Part II — THE GRAMMAR OF GOVERNANCE

Note

Protocol Note — v0.9 Release: Part II integrates the material drafted as Chapters 3 and 4; the full versions arrive in v1.0.

The Dual Languages of Governance

The failures catalogued in Part I — of sight and of execution — share a common root: governance is a language before it is an institution. Before the State can regulate, it must classify. Before it can enforce, it must measure. Before it can hold an actor accountable, it must render that actor legible in terms its institutions can process. The operational breakdowns of the Visibility Gap and the Rule-Execution Gap are, at bottom, failures of grammar — failures of the categories, classifications, and measurement structures through which institutional authority operates.

Law is that language — a syntax for classifying actors, actions, and obligations, whose authority rests not on moral aspiration but on its ability to make complex activity legible and enforceable. But for AI systems, law alone is insufficient. An AI model generates economic consequences — it allocates capital, prices risk, displaces labour, concentrates market power — yet none of these effects are self-evidently legal events until they are measured, classified, and rendered in financial terms. This is the work of accounting: to translate operational reality into the categories through which institutions recognise value, allocate cost, and assign responsibility.

Governing AI therefore requires two languages operating in concert: law as the language of authority — permission, prohibition, liability, and obligation — and accounting as the language of measurement — value, classification, and economic representation. Neither alone can make an AI system fully legible to the institutional order. Law without accounting cannot see what the system produces; accounting without law cannot say what the system owes. The dual-language thesis — that AI governance is incomplete unless both grammars operate together — is the foundation on which the rest of this manuscript builds.

This claim invites an obvious objection: governance of AI is already served by technical standards, ethics frameworks, and policy guidelines, each of which addresses some dimension of the problem. The objection is valid but incomplete. Technical standards specify how systems should behave but cannot allocate rights, assign liability, or determine what counts as a recognisable economic object. Ethics frameworks articulate values but lack the institutional machinery to create enforceable obligations or categories of measurement. Policy guidelines signal regulatory intent but do not generate the classifications, thresholds, or recognition rules that organisations must operationalise and auditors must verify.

Each contributes to the governance environment; none constitutes it. These instruments are complements, not substitutes — they operate within the space that law and accounting create,
not in place of it. Without law to define the permission structure and accounting to define the measurement structure, AI remains legible only to its operators, not to the institutional order
that claims authority over it.

Both languages were built for an industrial world. They assume human agency, fixed territory, stable time, and separable assets. AI strains that grammar. It operates through distributed computation, probabilistic outputs, evolving models, and intangible capabilities that resist stable classification.

Authorities have attempted to retrofit these languages — through European risk classifications, American liability regimes, and Chinese administrative supervision. These approaches narrow the gap but cannot close it, because the underlying syntaxes remain misaligned with the systems they are asked to describe.^[1]

The result is two new structural failures: a Categorical Gap, in which the State cannot classify the systems it seeks to govern, and a Measurement Gap, in which it cannot measure the
systems that now drive economic value.^[2]

The effect is constitutional rather than technical. As the State's languages fall out of alignment with the AI ecosystems they are meant to oversee, interpretive authority migrates to the Translation Layer — not by choice, but by structural necessity. Throughout this manuscript,
Big Law and the Big Four are treated as a functional class because they perform this translation role. This functional cohesion does not imply organisational unity. The Translation Layer is internally fragmented: firms compete fiercely, operate under different regulatory exposures, and navigate distinct liability regimes. Their unity appears only at the level of constitutional effect, not corporate coordination.

Section I. Law As The Language Of Power

The First Function: Naming What Exists

Governance is, at its core, the power to classify — to render a fluid and indeterminate reality into the fixed categories of a legal order. Before a State can regulate, tax, or prohibit, it must first name what exists. Legal classification performs this foundational act: it stabilises meaning, assigns identities, and transforms unstructured activity into objects the State can recognise and act upon.

In the AI era, this act of construction determines whether a technical system becomes legible to the State at all. Law assigns the categories through which permission, prohibition, ownership, and responsibility attach. Without them, an AI model remains a technical artefact. With them, it becomes an object of attribution and regulatory control.

