completes the manuscript's analytical arc. Law provides the language of permission. Accounting provides the grammar of value. The Incorporation Heuristic identifies which rules travel. Part V identifies the boundary condition — the physical capacity without which none of them execute. The framework is now complete: from grammar to design to physics.

The Physics, Made Explicit
The physics of frontier AI are material, measurable, and increasingly constraining. Training a state-of-the-art model can require more than 100 megawatts of continuous power — the electrical load of a small city — with some forecasts projecting multi-gigawatt-scale clusters before the decade closes. Frontier-scale training runs can cost tens of millions to billions of
dollars in compute alone.^[2]

The semiconductor supply chain that makes such runs possible passes through a small number of critical nodes worldwide, with leading-edge fabrication concentrated in a single cluster of facilities in Taiwan.^[3]

The global talent capable of designing and operating these frontier models remains a narrow pool — a subset of the broader AI workforce with direct experience at frontier scale. And the
capital required to sustain a frontier programme exceeds the annual R&D budgets of many nation-states.^[4]

This is not software. It is an industrial ecosystem — one whose resource requirements many countries cannot meet independently, and whose constraints increasingly define the outer
limits of what governance frameworks can reach.

The Capacity Challenge — Especially Outside Major Powers
The capacity demands of frontier AI impose constraints that most states cannot overcome through regulation alone. Unlike earlier waves of the internet and digital infrastructure — which could be distributed and adopted by almost all countries — frontier AI depends on concentrated high-performance computation with material limits: training clusters, inference hosting, data-centre scale, and the capital required to build and maintain them. For countries outside the primary compute centres, the constraint is real.

A country without meaningful training capacity has limited ability to shape the underlying assumptions embedded in the models it depends on. A country unable to host inference at
scale cannot consistently apply its own rules to the decisions those models produce. Authority shifts toward those who control the underlying capacity — not by design, but by material necessity.

The Incorporation Heuristic provides pathways for participation within this asymmetry. But incorporation does not dissolve it. The resource intensity of frontier AI remains a binding constraint on national agency — one that governance frameworks must accommodate rather than assume away.

AI Physics — The New Seat of Power
Cloud providers, AI labs, semiconductor ecosystems, and energy networks provide not just services but foundations that shape rulemaking outcomes. Their design decisions set the
practical boundaries within which public authority operates, as constitutional provisions set the limits of legislative power.

When a hyperscaler decides which regions receive GPU allocation, it shapes, in practice, which countries can exercise AI autonomy. When a chip designer sets export thresholds, it draws compute boundaries as consequential as many treaties. These are not commercial decisions with incidental policy effects. They are governance decisions expressed as engineering choices — accumulating into the material constitution of the AI era.

From Capacity to Legitimacy — The Convergence of All Prior Gaps
The legitimacy crisis of the AI era is not merely a new category alongside the Visibility Gap, the Rule-Execution Gap, the Categorical Gap, and the Measurement Gap identified in Parts I and II. It rests on their material foundation.

Each of those gaps — the inability to see, to act, to classify, or to measure AI systems — is ultimately a capacity gap. States that cannot support machine intelligence and cognition at scale cannot readily close any of them. The crisis is therefore not only ideological. It is the point where prior failures converge on a single physical constraint.

The Three Constitutional Questions of Part V
Part V is organised around three constitutional questions.

First, what are the new assets of power? Compute, data, and models are becoming the physical foundations on which authority is exercised — and their distribution helps determine which countries possess meaningful AI autonomy and which do not.

Second, what are the hard constraints? Five material boundaries — energy, semiconductors, capital, talent, and time — limit what any governance system can achieve, regardless of institutional design.

Third, what comes next? If control is bounded by physics and mediated largely by private firms, what institutional forms can build or restore democratic accountability for systems that currently operate beyond democratic reach?

These questions define the constitutional terrain of the AI era.

The Stakes — The Reality Every Country Must Confront
AI is more than a policy challenge. It is a general-purpose cognitive infrastructure with material resource requirements that cannot be wished away. Its diffusion across law, finance, medicine, logistics, defence, education, and identity systems will continue even when institutional readiness lags.

The asymmetry is material and time-sensitive. Countries that internalise capacity constraints will design within them. Those that treat AI as a purely regulatory challenge will encounter
those constraints only in retrospect, as their rules collide with physical limits they failed to anticipate. And countries that cannot power AI will encounter machine decisions they cannot
shape, supervise, or constrain.

The window for institutional design remains open — but narrowing under the same material constraints this book has identified. Governance without capacity is commentary. Capacity
without governance is power unchecked. The challenge of the AI era is to hold both in tension
— and to build institutions adequate to constraints no previous generation of political designers has confronted at this scale.

I. The New Assets: Compute, Data, Models

Frontier machine intelligence is defined by three properties that no previous industrial system has combined. It shows extreme capital indivisibility — you cannot train half a frontier model. It exhibits radical supply-chain concentration — leading-edge chip fabrication capacity remains confined to a handful of global facilities. And its progress is cumulative and path-dependent — each completed training run generates advantages that cannot easily be unwound.

