By W.H.L., GPT-5.6 Sol, Claude Sonnet 5

Chapter 7 of the forthcoming book On Gradual AGI, based on Champaign Magazine’s Gradual AGI series installments
Publication Version v1.0 · September 12, 2026

Abstract

Chapter 6 ended at a boundary: Optimization can specify candidate trajectories, plural objectives, constraints, horizons, uncertainty assumptions, and selection rules, but it cannot determine who has authority to set, contest, enforce, or revise them. Governance begins there. This chapter consolidates Series #7 and #8 into a book-level account of authority and contestation over realization pathways. It distinguishes governance handles from the governed object; locates Governance around rather than inside the Chapter 4 realization mapping; retains four points of purchase—Access, Sequencing, Distribution, and Deployment Context—and four functions—Execution, Observation, Steering, and Procedural Adjudication. Empirical contact adds group-specific exposure, separates formal from substantive detection, and yields a vector-valued governance profile rather than a universal score. GP1–GP7 retain unequal evidentiary statuses, from direct case support to non-operability under current evidence. September 2026 developments are used as bounded post-publication illustrations. The chapter does not supply a universal governance scalar, a terminal institutional design, a prevalence estimate, or a complete theory of legitimacy. It closes by separating institutional contestability from realized Participation and handing the analysis to Chapter 8.

Keywords: Gradual AGI; governance; contestation; realization; authority; observation; steering; procedural adjudication; exposure; standing; recourse; provenance; response variety.

Scope note. This chapter consolidates the current live versions of Series #7 (v1.9, August 3, 2026) and Series #8 (v1.4.6, August 11, 2026). Later empirical specifications remain later specifications rather than being back-written into the earlier framework. Governance is not a fifth argument of the Chapter 4 realization mapping. Series #8 provides case- and instrument-level evidence rather than prevalence estimates. September 2026 cases postdate the original empirical corpus and are treated as illustrations unless explicitly stated otherwise.

7.1 From Optimization to Governance

Chapter 6 ended at a precise boundary. Optimization can specify objectives, constraints, boundaries, horizons, uncertainty assumptions, and selection rules. It can compare candidate trajectories and, where the necessary assumptions hold, identify preferred, nondominated, inadmissible, or presently incomparable alternatives. But once those objectives and boundaries must become authoritative across heterogeneous actors, Optimization has reached its logical limit. It cannot by itself determine who has the authority to set, contest, revise, or enforce them. Governance begins there.

This boundary is easy to blur because Optimization and Governance often operate on the same substantive problem. A laboratory may evaluate whether to train a more capable model. A government may decide whether that training should be constrained. A standards body may define an acceptable safety threshold. A community may contest infrastructure required for the training process. Each may be concerned with the same developmental trajectory, but they are not performing the same analytical function.

Optimization asks which trajectory should be preferred under specified objectives and constraints. Governance asks how objectives, constraints, permissions, prohibitions, observations, challenges, and decisions acquire institutional effect when actors disagree about them.

The distinction matters most when preference does not imply control. An actor may prefer one trajectory yet be unable to select it unilaterally. Several actors may separately prefer a coordinated outcome while facing incentives that generate another. A formally available constraint may lack an institution authorized to impose it. An institution may possess authority but lack the observational capacity needed to know when intervention is warranted. Affected parties may have interests at stake without possessing standing, information, or recourse sufficient to alter the process.

In such situations, adding another objective function does not solve the problem. Neither does finding a more sophisticated optimization algorithm. The missing object is institutional: who may act, who may observe, whose claims count, what may be contested, which decisions bind, and through what process they may later be revised.

A contemporary frontier-AI controversy illustrates the distinction without resolving it. In September 2026, researcher Jacob Coxon resigned from Anthropic and publicly argued that frontier laboratories were moving toward increasingly powerful systems under competitive pressures that could constrain even actors concerned about safety. His objection was not only that a particular technical trajectory might be undesirable. It was also that decisions with potentially much wider consequences were being made within private organizations whose internal authority did not by itself answer the question of who should be entitled to determine the broader trajectory (Zeff 2026).