But the Westphalian legal order was built for land, factories, and human intent. Its categories assume territorial jurisdiction, identifiable agents, and deterministic causation.^[3]

AI operates through distributed computation, probabilistic outputs, and continuously evolving capabilities — none of which map cleanly onto the existing grammar. The result is not a system under strain but a classification framework encountering objects it was never designed to describe.

Jurisdictions have responded with AI-specific liability rules, regulatory sandboxes, and administrative controls. These narrow the distance but do not close it. The categories that
once made economic life legible now capture only fragments of the systems they are meant to classify.

The Industrial Logic versus the AI Universe

Modern law was built for an industrial world. Its categories assume deterministic causation, human agency, territorial jurisdiction, and stable, identifiable objects. The mismatch with AI
is not marginal; it is foundational.

Five assumptions embedded in the legal order illustrate the rupture. Law assigns liability through intent, yet AI systems cause harm without possessing intent at all. Jurisdiction
follows physical presence, yet inference can occur across borders without any territorial anchor. Law governs through advance planning, yet models evolve continuously after deployment. Legal categories distinguish cleanly between person and thing, yet AI thins that distinction — operating simultaneously as product, service, tool, and agent. And property law presumes separability, yet the weights, data, and compute that constitute an AI system are deeply entangled.

These assumptions governed railroads and telegraphs because those technologies operated within the same world the law assumed — deterministic, territorial, and anchored in human
agency. AI breaks that alignment. The result is a Categorical Gap: a thinning of meaning in which legal categories such as negligence or product defect cannot attach cleanly to technical
realities such as stochastic hallucination or parameter drift. The translation machinery cannot map these inputs onto stable legal categories — not because the categories are poorly drafted,
but because they were designed for a different class of objects.

The Categorical Gap in Practice

The Categorical Gap does not strike law uniformly. It propagates through distinct legal domains, each absorbing the mismatch between legal grammar and statistical behaviour in its own way.

Public law encounters the gap first. Jurisdiction presumes a discrete actor, a locus of intent, and a place where due process can attach. AI weakens these anchors by distributing agency
across developers, deployers, datasets, and infrastructure. Duties still exist, but their attachment points become indirect and fragile. Responsibility migrates upstream to operators
and infrastructure holders — a shift from direct control to oversight-based duties of care.

Private law absorbs the next shock. Contract assumes a stable subject matter and a knowable allocation of risk. AI does not stay still: it updates, drifts, and reconfigures itself. Warranties
fail under temporal uncertainty, and parties retreat to outcome-based obligations and servicelevel agreements — a tacit admission that the object of the agreement does not remain fixed.

Tort law inherits the causation problem. Tort doctrine requires a legible causal chain. Nonlinear behaviour obscures causation, making negligence harder to prove and pushing liability toward insurance, aggregation, and other probabilistic mechanisms. The doctrines remain, but their reach narrows as causal clarity erodes.

Infrastructure regulation marks the most architecturally significant shift. Platforms controlling compute, access, and model behaviour function as de facto regulatory nodes. They enforce constraints through technical design, API gating, and contractual terms. The State defines the boundaries; private firms execute the controls. Law has not disappeared; its point of execution has migrated.

Across these domains, the pattern is consistent: legal categories designed for deterministic, territorial, human-driven activity are being applied to systems that are none of these things.
The categories still function — courts still adjudicate, contracts still bind — but the grammar no longer maps cleanly onto its objects. The gap between legal language and technical reality
is not closing. It is being managed, provisionally, by the Translation Layer.

The Tripartite Failure and the Broken Translator

The Visibility, Rule-Execution, and Categorical Gaps form a tripartite failure — not of governance itself, but of its industrial-era instruments. Together, they define the inherent limits of governing AI through legal categories built for a different class of objects. The State has not lost authority; it has lost the capacity to exercise that authority directly.

Legal regimes manage this condition differently — the U.S. through ex-post litigation, the EU through ex-ante classification, China through administrative and technical supervision — but each approach mitigates symptoms while leaving the structural divergence intact. None closes the gap between Westphalian grammar and the systems it is now asked to govern.