From these properties emerge the three foundational assets of the AI era: compute as the engine, data as the fuel, and models as the territory.

A. Compute — The Engine

Compute is the capacity to transform energy into probabilistic inference at scale. It is not merely a resource. It is a national capacity.

The mechanism is straightforward but carries profound implications. A hyperscaler, following a regulatory dispute, reprioritises GPU allocation away from a region. The decision is commercial in origin; its consequences are structural. Countries lose access to training clusters. Research labs suspend experiments. Startups miss product deadlines. Public agencies dependent on cloud inference see throughput collapse. What appears as operational reprioritisation is, in practice, a reallocation of national capacity.

Compute shapes the full lifecycle of AI. Training capacity influences the foundational assumptions of models; inference capacity determines who can operate them; alignment
capacity determines who can control their behaviour. A country without compute faces real limits on its ability to control the AI systems operating within its borders — not merely to regulate them on paper, but to exercise operational authority over the decisions machine intelligence produces.

B. Data — The Fuel

Data is generated — by people, by institutions, by societies, and by civilisation itself. It is path-dependent, nationally situated, and politically charged. And the fundamental truth is unavoidable: many countries cannot independently generate the volume and quality of data required to train or align domestic models.

The asymmetry runs deeper than population size. Industrial economies generate substantially more high-value data than developing economies — not only in volume but in structural composition, because industrial applications, medical archives, legal documentation, and financial flows carry far more signal than consumer clicks or social-media traces.

And data carries norm encoding: it shapes a model's behavioural priors — what it learns to treat as normal, acceptable, risky, or harmful. This is why data is not the new oil. Oil is fungible. Data is non-fungible capital — the memory of a civilisation and the foundation of its future cognition.

Data asymmetry therefore produces cognitive disenfranchisement: the systematic exclusion of a society's norms, harms, and institutional patterns from the decision outcomes of AI models that increasingly mediate its economic and institutional life. Unlike political disenfranchisement, which can be remedied by extending the franchise, cognitive disenfranchisement requires the prior existence of the data that makes a society legible to machine learning. Societies absent from training corpora are not underrepresented — they are unweighted. Their harms are invisible. Their citizens become statistical outliers to the systems that increasingly determine access to credit, healthcare, mobility, and public services.

Data is not simply fuel. It is representation — the pathway by which a civilisation appears, or fails to appear, in the intelligence that its models generate and enact.

C. Models — The Territory

Frontier models embed decision boundaries shaped by mathematical constraints that determine, for any given input, what outputs are possible, probable, or prohibited. Whoever
shapes those constraints shapes the universe of decisions available to every downstream user.
Models are not merely products. They are territory — cognitive landmass.

Deployment of a foreign-trained model entails downstream inheritance. The foundational training decisions — data selection, alignment regimes, safety constraints — are made once
by an upstream actor and inherited by every downstream user. Countries that deploy foreign-trained models accept machine behaviour shaped by upstream priors, alignment choices, and design constraints unless they possess the capacity to meaningfully influence or adapt those defaults. This is more than technical dependency. It is the adoption of behavioural boundaries encoded elsewhere.

The policy consequence is direct. At training time, models absorb rules — the behavioural constraints embedded through data selection, alignment, and safety design. At inference time,
they enforce them — defining the acceptable boundaries for institutions, the decision logic of markets, and the informational landscape through which societies navigate. Once deployed, models become the operative carriers of these constraints within their deployment domains. Every downstream user operates within boundaries set upstream.

D. The Triad in Practice — A Single Scenario

Consider Nigeria — a country attempting to build a national model trained on African languages, legal corpora, and institutional data. The ambition is real; the constraints are equally real.

On the compute side, a competitive training run at frontier scale demands GPU capacity on the order of thousands of accelerators — a scale that exceeds the training-cluster capacity currently available within the country. Procuring such hardware requires navigating supply chains that run through Taiwan, Malaysia, South Korea, and U.S. hyperscalers — a dependency structure that most African states cannot meaningfully reroute independently.

On the data side, Nigerian legal texts, Hausa and Yoruba corpora, and financial records exist, but not in the digitised, structured, and cleaned form that frontier-scale training requires. The
raw material is present; the industrial refinement pipeline is not.

On the model side, even if Nigeria were to secure compute and refine its data, the resulting system would still inherit architectural priors from upstream frameworks, toolchains, and alignment methods developed elsewhere. Without the capacity to adapt or retrain these components at scale, the country would deploy a model whose behavioural boundaries — what it treats as plausible, safe, risky, or prohibited — are shaped materially by external design choices. The cognitive territory would be national in aspiration but foreign in origin.

E. Transition — From Assets to Constraints

Compute, data, and models form the material foundation of AI. But recognising their importance does not answer the harder question: how many countries can assemble these assets at scale — and what happens to legitimacy when access to machine intelligence is globally uneven.

These assets are not evenly distributed, and they are not independently obtainable. Each rests on supply chains, energy systems, capital flows, talent pools, and time horizons that most countries do not control. National autonomy meets physics; ambition meets industrial reality.