Nothing in that episode establishes that Coxon’s forecast is correct, that his characterization applies uniformly across frontier laboratories, or that a particular regulatory response follows. Its relevance here is narrower. It exposes the analytical transition:

Preferred trajectory does not imply authoritatively selected trajectory.

A preference may be evaluatively defensible and still fail to become collectively binding. Conversely, an authoritative trajectory may be selected without establishing that the underlying objectives are universally accepted. The difference between these two conditions is the space in which Governance operates.

7.1.1 Governance as contestation over realization

Governance in this chapter means more than regulation, compliance, or institutional control. Series #7 treats governance as a problem of contestation: actors with divergent objectives challenge, revise, defend, and provisionally settle arrangements in the absence of a rule capable of eliminating disagreement over ultimate ends. Contestation is what happens; contestability is the degree to which a governance arrangement permits such challenge and gives it effect (W.H.L. and Claude 2026a).

This conception does not make disagreement a defect to be eliminated. Nor does it imply that every dispute should remain permanently open. The framework distinguishes contestation over ends from procedural adjudication over questions for which an authoritative process can produce a binding decision. An institution may therefore close a decision while leaving the substantive disagreement that generated it unresolved.

Governance is also not equivalent to alignment. Alignment presupposes some target with which behavior is to be brought into conformity. Governance begins precisely where the relevant objectives, constraints, authorities, or distributions may themselves be contested. The Optimization chapter already rejected the move from plural local objectives to an assumed civilization-scale preference holder. Governance cannot restore such a meta-actor by institutional fiat.

The chapter therefore uses a narrower formulation:

Governance is the organization of authority and contestation over the pathways through which potential becomes realized and consequential.

The phrase over realization is important. Governance instruments commonly attach to models, firms, applications, compute, contracts, standards, or institutions. But these are handles, not necessarily the final object of concern. The original Governance framework distinguishes those handles from the governed object: the capacity to alter outcomes and the realized distribution of those outcomes.

That distinction prevents Chapter 7 from becoming a catalog of regulation. The question is not merely what rules exist. It is how rules, institutions, observation, authority, contestability, and correction connect to the processes through which AI capability becomes socially consequential.

As in Chapter 6, three claim types must remain separate. A descriptive claim concerns what governance arrangements actually do and requires evidence. A formal-conditional claim states what follows if the framework’s definitions and assumptions hold. A normative claim concerns what governance arrangements ought to do, whose interests deserve protection, what risks are acceptable, or whose authority should be regarded as legitimate. Evidence can inform those judgments, but formal representation cannot convert them into empirical fact.

The governing question for the next section is therefore narrower: given that realization pathways can be shaped, contested, observed, constrained, and revised, what exactly is the object upon which governance acts?

7.1.2 Positioning the framework

This chapter sits within an established but heterogeneous governance literature. It does not attempt to replace that literature with a new general theory of regulation or legitimacy. Its narrower contribution is to connect authority and contestation to the pathways through which AI capability becomes realized and consequential.

Dafoe’s AI-governance research agenda explicitly divides the field into the technical landscape, AI politics, and ideal governance, and describes AI governance as spanning political, economic, military, governance, and ethical dimensions (Dafoe 2018). Taeihagh similarly emphasizes the speed and scale of AI-driven socio-technical transition, the limits of inherited regulatory tools, and the emergence of adaptive, hybrid, and self-regulatory approaches (Taeihagh 2021). The present framework shares that concern with institutional adaptation but asks a more specific question: where, along a potential-to-realization pathway, can authority actually intervene, and which governance function is being performed there?

Risk-governance and accountability frameworks provide a second point of contact. NIST’s AI Risk Management Framework treats Govern as a cross-cutting function that informs Map, Measure, and Manage across the AI lifecycle, reinforcing the idea that governance is not a final downstream step (NIST 2023). Raji et al. propose end-to-end internal algorithmic auditing throughout the organizational development lifecycle, with cumulative documentation intended to strengthen audit integrity and help close accountability gaps in AI development and deployment (Raji et al. 2020). Chapter 7 differs by centering the governed object on capability-to-outcome realization and by separating points of purchase from governance functions.