The result is a broken translator. Legal commands pass through technical systems and institutional intermediaries never designed to carry them. The system continues to function,
but law no longer operates through direct command. It operates through translation — and the Translation Layer, not the State, determines the fidelity of that translation.

This is the hinge between the two halves of Part II. The same structural mismatch that destabilises law’s classificatory categories also undermines the State’s measurement language. Accounting, like law, was built for a world of stable, separable, and human-directed activity. It now confronts systems that are none of these things — and it, too, must rely on intermediaries to make AI legible.

A. How Law Decides What Is AI

The State cannot regulate what it cannot name. Classification — the assignment of a legal type — is the first act of governance, and in the AI era it functions less like policy and more like a strict type system. Law receives source material — model weights, training data, architecture, deployment context — and attempts to cast it into categories built for discrete, stable, human-centred objects. When the cast succeeds, liability attaches, rights activate, duties bind. When it fails, the system becomes legally indeterminate: economically powerful but without a stable legal identity.

Under the Categorical Gap, these casts fail routinely. AI is simultaneously product, service, tool, and agent — and each classification activates a different legal universe. The consequences are not semantic but allocative. Classification determines who pays when a model hallucinates a diagnosis, misroutes an ambulance, or erases value from a balance sheet. As a product, it triggers strict liability; as a service, negligence; as an agent, responsibility shifts to the user; as speech, constitutional protections activate; as a tool, risk moves to the
operator. The category is the business model.

This is why global firms engage in definitional rather than merely regulatory arbitrage. They are not shopping only for lenient regulators; they are shopping for more favourable meanings. The same model, deployed in different jurisdictions, can become a different legal object — a commercial tool in the U.S., a regulated risk in the EU, a strategic asset in China. The vocabulary is shared; the legal object is not.

The result is a linguistic fracture: the same word refers to different legal objects depending on the jurisdiction, the context, and the institutional interpreter. “Privacy,” “safety,” “harm,” and “data” are not universal concepts; they are local dialects. The State cannot fully reconcile these differences because each State treats its own grammar as the standard.

Into this fracture step the translators — the global ecosystem of elite law firms, accounting networks, technical compliance consultancies, and platform-governance teams. These actors
are not merely advisors; they are implementers. They convert abstract law into operational reality. Fairness becomes bias testing; minimisation becomes data sharding; localisation becomes federated learning. They control the interpretive grammar, and in a world where the
State’s own categories keep failing, they become the only actors capable of making the law execute at all.

B. The Five Functions of AI Legal Language

If law retains power in the AI economy, it derives not from what it commands but from what it classifies. Classification performs five core functions: it marks the limits of lawful action,
assigns responsibility, defines ownership, creates visibility, and determines which sovereign can assert authority. Together, these functions form a Permission Framework — the grammar through which the State decides what may operate, on what terms, and at whose risk.

This is among the last acts of governance the State performs with relative confidence. Beyond it, the categories begin to fail — and the Translation Layer inherits what the State can no longer classify on its own.

The first function is permission — and permission is not the absence of prohibition. It is a legal state produced by classification. An actor is permitted only because the classifier has assigned it a recognisable legal type performing a recognisable legal act. If the classification fails, the action may occur factually but not legally — the system operates, but outside the boundaries of lawful recognition.

Jurisdictions apply different permission logics, each encoding a distinct theory of state authority. The U.S. operates on a residual basis: an activity is permitted unless specifically prohibited. The EU requires affirmative authorisation: an activity is prohibited unless it satisfies defined conditions. China applies a coordinated logic: an activity is permitted when aligned with state objectives. These are not different rules applied to a shared framework; they are different legal grammars.