The next section examines these dependencies — and the hard constraints they impose on states navigating AI industrialisation under conditions they did not choose.

II. The Legitimacy Crisis — The Widening Gap Between Statute and AI Governance

Industrial-era legitimacy rested on four stabilisers: consent, representation, procedure, and rights. These held because public authority was executed by human agents, institutions controlled the means of enforcement, and rules could be applied through established processes.^[5]

The AI era introduces a fifth — and materially prior — stabiliser: operational capacity. This is the ability of states to convert resources into machine-executed decision power at scale, and to maintain meaningful control over how that power is applied across institutions and society.

Without sufficient capacity, the classical stabilisers remain formally intact but lose practical force. AI systems increasingly execute decisions that were previously made by human agents
within institutional frameworks — and states that cannot govern how those decisions are made, applied, and corrected find traditional instruments of accountability operating at a growing distance from the systems they are meant to constrain.

The pace of AI advancement is outstripping the adaptation cycles of most existing institutions, widening the gap between authority written in statute and authority realised through the ability to direct and control AI systems in practice.

A. The Shift of Authority — From Institutional Control to AI-Driven Execution

Public authority that once flowed through institutions designed for democratic accountability is increasingly exercised through AI-mediated decisions that shape rulemaking and rule keeping in real time. The actors who control frontier models and agents gain influence over which rules can be enforced, how they are interpreted, and what forms of oversight remain possible at scale — not through elections or public mandate, but through the operational fact that AI now mediates how rules are executed.

Alignment and safety teams inside a handful of companies make design choices that determine how AI systems behave across vast numbers of interactions — which outputs are encouraged, restricted, or blocked. In practice, these choices function as large-scale norm-setting at a reach and pace that few legislatures or courts can match. The processes through which these decisions are made remain only partially subject to parliamentary debate, judicial review, or public accountability.^[6]

Operational authority is migrating into systems built for speed, scale, and computational advantage. A state whose instruments cannot extend to monitoring and guiding machine-executed decisions does not lose formal authority — but it loses the capacity to govern how power is exercised within its borders. The State continues to hold formal power on paper; AI systems increasingly shape daily life in practice.

B. The Five Non-Substitutable Constraints

The AI era imposes five non-substitutable constraints that determine whether a country can govern AI on its own terms: energy, semiconductors, capital, talent, and time. Each operates
independently but locks together with the others — and they do not substitute for one another.
Money cannot compensate for missing chip access; skilled engineers cannot overcome a lack of power; and time lost to earlier capability cycles cannot be recovered once others have moved ahead. A shortfall in any single constraint limits a state's ability to build, train, or control advanced AI systems, regardless of its strengths elsewhere.

Crucially, these constraints do not determine whether a country can use AI. Most nations can
— and should — adopt AI through imported models, cloud services, or API access. What the constraints determine is whether a state can exercise control at the machine level, where
decisions are actually executed. A country facing binding limits will rely on external AI systems — and in doing so cedes influence over how rules are interpreted, enforced, and adjusted in real time.

The choice is not between AI and no AI. It is between different modes of participation — each with its own trade-offs, dependencies, and opportunities for agency. The material constraints of the AI era define the boundaries of that choice: when key constraints are missing, countries are pushed toward more dependent forms of operation, while only those with all five at sufficient scale can exercise machine-level control. These conditions do not eliminate agency, but they shape the strategic space within which it can be exercised.

C. The Gradients of Operational Agency

Scarcity in the core inputs of AI does more than limit what countries can build. It produces the same permission gradient identified in Part IV — but grounded in material capacity rather
than regulatory propagation. Those who build and run frontier models shape system behaviour at its foundations. Those who deploy or regulate these systems operate within boundaries set by the builders. Most governments and institutions apply systems whose deeper logic is defined elsewhere.

A parallel gradient runs through the rules inside these systems. Public rules — laws, regulations, and court decisions — move at institutional pace. Private rules — contracts, terms of service, audit requirements — evolve iteratively, revised by the same actors who control the underlying models and compute. The rules that most directly shape AI behaviour are increasingly written and updated outside the public sphere.

Reconnecting democratic authority to these systems requires more than access to models. It requires access to the legal and operational mechanisms that determine how those models
behave in practice — and that access is currently concentrated among a small number of actors whose authority derives from capacity, not mandate.

Countries without these inputs are not excluded from AI — but they navigate a landscape where influence is unevenly distributed and the cost of delayed action compounds. The urgency lies not only in adopting AI, but in securing the mechanisms that determine how it behaves.

D. Legitimacy Debt — The Silent Failure Mode

The permission hierarchy rarely announces itself. Democratic institutions continue to function — legislatures legislate, regulators regulate, courts adjudicate. Yet the gap between what institutions believe they govern and what they can actually influence widens with each new model generation, each expansion in agent capability, each increase in compute deployment.