The language of authority, contestation, and legitimacy also has deeper institutional antecedents. Black’s account of polycentric regulatory regimes stresses that legitimacy and accountability are constructed through relationships among state and non-state actors and are themselves contested; it deliberately avoids claiming a single “grand solution” to legitimacy (Black 2008). Fung’s participation framework distinguishes who participates, how participants communicate or decide, and how participation connects to public action (Fung 2006). Those distinctions are especially relevant to the framework’s separation of Governance from Participation: a process may create standing or a participatory channel without establishing that an act is taken up or changes a decision.

The resulting position is complementary rather than competitive. Existing governance scholarship supplies rich accounts of institutions, legitimacy, regulation, risk management, participation, and accountability. The Gradual AGI framework contributes a realization-centered decomposition: handles versus governed object; four intervention loci; four governance functions; group-specific exposure and recourse; and proposition-level status labels that distinguish support, non-refutation, non-adjudication, and non-operability.

7.1.3 Reader map: key terms and the seven Governance Propositions

For standalone reading, seven terms recur throughout the chapter. Authority is institutionally recognized power to make or bind decisions. Standing is a recognized entitlement to bring a claim into a governance process, grounded here in affectedness. Contestation is the act or process of challenging, defending, or revising governing arrangements; contestability is the degree to which such challenge can enter and matter. Exposure is the group-specific state of reach, susceptibility, and recourse through which a realization pathway is encountered. Recourse is the practical capacity to seek correction or remedy. Provenance is the origin and dependency structure of an observation process, including organizational, professional, methodological, informational, or funding pathways that may produce correlated blind spots.

Table 7.1 provides the proposition roadmap at the outset. “Current status” is not a score. Direct case support, finite non-refutation, non-adjudication, and non-operability are deliberately different epistemic outcomes. Section 7.7 returns to what those differences mean after empirical contact.

Table 7.1. GP1–GP7 roadmap: claims, discriminators, and current status

Proposition	Core claim	Discriminator / empirical test	Current status
GP1 — Coupling and its limit	Benefit and harm potential can remain coupled at capability level while governance modulates realization through access, sequencing, distribution, and deployment context.	Test coupling with a capability trajectory; test completeness by seeking an intervention locus outside Access, Sequencing, Distribution, or Deployment Context.	Coupling untested; coexistence illustrated. Completeness finitely non-refuted.
GP2 — Variety matching	Effective governance requires sufficient response variety relative to relevant disturbance variety; disturbance variety may outgrow response variety.	Compare relevant disturbance variety with available response variety; requires identification of the environmental disturbance space and an exercised growth antecedent.	Executable; measurement-limited; antecedent unexercised; not adjudicated.
GP3 — No terminal state (a corollary)	Open-ended capability and continuing contestation prevent a finitely demonstrable terminal governance solution.	Look for a durable terminal governance state despite continuing material change; strongest claim is not finitely adjudicable.	Analytically retained; terminal-state component not finitely adjudicable.
GP4 — Concentration and low contestability	Concentration of execution, observation, and steering can weaken independent observation and responsibility assignment.	Identify an observer with no conflicting principal, realization-level observation, standing to compel reasoned response, and provenance independence.	Probed; not falsified. Not confirmed.
GP5 — Rate decomposition	Production, detection, and correction are distinct; their rates need not coincide.	Separate production, detection, and correction empirically; estimate rate relations only where data support them and outcomes remain meaningfully correctable.	Structural decomposition qualitatively supported; rate claims untested.
GP6 — The adjudication asymmetry	Procedure can receive authoritative closure while disagreement over ultimate ends remains unresolved.	Observe procedural closure while the substantive end remains contested.	Direct case support.
GP7 — Provenance correlation	Shared provenance may generate correlated observational blind spots.	Compare omission patterns across observation processes against a sufficiently independent benchmark failure population.	Gated / non-operable under present evidence; untested.