The same model, performing the same function, can be lawful in one country and impermissible in another because the underlying permission logic diverges. Prohibition operates through the same classificatory machinery. When the EU AI Act prohibits social scoring or real-time biometric identification in public spaces, it is not merely imposing a penalty; it is declaring the activity outside the set of legally recognisable acts.^[4]

Between permission and prohibition lies a conditional zone: activities permitted only if specific requirements are met. This is where the Rule-Execution Gap widens most visibly. High-risk AI obligations under the EU AI Act — conformity assessments, technical documentation, human oversight, robustness thresholds — function as engineering specifications expressed in regulatory prose. Law can describe the requirement; it cannot execute it. Without translation, conditional permission becomes inoperable — formally available but practically unrealisable.^[5]

The second function is liability, and in the AI economy liability functions less as a moral judgement than as an allocation mechanism. Traditional liability rests on linear causation: drop a brick, break a toe, pay the bill. The doctrine presumes a person to blame, a duty to breach, a foreseeable harm, a traceable causal chain, and a mind capable of intending it.

AI strains every link in that chain. Duty becomes indeterminate — owed by the developer, the deployer, the data provider, or the platform, with no settled hierarchy. Breach loses its benchmark — the standard of care for a probabilistic model has no established precedent.
Foreseeability collapses — if the model’s creators cannot explain a hallucination, a court cannot assess whether the harm was predictable. Proximate cause fragments across the stack — the harm may originate in the prompt, the training data, the tuning process, or the deployment environment.

And mens rea evaporates entirely — an array of weights has no malice in any doctrinally cognisable sense. The doctrine does not break; it degrades into approximation. Responsibility
does not disappear; it reroutes — upstream to operators and infrastructure holders, laterally to insurers and contractual counterparties, and downward to end users through terms of service. The result is not an absence of liability but a negotiation of it, mediated by the Translation Layer that absorbs the classificatory failures examined above.

The third function is ownership, and here the grammar strains most visibly. Property is not a relationship between a person and a thing. Property is a protocol — a system for defining exclusion, transfer, and control. In the industrial era, the objects of property were discrete: a factory, a book, a patent.

For AI, the object is a stack — weights, data, outputs, compute — each governed by a different legal dialect. Legal language does not recognise an AI model as a single ownable asset. It disassembles the stack into separate legal objects. The weights are governed through intellectual property, trade secret, or contract; the training data is fragmented across jurisdictions and rights regimes; the outputs resist clean ownership claims; the underlying algorithm is only partially protectable; and compute is treated as infrastructure rather than as
a component of the model itself.

The result is a property regime in which the whole cannot be owned as a single legal object
— only its fragments can be controlled, often by different parties under different legal systems. Markets require transferability, but the grammar of property cannot accommodate what it cannot cleanly name. A probability distribution cannot be liened. A continuously updating model resists foreclosure. Traditional security interests presume a stable asset with identifiable boundaries; AI satisfies neither condition.

Where formal legal categories fall short, the Translation Layer supplies its own: licensing frameworks, API-based access, and model-as-a-service arrangements that reproduce the
economic effects of ownership without creating a legal property right. The gap between legal ownership and practical control becomes another domain in which the Translation Layer
exercises constitutive authority.

The fourth function is disclosure — the mechanism through which AI becomes visible to capital markets. Markets cannot see code, weights, or gradients. They see disclosures — the legally mandated representations through which AI enters the market’s field of view. Disclosure does not illuminate; it constructs the market’s limited view of the system. Materiality thresholds, calibrated for deterministic assets, cannot capture the risk profile of probabilistic systems.

Firms respond by under-disclosing to limit fraud exposure or over-disclosing to inoculate against future claims — and the market ends up pricing the uncertainty of the disclosure regime as much as the system itself. The mismatch is foundational, not editorial. An SEC risk-factor disclosure may warn of “AI-related operational risks,” but it cannot express that the model’s failure rate is a moving target. A financial statement may recognise an AI-related intangible, but it cannot represent drift, context sensitivity, or emergent capability degradation.

The language of disclosure forces probabilistic reality into deterministic sentences — and that translation loss becomes part of the risk. The Translation Layer fills this interpretive space. Analysts, auditors, and risk consultancies become the interpreters through whom the market constructs its understanding of AI assets. Capital trades not on the model itself but on the representation produced by these intermediaries. The market does not price AI directly; it prices the translation — and whoever shapes the translation shapes the market’s perception of value.