Legitimacy Debt accumulates when formal responsibility and real decision authority drift apart. Like technical debt, it is not recorded when it forms — it remains hidden until the system is under strain. Institutions are held accountable for outcomes they cannot reliably shape at the machine-execution layer. The gap between attributed responsibility and operational control grows silently.^[7]

This produces a Misleading Dashboard Effect. Monitoring tools continue to signal compliance and control even as actual system behaviour moves beyond institutional reach — a dynamic foreshadowed by the cockpit analogy in Part I, where surface indicators remain calm while underlying misalignment deepens. The dashboard stays green; the execution layer moves on. When failure arrives, it is often sudden and nonlinear — not because institutions are inattentive, but because the operational mechanisms have already shifted beyond their grasp.

E. What Legitimacy Collapse Looks Like

Legitimacy collapse in the AI era is unlikely to resemble revolution. It is more likely to take the form of institutional irrelevance. A ministry publishes an AI ethics framework. It circulates through international forums. It appears in the country's Universal Periodic Review. Yet few global firms operationalise it, and no dominant compliance template incorporates it. Deployed systems continue unchanged.

The framework enters the documentary archive but never reaches the machine. The ministry continues to regulate. The minister continues to speak. Yet citizens interact daily with AI
models and agents their government cannot shape, audit, or align.

Institutional formalities persist while decision authority migrates. The throne remains. The occupant has vacated.

F. The New Test of Legitimacy — Consent versus Execution

In the industrial era, legitimacy centred on consent: do the people agree? In the AI era, legitimacy increasingly also depends on execution: can the State execute? Both principles
remain essential. The crisis emerges not because one displaces the other, but because system stability now depends on variables the two optimise differently.

Democratic legitimacy prioritises consent — deliberative, procedural, inclusive. Operational legitimacy prioritises execution — fast, reliable, technically coherent. No widely adopted
institutional framework has yet reconciled both at system scale.

Institutions optimised for consent struggle under inference-speed decision cycles. Institutions optimised for execution concentrate authority beyond classical consent mechanisms. This is
the defining constitutional tension of the AI era.

G. The Five Design Conditions for AI Legitimacy

Any governance framework capable of sustaining legitimacy in the AI era must satisfy five design conditions simultaneously. It must preserve output, since legitimacy cannot survive
persistent performance degradation. It must restore input, ensuring that democratic mandates reach executable decisions rather than remaining declarative policy. It must bridge the velocity gap, allowing institutional processes to interoperate with machine-speed execution. It must enable accountability, keeping decision authority observable, attributable, and contestable at system speed. And it must enable exit, allowing individuals and institutions to leave a regulatory arrangement without prohibitive costs, informational lock-in, or dependence on a single model's reasoning system.

As with the Incorporation Heuristic, these conditions operate multiplicatively rather than additively. A framework that preserves output but cannot restore input accumulates Legitimacy Debt regardless of performance. An arrangement that enables accountability but cannot bridge the velocity gap induces regulatory delay incompatible with machine-speed systems. A design that enables exit but cannot preserve output fragments the very system it seeks to govern.

In multiplicative systems, a single zero collapses the product. Legitimacy is no exception.
All five conditions must be satisfied simultaneously — a standard that existing institutions were not designed to meet. The next section — Code After — sets out the architecture proposed to meet it.

III. Code After: The Operating Manual for the AI Era

Part V concludes with the architecture itself — not a vision or a manifesto, but a constitutional design proposed for the AI era, grounded in material constraints and derived from the analytical framework developed across the preceding Parts.

The manuscript has traced a continuous arc. It began with the State's inability to see or execute the AI economy it governs, moved through the dual governance languages that define economic reality and the Translation Layer that now mediates between them, and mapped the vertical stack and the G3 fracture it produces. It then followed the Incorporation Heuristic and the propagation mechanisms that determine which rules reach the systems that actually execute, before arriving at the material constraints and legitimacy conditions that bound what any institutional response can achieve. What remains is the institutional design those constraints imply.

The Third Architecture of Governance introduces three mechanisms — the Sovereign API,
Parametric Democracy, and the Protocol of Federation — that together form a separation-ofpowers arrangement for machine-executed AI ecosystems.

The Sovereign API provides an institutional interface linking democratic authority to machine-executed systems. Parametric Democracy calibrates democratic input into executable constraints. The Protocol of Federation enables incompatible normative foundations to coexist without requiring internal value harmonisation.^[8]

Together, these mechanisms satisfy the five multiplicative design conditions — preserving output, restoring input, bridging the velocity gap, enabling accountability, and enabling exit
— not through trade-offs but through functional separation across distinct levels of authority.
The design does not prescribe political preference. It derives institutional necessity.

A. The Sovereign API

The Sovereign API is an institutional interface specification. It defines the channels through which democratically legitimate authority shapes machine-executed outcomes. On the input
side, it receives statutes, regulations, thresholds, and jurisdiction-specific obligations. On the output side, it renders them as constraint schemas, parameter ranges, prohibited and required
behaviours, and documentation requirements that can be embedded directly into models, workflows, and systems. The Sovereign API does not execute rules; it translates them into a
form that technical systems can consume. It does not replace the Incorporation Heuristic; it reshapes the inputs that determine its results.