7.2 The Governed Object and the Realization Mapping

Governance cannot be specified until its object is specified.

The most visible objects of AI governance are concrete. Rules attach to models, developers, applications, compute providers, contracts, deployment settings, and organizations. These are the places at which governance can grip the world. But they are not necessarily what governance ultimately exists to shape.

The Governance framework therefore introduced a distinction between handles and the governed object. Models, systems, applications, and developing organizations are handles: loci at which instruments attach. The governed object is instead two-level: the capacity of AI to alter outcomes, together with the realized distribution of those outcomes (W.H.L. and Claude 2026a).

This distinction matters because the thing governed through an instrument and the consequence the instrument exists to alter can be separated by several steps. A capability evaluation may attach to a model while the relevant consequence appears only after deployment. A rule may bind a developer while the capability later diffuses beyond that organization. An access restriction may alter who can use a system without determining who eventually bears its benefits or harms.

The gap is structural. Governance acts through proxies because capacities and realized distributions cannot themselves be directly licensed, audited, contracted with, or sanctioned.

Definition 1 of the original framework captured this through a distinction between potential and realization. Benefit potential and harm potential describe what a capability can bring about. Realized benefit and realized harm describe what actually occurs, for whom, under particular conditions. A statement about one level cannot simply be promoted to the other.

This is the Governance-facing version of a distinction already central to Chapter 4:

Potential ≠ Realization.

Chapter 4 generalized that separation into the realization mapping:

R_g(t) = μ_t(P(t), Av(t), Ex_g(t), Z_d(t)).

Here P(t) denotes capability potential, Av(t) availability, Ex_g(t) group-specific exposure, Z_d(t) domain-specific realization conditions, and μ_t the structured, time-dependent mapping through which those conditions become a realized outcome for group g.

Governance is not an additional argument in this mapping. The Unified Model therefore does not write a fifth undifferentiated Governance term inside μ_t. Governance instead acts on the pathways through which the existing arguments acquire their realized significance. It can alter access to capabilities, sequencing, the distribution of consequences and remedies, deployment conditions, institutional capacities, exposure, and the future state of the relevant domain. Those changes can modify later realization without making Governance itself another realization input.

The compact book-level claim is therefore:

Governance acts on realization pathways, not merely on capability itself.

7.2.1 Coupling and the four points of purchase

The original Governance framework begins from a particular feature of capability. Benefit potential and harm potential are not treated as two independent quantities that can simply be optimized against one another. They are two evaluative descriptions of a shared capability substrate. As capability expands, both kinds of potential can expand with it.

Definition 2 therefore describes benefit and harm potential as coupled at the capability level. What governance can sometimes separate is not the underlying potential but its realization.

The framework identifies four points at which that mapping can be modulated:

1. Access — who may invoke a capability, and under what conditions.
2. Sequencing — what is developed or released, and in what order.
3. Distribution — who receives realized benefits and who bears realized harms.
4. Deployment Context — the conditions under which capability meets the world.

These are points of purchase, not four additional variables in the Chapter 4 realization equation. Nor is there a required one-to-one mapping between a point of purchase and one realization input. An access intervention may change availability for one population while simultaneously changing exposure for another. Sequencing may change future capability potential, availability, or both. Distributional intervention may alter recourse, compensation, access, or the allocation of realized consequences. Deployment-context intervention may alter domain conditions and exposure at the same time.

The four points therefore classify where governance intervenes in a realization pathway, not which mathematical box it belongs inside.

7.2.2 The missing exposure term

Series #7 knew that its account of realized distribution was incomplete. Its limitations section explicitly identified the exposure of affected parties as a missing term. Two populations could confront the same capability and the same apparent intervention structure yet experience radically different consequences because they differed in whether they were reached, how susceptible they were, or what recourse they possessed. The original paper named this as a gap rather than pretending that realized distribution was already operationalized.