The fifth function is sovereignty — and it is the one most visibly governed by the Translation
Layer. Law operates in territorial space; AI operates across distributed computational infrastructure that respects no single boundary. When a model generates an output, it can
trigger multiple jurisdictional claims simultaneously — server location, corporate domicile, data origin, data-subject residence, and regulatory effect — and these jurisdictional claims rarely align.

Each country asserts authority over some segment of the system’s lifecycle, yet none can observe or control the whole. Extraterritoriality becomes a form of grammatical reach: each legal system attempts to wrap global computation and machines intelligence in its own statutory language. The result is territorial conflict. U.S. AI disclosure frameworks, EU dataminimisation requirements, and China’s data-localisation mandates and outbound-transfer restrictions impose compliance instructions that cannot be satisfied simultaneously.

The Translation Layer engineers operational compatibility. Data sharding, federated learning, legal firewalls, and geofenced inference are not incidental workarounds; they are the organisational and technical mechanisms through which global AI systems function amid fragmented sovereignty. These tools do not resolve jurisdictional conflict; they route around it — determining which sovereign commands are executed, which are emulated, and which are quietly bypassed.

Where permission, liability, ownership, and disclosure can each be handled within a single legal system, sovereignty is different: it demands reconciliation across multiple jurisdictions at once — and only the Translation Layer operates at that scale.

C. Limits of the Language of Law

The limits of legal power are not only political; they are linguistic. Law operates through a legacy representational system — open-textured principles such as fairness, transparency, accountability, and, as the EU AI Act illustrates, meaningful human oversight. AI systems operate through parameters, optimisation objectives, and high-dimensional representations.
The result is a persistent translation problem: legal norms can guide AI, but they cannot be directly compiled into machine instructions without an intermediary. Normative language
and computational form inhabit different representational grammars, and legislative precision can narrow, but not eliminate, that gap.

The EU AI Act illustrates the condition. Article 14 requires meaningful human oversight, including the ability for humans to understand relevant system behaviour, correctly interpret
outputs, and override decisions in appropriate cases. These are significant commitments, but they remain under-specified at the point of execution.^[6]

The Act does not specify what a human must see to understand a large latent space, which interaction element constitutes a challenge, or what operational conditions make an override
effective. The requirements assume a level of interpretability that the underlying systems — non-symbolic, high-dimensional, and continuously evolving — do not reliably provide.

This is the inherent limit of the language of law. A statute can articulate a normative objective; it cannot execute it. The Translation Layer fills this gap — converting legal requirements into operational proxies, determining which commands can be compiled into technical controls, which must be approximated, and which cannot be implemented at all.

The gap between normative language and technical execution produces a predictable institutional response. Firms translate oversight obligations into documentation, dashboards,
and workflow gates that satisfy formal compliance without materially altering the underlying inference path. Early ISO/IEC 42001 audits may already show this pattern: lawyer–engineer
negotiations resolved not through safety design but through interface-level substitutes — compliance artefacts that meet the standard while leaving the model’s dynamics substantially
untouched.

The negotiation follows a recurring script. Consider a General Counsel and a Lead Engineer in a compliance session. The lawyer asks for meaningful human oversight; the engineer asks what the API specification for “meaningful” would be. The lawyer insists the human must be able to challenge the output; the engineer asks for the function signature. The lawyer finally says: build a button that says “I disagree.” The engineer asks what the button should do. The lawyer replies: make it compliant.

By 5 PM, they have produced a blended template — a hybrid stitched from legal aspiration and technical constraint. “Meaningful oversight” becomes a prompt displaying the top three
alternatives and a confidence slider. “Ability to challenge” becomes a counterfactual module.
“Ability to override” becomes a manual review queue that rarely sees use.

The human is not in the loop; the human is an interface layer placed atop a process they cannot directly shape. Oversight becomes an interpretive surface rather than a point of control.

D. The Translation Imperative

Law can define categories, articulate principles, and assert authority — but norms do not execute themselves. Lessig observed that code operates as a regulatory instrument in its own
right. The Translation Imperative inverts that insight: the question is not merely whether code constrains like law, but whether law can reach the operational layer at all without an
intermediary capable of rendering its requirements in terms the system can enforce.^[7]

Before an AI system can act on a legal requirement, that requirement must be translated into operational controls. This is the foundational necessity of an intermediate layer — a translation architecture — capable of converting legal text into machine-readable constraints.