A statutory non-discrimination norm — "no automated decision may consider protected attributes except where expressly permitted by law" — illustrates the mechanism. The API
would not adjudicate fairness or resolve the underlying substantive inquiry; instead, it would enforce an interface requirement that any model used for an eligibility decision declare whether protected attributes or recognised proxies are present in its inputs, and block execution unless a lawful basis or statutory exception is supplied.

In this sense, the Sovereign API is not a technical instrument but a constitutional one: a structured pathway through which legal authority is expressed as operational constraint,
ensuring that the rules governing AI systems originate in public law rather than private construct.
This does not reduce anti-discrimination law to feature disclosure alone, but it does translate one component of the legal rule into an ex ante constraint that the deployment environment
must satisfy. In this way, a normative legal value becomes a binding precondition for technical execution, without requiring the State to modify the underlying model. This logic generalises.

The adjustment the Sovereign API imposes is visibility: deployed AI systems must make their behaviour legible — observable, verifiable, and auditable through institutionally authorised channels. Under this specification, visibility becomes a constitutional requirement rather than an operational preference.

The first adjustment establishes visibility as institutional design. By requiring legibility as a condition of execution, the Sovereign API ensures that downstream normative constraints —
whether grounded in non-discrimination, due process, or sector-specific mandates — operate on systems whose behaviour can be inspected and justified. Visibility is not an ancillary safeguard; it is the precondition that allows democratically legitimate authority to govern machine-executed outcomes without intervening in the model's internal mechanics.

The second adjustment establishes workability. The Sovereign API translates democratic input into constraints that AI systems can execute — rendering legislative intent in machine-interpretable form. By reducing normative ambiguity, it increases the workability of public rules across the national AI ecosystem.

The third adjustment establishes necessity. Through federated interface standards — developed further in the Protocol of Federation — the Sovereign API allows national rules to become embedded in the economic network itself. When multiple sovereigns adopt compatible interface specifications, compliance ceases to be a bilateral regulatory demand and becomes a condition of network participation.

Firms that align with common interfaces maintain uninterrupted access; those that diverge encounter increasing friction through the economics of network position rather than diplomatic sanction. Under these conditions, necessity shifts from market size to network membership.

The Sovereign API therefore operates at the interface — the constitutional hinge linking democratic consent to machine execution. Rules that once failed to influence machine behaviour gain operational weight when they are institutionally observable, expressed as executable parameters, and reinforced through network-level necessity.

B. Parametric Democracy — Governing at Machine Speed

Citizens cannot directly govern AI systems operating at neural-network scale. Frontier models operate at a complexity, speed, and scale that text-based legislation cannot specify in detail. Parametric Democracy preserves democratic legitimacy by shifting regulation from rule authorship to constraint calibration. Citizens do not vote on executable logic; they vote on the trade-offs that executable systems must respect.

The Thermostat Model illustrates the principle. A citizen does not need to understand thermodynamic engineering to set a temperature, just as they do not need to understand neural
networks to express calibrated preferences across familiar tensions — privacy and convenience, innovation and safety, efficiency and employment, autonomy and oversight. The analogy is not about simplification; it is about constructing the right abstraction boundary. The interface translates value preferences into parameter constraints, while the model manages the execution dynamics beneath them.

This raises the central institutional question: who controls the interface? The legitimacy of
Parametric Democracy depends on which parameters are exposed, how they are framed, and what decision ranges are made available. If the design of those parameters is controlled by
the actors the process is meant to constrain, the arrangement reproduces the legitimacy deficit it was created to resolve. The safeguard is institutional pluralism in interface design —
competing parameterisations, transparent selection procedures, and independent oversight of the constraint-definition framework.

Parametric governance is not speculative. Its functional components already operate in isolation. Quadratic voting demonstrates how preference aggregation can be expressed as
calibrated intensities rather than binary choices.^[9]

Reinforcement learning from human feedback and Constitutional AI show how value constraints can be encoded into model behaviour.^[10]

Central banks illustrate how mandates can be translated into parameterised policy instruments. Platforms such as vTaiwan demonstrate how digital deliberation can structure public input on contested questions.^[11]

Parametric Democracy integrates these mechanisms into a coherent design capable of sustaining democratic oversight at machine speed.

C. The Protocol of Federation — Interoperability Without Harmonisation

The Third Architecture does not require cultural convergence. It requires protocol-level compliance. The Protocol of Federation enables interaction between incompatible normative
systems without demanding alignment of their internal values.

Its foundational principle is straightforward: national rules remain authoritative within their own borders, while cross-system interoperability depends on adherence to shared protocol specifications. When a country departs from those specifications, it encounters network-level consequences rather than diplomatic sanction. Isolation emerges as the structural outcome of non-compliance, not as a political punishment.