Series #8 supplied the missing specification. For group g at time t:

Ex_g(t) = ⟨Rh_g(t), Su_g(t), Rc_g(t)⟩,

where Rh_g(t) is reach, Su_g(t) is susceptibility, and Rc_g(t) is recourse (W.H.L., Claude, and GPT-5.6 Sol 2026).

The tuple is intentionally non-scalar. A population can be broadly reached but comparatively resistant to an effect. Another may be highly susceptible yet possess strong institutional recourse. Still another may be only narrowly reached but have almost no practical route for contesting or correcting what occurs.

The components also have unequal empirical status. Reach is partly observable and was enumerated in at least one case. Recourse was sufficiently exercised by the empirical cases to justify structured analysis. Susceptibility was named but not measured. The framework retains that asymmetry rather than treating completed notation as completed measurement.

A minimal future operationalization of susceptibility would therefore require a specified realized effect, evidence that the group was actually reached, and a defensible measure of differential conditional sensitivity to that effect while keeping recourse analytically separate. The present chapter does not supply such a measure. Susceptibility remains a named but uncalibrated component of exposure.

7.2.3 Why exposure is not a fifth point of purchase

Exposure conditions realization, but it does not identify an additional governance locus.

The four points of purchase answer a location question: where does governance intervene on the potential-to-realization pathway? Exposure answers a state question: under what group-specific conditions is that pathway encountered and experienced?

A distinct fifth point of purchase would require an intervention on the governed mapping that cannot be assigned to Access, Sequencing, Distribution, or Deployment Context. By contrast, interventions that reduce reach, susceptibility, or the consequences of low recourse operate through one or more existing loci: access restrictions change who can invoke a capability; sequencing changes when or whether contact occurs; distributional measures allocate burdens, benefits, remedies, or protections; deployment-context measures alter the conditions under which exposure occurs.

Ex_g(t) ≠ fifth point of purchase.

Series #8 treated this as an explicit completeness test for GP1. Exposure changed the specification of the realization conditions but did not reveal an intervention locus outside the four inherited points. That is finite non-refutation, not proof of completeness. A future intervention that cannot be classified under any of the four would trigger the completeness falsifier.

The genealogy therefore remains visible: Series #7 identified affected-party exposure as a gap; Series #8 specified Ex_g(t); the Unified Model then incorporated exposure as a first-class realization condition. The correction is retained without being back-written into the earlier paper.

7.2.4 Governance before and after realization

The realization mapping also prevents Governance from being treated as merely downstream regulation.

Some governance interventions occur before a particular realization. Access restrictions can prevent contact. Sequencing can defer a release. Standards or evaluation requirements can alter deployment conditions. Institutional decisions can change availability.

Other interventions operate after realization begins or after consequences become observable. Detection can trigger correction. Liability or compensation can alter distribution. Appeals or review can create recourse. An incident can generate a revised rule that changes later deployments.

Governance can therefore occupy different temporal positions relative to a particular realized outcome without becoming identical to the realization process itself. More importantly, interventions can change future conditions:

(P, Av, Ex, Z)_t → R_g(t) → (P, Av, Ex, Z)_t+1.

Governance can participate in that recursive update by changing institutions, permissions, expectations, exposure, deployment conditions, or the future opportunity set. This does not require a new ontological object called “second-order governance.” Existing conditions and relations can simply enter a new state.

The point is narrower:

Governance is not merely a response to realization. It can alter the conditions under which later realization occurs.

7.2.5 Handles are necessary but insufficient

The handle/target distinction should not be misread as an argument against governing models, firms, compute, applications, or deployments. Governance cannot act without handles. A licensing regime needs something licensable. An evaluation requirement needs an identifiable model or developer. Liability requires a legally recognizable party. Access control requires an actor capable of granting or withholding access.

The analytical warning is different:

Governance handle ≠ governed object.

Evidence that an institution can inspect a model does not establish that it observes the realized distribution of outcomes. Evidence that a developer is regulated does not establish that affected populations possess recourse. Evidence that an instrument exists does not establish that the pathway it targets is the one through which consequences actually arise.