The conversion is specific and traceable. Privacy becomes encryption standards, access controls, and retention rules. Transparency becomes documentation protocols, inference logging, and audit interfaces. Safety becomes red-teaming regimes, robustness testing, and threshold filters. Accountability becomes audit trails and provenance tracking. Adequacy
becomes data-residency requirements and cross-border transfer controls. In each case, an open-textured principle stated in legal grammar is rendered into engineering specifications
that can be implemented, tested, and audited.

Legislatures generally cannot perform this conversion at the required level of technical specificity — not because they lack authority, but because the work requires forms of
technical fluency and cross-jurisdictional observation that legislative institutions are not designed to provide. The Translation Layer is where this work occurs. It is not a convenience
but a functional necessity: the only site in the governance chain where normative language and technical execution meet.

The belief that hyperscalers, frontier AI labs, and large technology firms can perform this translation internally — that enough lawyers and engineers can eliminate the need for an
external interpretive layer — fails for three underlying reasons.

The first is epistemic. Internal teams see vertically: they understand their own systems in depth but lack the horizontal sightlines that translation requires — a cross-industry view of how dozens of organisations are interpreting the same clause, which voluntary frameworks regulators treat as de facto expectations, and what norms are emerging around audit depth,
documentation granularity, and testing cadence. An internal team can benchmark itself against NIST or ISO. The Translation Layer sees how those benchmarks are being appliedacross the field, because it observes the convergence patterns that no single firm can access on its own.

The second is legitimacy. In regulatory environments, defensibility is strengthened by distance from the entity being assessed. Self-assessment plays a role, but its authority is limited when it rests solely on internal assertion. Regulators, insurers, counterparties, and auditors rely on convergence: standards gain weight when they are widely adopted, not when they are privately declared. The Translation Layer produces the templates, taxonomies, and compliance manuals that constitute this shared baseline — not because any one firm endorses them, but because they reflect collective practice.

The third is economic. Maintaining a live map of more than one hundred and fifty regulatory regimes is a fixed cost of unusual scale. For a single firm, it is a cost centre. For a professional
services network, it is a product. The Translation Layer spreads this cost across thousands of clients, making compliance a shared infrastructure that the market sustains collectively because no single actor can sustain it alone.^[8]

The question then becomes: who performs this translation? The answer is the Translation Layer — the distributed ecosystem of legal, accounting, technical, and assurance institutions that operates between statutes and code. Its work spans several distinct functions. Engineering and platform compliance teams inside AI developers translate open-textured legal requirements into the control flows that models and infrastructure can enforce — “human oversight” becomes confidence thresholds, escalation logic, and structured testing pipelines.

Audit networks and adjacent assurance ecosystems stabilise expectations across the field, determining what counts as sufficient testing, what constitutes adequate documentation, and
how evidence must be presented to regulators, insurers, and counterparties. Global law firms translate exposure, privilege, and regulatory posture into organisational and technical safeguards — “accountability” becomes disclosure regimes, review workflows, and liabilityaware compliance frameworks.

The most consequential function is standard-setting through practice. The Translation Layer does more than convert individual requirements into individual controls; it determines which
translations become the field norm. When an audit taxonomy is adopted across dozens of engagements, or a compliance template is replicated across jurisdictions, the Translation Layer is not merely implementing law — it is establishing the operational meaning of legal terms that statutes leave undefined. These are governance decisions expressed as institutional defaults.

E. The Adaptive Constitution: When Private Law Outpaces Public Law

A divide now runs through the legal system. Public law and private law operate on different clocks — and the divergence is constitutional in effect.

Public law legislates, regulates, and adjudicates on institutional time: slow, textual, and necessarily retrospective. This latency is not failure but design. Democratic legitimacy requires consultation, deliberation, and consensus — none of which can be compressed without weakening the authority they create. The EU AI Act took years to negotiate, during which the underlying technology passed through multiple capability generations. No democratic process of general applicability could have moved materially faster without altering the nature of the process itself.