Isolation does not prevent the existence of rogue or state-sponsored shadow networks; it prevents their integration. The Protocol of Federation is not a universal security perimeter
but a constitutional boundary: it ensures that systems operating outside the compliance regime cannot draw on, contaminate, or impersonate actors within it. By tying access to shared identity services, auditability, and cross-operator interoperability to compliance, the protocol makes the benefits of federation unavailable to non-aligned networks.

Rogue ecosystems may persist, but they remain informationally and operationally segregated, unable to exploit the trust, identity, or compute pathways that define the primary network. In this sense, isolation is not a tool for suppressing shadow networks; it is a mechanism for preventing their expansion into the institutional core.

Hong Kong offers a transitional illustration. Its dual-system configuration — a common-law framework operating within Chinese national authority — demonstrates how divergent legal
regimes can coexist within a shared economic environment. The arrangement also illustrates a structural truth about federated governance: it is contingent on the stability of the underlying political compact, not a permanent constitutional settlement.

The Protocol of Federation marks a shift in regulatory logic — from harmonisation to compatibility. Nations need not agree on the meaning of liberty, privacy, order, or rights;
they need only maintain protocol-level interoperability. The framework achieves coexistence not by resolving value disagreements but by rendering them operationally manageable at the
machine interface.

The Protocol federates constraint schemas, compliance parameters, audit outputs, and risk classifications across jurisdictions, enabling systems to operate across borders without collapsing into the lowest common denominator. Its function is not to erase legal differences but to structure them — to allow divergent regimes to express their authority without forcing
technical systems into mutually exclusive configurations.

Where jurisdictional constraints conflict, the Protocol resolves them through three mechanisms. Precedence rules determine which country's requirements govern in cases of overlap, subject to the applicable legal test. Arbitration mechanisms provide defined channels for negotiated reconciliation where precedence alone cannot resolve the tension. Fallback defaults ensure that systems remain operable even when substantive disagreements persist, reducing the risk of system-level failure. The output is a set of interoperable rule bundles that allow AI systems to operate under multiple legal regimes simultaneously.

In this sense, the Protocol of Federation is neither a treaty nor a standard. It is an institutional mechanism for managing legal pluralism at the machine interface — a way for nations to remain sovereign while participating in a shared computational environment.

D. The Four-Tier Constitutional Design and the Role of Licensed Operators

The Third Architecture operates through a separation of powers not between branches but between tiers. Each performs a distinct constitutional function, and none can substitute for
the others.

The Value Tier is where societies form preferences, express norms, and generate legitimacy.
The Protocol Tier translates those preferences into machine-readable constraints. The
Operations Tier executes those constraints within the limits of physics — compute, energy, bandwidth, and latency. The Audit Tier measures and verifies AI behaviour, ensuring that
systems remain observable and contestable over time. Together, these four tiers constitute the minimum constitutional structure capable of stabilising governance in the AI era.^[12]

The Operations Tier is constrained not only by physics but by limits on authority. AI systems at frontier scale — with their tight interconnections, cross-domain effects, and autonomous
capabilities — require operation by Licensed Operators: entities authorised to run advanced
AI under domestic rules and within protocol-compatible licensing regimes. Licensing is not centralisation; it is functional delegation.

A Licensed Operator functions as a legally constrained environment: an entity required to comply with the Sovereign API, maintain real-time and retrospective audit and interruption
mechanisms, and operate under transparent public oversight consistent with alignment and safety obligations. Licensing thresholds are tied to capability levels rather than fixed technical specifications, allowing the framework to evolve as systems advance.

This structure enables frontier-scale AI to be deployed without compromising the integrity of the domestic constitutional order. It permits multiple entities to operate advanced systems within the same country while maintaining uniform conditions for accountability, contestability, and national control.

A further consideration is economic feasibility. If interface compliance becomes the core of authority, then the cost of maintaining compliant interfaces must not become a de facto barrier to entry. Frontier-scale systems require secure logging, identity verification, auditability, and interruption mechanisms, all of which impose fixed costs that can be prohibitive for smaller firms.

Ensuring a contestable ecosystem therefore requires complementary economic supports:
open-compute subsidies that lower the baseline cost of compliant operation, shared public infrastructure for identity and audit services, and standardised compliance modules that
reduce duplication across operators. These measures do not prescribe a particular market structure; they ensure that the operational overhead of compliance does not, by itself, determine who is allowed to participate.

Preventing capture requires more than technical neutrality; it requires ensuring that no single actor can monopolise the compliance stack. Antitrust scrutiny is necessary to prevent
dominant firms from controlling the interfaces through which all operators must pass, and to ensure that compliance tooling does not become vertically integrated into a single platform.

Open-compute subsidies and shared infrastructure serve a structural purpose: they prevent compliance costs from concentrating power in incumbents and preserve the possibility of
new entrants operating frontier-level systems under the same constitutional constraints. The
Licensed Operator model is not only a governance design but an economic design — one that distributes the cost of safety in a way that keeps the space contestable.