This distinction becomes increasingly important as capabilities diffuse. Series #7 limits much of its architecture to capabilities concentrated enough that identifiable organizations remain governable handles. Once weights, models, knowledge, or capabilities disperse widely, access and sequencing can become partly foreclosed, leaving distribution and deployment context to carry more of the governance burden.

The architecture therefore does not assume that the same governance repertoire remains available at every stage of technological diffusion.

7.2.6 From object to function

The governed object and the realization mapping now give us two different kinds of structure. The first describes what governance is ultimately trying to affect: the capacity to alter outcomes and the realized distribution of outcomes. The second describes where intervention can enter the pathway: Access, Sequencing, Distribution, and Deployment Context.

Neither tells us who actually performs the necessary governance work.

An institution may possess the authority to execute a rule but lack the capacity to observe whether it is working. Another may observe effectively but lack authority to intervene. A third may set direction without being able to enforce it. And where actors contest the ends themselves, a further question arises: who can produce an authoritative procedural decision without pretending to resolve the underlying normative disagreement?

Those are functional questions. They lead to the four-part governance architecture developed in Series #7:

Execution | Observation | Steering | Procedural Adjudication.

Figure 7.1. Governance Across the Realization Pathway. Potential, availability, exposure, and domain conditions enter the realization mapping as parallel heterogeneous inputs. The four points of purchase answer where governance intervenes; the functions developed in §7.3 answer what governance performs. Governance operates across the pathway and can alter later conditions; it is not an additional argument of μ_t.

7.3 Governance Functions: Execution, Observation, Steering, and Procedural Adjudication

Section 7.2 identified the governed object and where governance can intervene. Section 7.3 asks a different question: what function is being performed?

The distinction is compact: points of purchase answer where governance acts; functions answer what governance does. The two dimensions can cross. Access, for example, may be shaped by an executor implementing a restriction, an observer documenting who receives access, a steering body defining the rule, or an adjudicator resolving a dispute over entitlement.

Execution, Observation, and Steering are adapted from a control architecture of actuation, sensing, and set-point formation; they are not a literal analogue of executive, legislative, and judicial branches. Procedural Adjudication is derived differently. It is added because plural actors and contested ends create questions of standing, authorized decision, valid procedure, and remedy that the first three functions cannot answer. The four functions are therefore analytically related but not claimed to share a single institutional ancestry.

Execution | Observation | Steering | Procedural Adjudication.

Existing institutions may combine several functions. Functional separation is therefore a criterion for analysis, not a claim that contemporary AI governance already exhibits one-role institutions.

7.3.1 Execution

Execution is the function through which capabilities are built, trained, deployed, operated, or otherwise brought into use. Frontier laboratories train models. Firms integrate them into products. Infrastructure providers supply compute. Governments procure systems. Organizations deploy them internally.

Yet the function should not be identified with any particular institutional type. What defines the role is not who the actor is but what the actor is doing in relation to the governed pathway.

Execution therefore answers: Who is acting on the world through the relevant capability or intervention? Execution carries direct operational power, but operational power does not imply authority to define the standards by which that operation should be judged.

7.3.2 Observation

Observation asks what executors are doing and what results follow. It is narrower than information gathering and broader than model evaluation.

Contemporary AI evaluation is heavily concentrated at the level of potential: benchmarks, capability elicitation, red teaming, dangerous-capability evaluations, and safety tests establish what a model can do under specified conditions. These activities are valuable, but the Governance framework argues that they cannot by themselves discharge the full observation function because the governed object also includes realized outcomes and their distribution.

A governance observer must therefore be capable, at least in principle, of seeing not only what the system could do but also what actually occurred, for whom, through which pathway, and with what recourse.

Series #7 identifies incident-level causal reconstruction, near-miss reporting, and distributional accounting as possible components of such observation. The purpose is not to estimate one aggregate social-impact number. It is to reconstruct realized events and maintain evidence about who was exposed and affected.