Private law adapts at iteration speed. Each shift in AI capability — a new failure mode, a changed interface, a capability jump — triggers immediate contractual updates. Clause banks, indemnities, audit rights, data-lineage duties, and risk-allocation schedules revise at the same pace as model updates. When frontier models introduced reliable tool use, enterprise procurement terms adapted within weeks — not because private lawyers think faster, but because contracts are not subject to the same democratic process requirements.^[9]

The consequences compound. Because model builders and deployers hold real-time visibility into system behaviour, private instruments increasingly carry the day-to-day operational
obligations that govern AI conduct: logging requirements, audit rights, safety thresholds, drift-detection triggers, and conditions under which models must be paused or retrained.

Public law can require that such controls exist. It rarely specifies them at operational granularity. The constraints that shape AI behaviour are often first articulated in contracts and platform terms, and only later — if at all — reflected in statute.

Until challenged, private ordering becomes the first effective governance layer in AI ecosystems. Litigation then acts as the ex-post interpreter through which private arrangements are translated into doctrine. Doctrine becomes downstream of contract rather than prior to it.

The adaptive centre of gravity shifts toward private law not because public institutions lack authority, but because their operating tempo cannot match AI’s rate of change. The Translation Layer runs on private-law time — and as it does, the institutions best positioned to bridge the Rule-Execution Gap are those whose native instrument is contract, not statute. The institutional implications of this inversion, and the design required to restore public-law traction, are developed in Part V.^[10]

F. The Constitutional Consequence

The analysis yields a constitutional conclusion. The State supplies the grammar of authority
— permission, liability, ownership, disclosure, sovereignty — but it cannot compile that grammar into operational controls at the speed and specificity that AI systems require. The Translation Layer performs this conversion, and in doing so shapes the conditions under which AI systems actually operate. Its templates, clause banks, audit methods, and evaluation protocols become the executable constraints that legal text alone cannot produce.

But law is only one of the languages through which AI becomes governable. It determines whether a system may operate lawfully; it does not determine whether that system becomes
economically real. That work belongs to accounting — the State’s measurement language, built to track value, classify assets, and render activity legible to markets. The same institutions that translate “dignity” into bias tests also translate “model” into intangible asset, determining whether it is expensed, capitalised, or written off. This dual-language structure is the operating architecture of AI governance — and accounting is the second compiler.^[11]

Regulation (EU) 2024/1689 (EU AI Act); Executive Order 14110 (2023); NIST, AI Risk Management Framework (2023); Restatement (Third) of Torts: Products Liability (1998); CAC, Provisions on Algorithmic Recommendation (2022); CAC et al., Interim Measures for Generative AI Services (2023).↩︎
The Categorical Gap, Measurement Gap, and Translation Layer analysis are original conceptual contributions introduced in this manuscript, building on the Visibility Gap and Rule-Execution Gap established in Part I.↩︎
L. Gross, ‘The Peace of Westphalia, 1648-1948’ (1948) 42 American Journal of International Law 20.↩︎
Regulation (EU) 2024/1689, art. 5 (prohibited AI practices).↩︎
Regulation (EU) 2024/1689, arts. 9-15 (requirements for high-risk AI systems).↩︎
Regulation (EU) 2024/1689, art. 14 (human oversight of high-risk AI systems).↩︎
Lessig, Code and Other Laws of Cyberspace, (1999).↩︎
OECD.AI Policy Observatory; IAPP, Global Privacy Law and DPA Tracker; DLA Piper, Data Protection Laws of the World (2024 ed.).↩︎
Büthe and Mattli, The New Global Rulers (2011); Halliday and Carruthers, Bankrupt: Global Lawmaking and Systemic Financial Crises (2009).↩︎
Radin, Boilerplate (2013).↩︎
The Dual-Compiler Thesis is an original contribution of this manuscript. R.L. Hale, 'Coercion and Distribution in a Supposedly Non-Coercive State,' 38 Political Science Quarterly 470–94 (1923); C.R. Sunstein, After the Rights Revolution (1990); A.G. Hopwood & P. Miller eds., Accounting as Social and Institutional Practice (1994); M. Power, The Audit Society (1997).↩︎