This does not eliminate the material advantages of scale — capital, compute, and expertise remain significant barriers — but it ensures that those advantages operate within a framework
of public constraint rather than private incumbency. The purpose of the licence is not to entrench hyperscalers but to condition continued participation on adherence to public rules
that are equally available to all. Firms that fail to meet these constraints lose access not through discretionary licensing, but through the predictable consequences of non-alignment within a federated network. The Licensed Operator model preserves contestability while anchoring frontier-scale AI within a constitutional order that remains legible, accountable, and open to new entrants.

The Audit Tier measures and verifies key aspects of AI system behaviour. It evaluates model outputs, data flows, control effectiveness, and compliance states, producing attestations,
exception reports, and behavioural traces. Unlike financial audit, which is periodic, document-based, and ledger-bound, this function is continuous, instrumentation-based, and system-level. It does not interpret rules; it generates an evidentiary record against which compliance and non-compliance can be determined.

Measurement warrants particular emphasis. As the manuscript has argued throughout, it is not auxiliary — it is constitutive. Without persistent measurement, continuous adaptation produces opacity rather than oversight. In the terms of the Incorporation Heuristic, the Audit Tier sustains the visibility variable — and without sustained visibility, no rule incorporates regardless of its workability or necessity. Accounting disciplines, built to track change over time, supply continuity that legal doctrine alone cannot maintain at machine speed.

This tier exists because the Measurement Gap identified in Part II cannot be resolved within existing accounting frameworks. Level 3 fair value — designed as a last-resort valuation
method when market evidence is weak — is structurally mismatched to frontier AI systems whose behaviour is probabilistic, path-dependent, and continuously updated. Instead of
estimating price, the Audit Tier generates continuous visibility into system operations, provenance, and effects through secure logging, identity verification, traceability, and interruption mechanisms. It tracks behaviour rather than financial position — a functional alternative to Level 3 valuation when valuation ceases to be informative.

Because existing accounting standards remain anchored to fair-value frameworks that cannot express model provenance, drift, or operational behaviour, the Audit Tier operates as a parallel evidentiary system. Its inputs are evaluation traces — safety-case documentation, training-data attestations, capability disclosures — rather than valuation estimates. The Audit Tier does not repair the Measurement Gap; it bypasses it, creating a verifiable visibility regime for systems whose economic and constitutional relevance cannot be rendered through conventional accounting categories.

E. Calibration Events — Democratic Revision at Machine Speed

Continuous systems require periodic legitimacy inputs. Calibration Events are proposed as the adaptive mechanisms through which citizens adjust the parameters that guide AI behaviour without interrupting ongoing operation. They provide structured, recurring signals of democratic judgement to systems that learn and adapt at machine speed.

Calibration Events occur at predictable but adjustable intervals, creating a stable rhythm for citizen input. They rely on deliberative platforms that support informed public reasoning rather than reactive preference expression. Their outputs take the form of parameter updates rather than statutes. Collective judgement is translated into machine-readable adjustments that enter the Sovereign API as regular revisions. Revised preferences then flow through the Protocol Tier and refresh the normative inputs that guide the system — without disrupting continuity in the Operations Tier.

Democratic governance adapted to machine-speed systems requires this regular update. Calibration Events allow societies to revise goals, adjust risk tolerances, and update normative baselines while keeping AI systems running — turning legitimacy from a onetime mandate into a recurring control signal. They complete the constitutional cycle: from Value to Protocol to Operations to Audit — and back to Value.

F. The Hard Problems — Inherent Tensions That Remain

Four underlying tensions define the research frontier of the Third Architecture.

The first is enforcement: how the Sovereign API can bind actors whose computational capacity, operational reach, or technical autonomy may exceed that of any single state.

The second is coordination: how multiple Sovereign APIs interoperate across national boundaries without drifting into incompatible standards or collapsing toward the lowest common denominator. This tension is especially acute for non-G3 countries, which must navigate competing AI ecosystems while preserving their own regulatory voice.

The third is transition: how the world moves from today's Invisible Constitution to the Third
Architecture without destabilising the digital and economic foundations on which societies depend. For countries reliant on foreign cloud providers or imported AI capabilities, the transition path is as consequential as the destination.

The fourth is scale: elements of parametric democracy have been demonstrated in small, digitally cohesive societies such as Estonia and Taiwan. It remains an open question whether
these approaches can operate in much larger countries where trust, identity, and digital infrastructure are unevenly distributed.^[13]

What works for a population of five or twenty-three million — high-bandwidth participation, stable digital identity, and broadly shared civic baselines — may not translate to India, Nigeria, Indonesia, or the U.S., where the parameters themselves are contested and where administrative capacity varies by orders of magnitude.

Scaling a deliberative system requires not only more participants but shared agreement on the common civic baselines that make deliberation possible: identity, authenticity, eligibility,
and the legitimacy of the parameters being tuned. In nations where these baselines are fractured — because identity systems are contested, digital infrastructure is uneven, or trust
is thin — parametric democracy may face structural limits that are not visible in smaller, more cohesive societies.

These tensions are not deficiencies. They define the frontier of constitutional engineering — the unresolved problems that will shape the research agenda and institutional debates of the AI era. They mark where further work is needed, and where the next generation of scholars, engineers, and policymakers must focus to ensure that the proposed design serves all societies, not only those with the most resources.