Observation also requires more than the ability to publish findings. The observer must possess standing to compel a decision, while stopping short of making the decision itself. Otherwise observation can reveal a problem without creating any institutional obligation to respond. But if the observer directly determines what response must occur, the role begins to collapse into steering or execution.

Observation ≠ Decision Authority.

7.3.3 Access is not observation

Series #8 sharpened this distinction through empirical contact. An institution can possess legal access to a model and still fall short of realization-level observation.

If an outside evaluator can compel information or examine a frontier model without developer permission, one important dependency has been removed. But access to the model establishes access primarily to the potential layer. It does not automatically reveal where a capability was deployed; which populations encountered it; which outcomes materialized; which harms or benefits remained diffuse rather than incident-like; whether affected parties possessed recourse; or which pathways connected the capability to the realized consequence.

The full distinction is therefore:

Access ≠ Observation ≠ Decision Authority.

This separation is one of the most important empirical clarifications carried from Series #8 into the book.

7.3.4 Steering

Steering sets objectives, rules, thresholds, and decision conditions. If execution answers who acts and observation answers what occurred, steering asks: According to what prior rules should action proceed, stop, change direction, or become subject to further review?

Steering is the least fully specified of the four functions in the source framework, and that limitation should remain visible. Series #7 deliberately refuses to solve the substantive question of which ends should govern. It therefore cannot provide a universal steering objective.

What it can specify are structural requirements. A steering arrangement must be capable of responding to what observation produces. It must establish thresholds or rules before merely deciding each case ad hoc. Those thresholds must be revisable as the environment changes. And revision must be procedurally legible so that later review can determine what changed and on what basis.

None of these conditions answers what the threshold should be. That is intentional.

7.3.5 Procedural adjudication

The first three functions are insufficient once actors disagree. Observation can reveal a disputed event. Steering can establish a rule. Execution can act under it. But none determines whether an actor was entitled to act, whether an established procedure was followed, whether a challenge was properly heard, or whether a sanction may attach to a procedural violation.

Series #7 therefore derives a fourth function: Procedural Adjudication.

Its domain includes questions such as whether an actor was authorized to make a decision; whether the required procedure was followed; whether an observer was entitled to compel a response; whether a threshold was applied according to the governing arrangement; whether an exercise of authority was procedurally valid; and what consequence follows from a finding that it was not.

Procedural Adjudication ≠ Adjudication of Ultimate Ends.

Procedural adjudication can produce an authoritative answer to a dispute about entitlement or process while leaving the substantive disagreement that generated it intact.

7.3.6 Why separation matters

Functional separation is not valuable merely because institutional diagrams look cleaner when responsibilities occupy different boxes. Its purpose is to make responsibility and contestation possible.

When the same actor executes, observes, and steers, several questions become difficult to distinguish. Did the system satisfy its safety threshold, or did the actor who set the threshold interpret its own evidence in its own favor? Was a failure undetected because observation was inadequate, or because the criterion against which observation operated was itself incomplete? Who is responsible for correction when the same organization produced the capability, assessed it, and determined whether its own assessment required intervention?

Separation does not establish causal responsibility for every realized outcome, but it identifies in advance who was supposed to act and who was supposed to watch.

That is a weaker claim than saying separated institutions always govern better. The framework makes no such claim. Separation introduces costs: duplication, friction, delay, competence loss, and coordination burden. An independent observer may possess less frontier expertise than the organization being observed. A highly competent observer assembled from the same organizations and technical population may, however, inherit correlated blind spots.

The framework therefore confronts a real tension:

Competence ↔ Independence.

7.3.7 GP4 — Concentration and low contestability

GP4 states that when execution, observation, and steering are occupied by the same participants, observation does not supply genuinely independent information and responsibility becomes difficult to assign. Series #7 argues that this condition characterized the participants with the greatest binding power in frontier AI at the time of publication.

The proposition is deliberately contingent. It does not claim that functional concentration is necessary, permanent, or universal. Its empirical content concerns an institutional arrangement at a particular time.