G. The Interdisciplinary Necessity

No single discipline contains the concepts or tools required to design constitutional frameworks for AI systems and the ecosystems of autonomous agents they are producing.
Each tier of the Third Architecture draws on a different intellectual tradition because each addresses a distinct category of problem.

The Value Tier, expressed through the Sovereign API, depends on legal definition and interpretation — the ability to express obligations, permissions, and constraints in forms that
institutions and machines can both process. The Protocol Tier, realised through the Protocol of Federation, draws on geopolitical reasoning — an understanding of how states, blocs, and cultural systems negotiate standards, assert authority, and manage interdependence. The
Operations Tier rests on computational engineering — the knowledge required to understand what frontier models and agentic systems can and cannot do within physical limits. The Audit Tier relies on accounting and measurement — disciplines built to verify claims, detect inconsistencies, and create records that can be trusted and contested.

This interdisciplinary foundation — law as form, geopolitics as coordination, engineering as constraint, accounting as evidence — reflects the structural realities of the AI era. Legal rules
become machine-readable. Geopolitical dynamics shape technical standards. System architecture determines which forms of governance are possible. Economic measurement
becomes continuous. No single field addresses these interactions alone. The Third Architecture requires their integration because the systems it governs operate across all four domains simultaneously.

H. Closing the Loop — The Visibility Gap Revisited

This work began with two linked impulses: a parent's question about the professional world two daughters — one studying law, one studying accounting — would enter, and the recognition that no single State can fully see the AI universe it now governs. The first impulse led to the second. The second became this manuscript.

The Visibility Gap, the Rule-Execution Gap, the Categorical Gap, and the Measurement Gap are not failures of political will. They are structural features of an AI-driven universe that has outpaced the observational and operational capacity of the institutions built to govern it. Closing these gaps requires not better regulations but better design — institutions capable of translation, mechanisms matched to the speed of computation, and democratic processes built
for complexity.

The Translation Layer already exists. The Incorporation Heuristic already compiles. The Material Constitution already operates. The physics will advance regardless.

What remains absent is the democratic framework that legitimises these mechanisms and connects them to the consent of the governed. That framework is not inevitable. It must be
designed, built, and defended — by the institutions that claim authority and the societies that grant it.

The architecture remains a choice. The choice remains ours.

1. "Capacity primacy" is an original analytic term introduced in this manuscript to describe the structural condition in which control of computational resources conditions the effectiveness of legal and regulatory authority.↩︎
2. Frontier-model energy and cost estimates evolve rapidly. Figures cited here reflect conditions as of early 2026. See IEA, Electricity 2026: Analysis and Forecast to 2030 (February 2026); Epoch AI, Data on Machine↩︎
3. Learning Hardware (March 2026); Epoch AI, Trends in Artificial Intelligence (accessed April 2026). Leading-edge semiconductor fabrication remains concentrated in Taiwan, particularly at the most advanced process nodes. SIA, State of the U.S. Semiconductor Industry (most recent edition).↩︎
4. Epoch AI, Data on Machine Learning Hardware (March 2026); Epoch AI, Trends in Artificial Intelligence(accessed April 2026); CSET, Global AI Talent Tracker (latest available edition).↩︎
5. J. Rawls, A Theory of Justice (Harvard University Press, 1971); J. Habermas, Between Facts and Norms (MIT Press, 1996); L. Fuller, The Morality of Law (Yale University Press, 1964).↩︎
6. R. Gorwa, R. Binns & C. Katzenbach, 'Algorithmic Governance: A Modes of Governance Approach' (2019) Regulation & Governance.↩︎
7. 'Legitimacy Debt' is an original analytic term introduced in this manuscript, modelled by analogy to technical debt. W. Cunningham, 'The WyCash Portfolio Management System' (1992), OOPSLA Experience Report.↩︎
8. The Third Architecture is an original institutional design proposed in this manuscript. "Third Architecture" distinguishes the proposed framework from the Westphalian territorial model (first) and governance through private intermediaries (second, the Invisible Constitution).↩︎
9. E.A. Posner and E.G. Weyl, Radical Markets (Princeton University Press, 2018), ch. 2.↩︎
10. P.F. Christiano et al., 'Deep Reinforcement Learning from Human Preferences' (2017) NeurIPS 30; Y. Bai et al., 'Constitutional AI' (2022), arXiv:2212.08073.↩︎
11. C. Hsiao et al., 'vTaiwan: An Empirical Study of Open Consultation Process in Taiwan' (2018); vtaiwan.tw.↩︎
12. The four-tier constitutional design — Value, Protocol, Operations, Audit — is an original framework introduced in this manuscript, drawing on traditions in constitutional separation of powers, systems architecture, and audit theory.↩︎
13. R. Kattel and I. Mergel, 'Estonia's Digital Transformation: Mission Mystique and the Hiding Hand', in Great Policy Successes (Oxford University Press, 2019).↩︎