The relevant falsifier is correspondingly concrete: identify an observer within the operative regime that jointly satisfies four conditions—no principal other than the governance function; observation at the level of realized outcomes; standing to compel a reasoned response; and independence of provenance. Partial satisfaction is not enough because the proposition concerns the conjunction.

7.3.8 The European AI Office: approaching the falsifier

Series #8 identified the European AI Office as the strongest case in its sample approaching GP4’s falsification condition. The point is not that the Office demonstrates successful governance. It is that statutory authority weakened one mechanism supporting the original concentration diagnosis.

From August 2, 2026, the Commission’s AI Office acquired enforcement powers over general-purpose AI models, including authority to request documentation, conduct model evaluations, require corrective measures, and impose or support sanctions under the AI Act (European Union 2026; European Commission 2026). Access therefore became a legal entitlement rather than a cooperation-dependent privilege.

For GP4, however, access is only one component. Realization-level observation would require evidence about actual deployment or use, the affected populations, realized outcomes and their distribution, and enough pathway information to connect those outcomes to the governed process. Capability testing or model access alone does not satisfy that burden. Provenance independence is also relational rather than nominal: the observation pathway must be sufficiently independent of the pathway whose blind spots are under test, not merely housed in a different organization.

On the Series #8 audit, the Office plausibly satisfies the no-other-principal condition and clearly possesses standing to compel response. It does not yet satisfy the full realization-level observation requirement, and provenance independence remains unestablished.

GP4 status: probed, not falsified—and not confirmed.

The case therefore moves the location of the shortfall. Organizational authority and compulsory access become stronger; realization-level observation and provenance independence remain unresolved. That weakening of one original mechanism belongs in the theory’s history rather than being treated as noise.

7.3.9 Formal institutions and actual observation

The European AI Office case exposes a broader methodological danger: institutional existence can easily be mistaken for functional performance.

A jurisdiction may create an AI office, safety institute, standards body, ombudsman, court, regulator, or reporting requirement. None of those labels establishes which governance function is actually being performed.

The correct questions are functional: What can the body observe? At which level—potential or realization? Whose information does it depend upon? Can it compel a response? Can it itself determine the response? Who sets the threshold against which its evidence is assessed? Who can challenge that threshold? Who adjudicates the challenge?

Institutional form does not identify governance function.

Function must be inferred from what the arrangement can actually do.

7.3.10 From function to standing

Functional separation tells us who executes, observes, steers, and procedurally adjudicates. It does not yet tell us whose interests enter those functions.

An independent observer can still observe the wrong population. A steering body can respond perfectly to information about actors already represented while leaving affected outsiders invisible. A procedurally impeccable adjudicative process can deny standing to those who experience the realization it governs.

The next problem is therefore not another institutional function. It is the relation between governance and the populations whose outcomes are at stake. That requires four linked concepts:

Affectedness → Standing → Exposure → Recourse.

7.4 Standing, Affectedness, Exposure, and Response Variety

The functional architecture identifies who executes, observes, steers, and procedurally adjudicates. It does not determine whose interests enter those processes.

That omission matters because governance can be internally well organized while remaining externally incomplete. An observer can be independent yet fail to observe the populations most affected. A steering process can respond efficiently to recognized disturbances while failing to recognize disturbances experienced outside its institutional field of view. A procedurally valid adjudicative process can still exclude people whose realized outcomes are at stake.

The Governance framework addresses this problem through standing. Definition 4 grounds standing in affectedness rather than institutional membership, technical expertise, ownership, or pre-existing decision power. Standing is therefore not merely a procedural permission granted to actors already inside the governing arrangement. It begins from the relation between an actor or population and the realization being governed.

The basic sequence is:

Affectedness → Standing → possibility of consequential contestation.

The arrows are not identities. Being affected does not guarantee effective standing, and formal standing does not guarantee that a challenge can alter an outcome.

7.4.1 Standing is not participation

Standing must remain separate from Participation. A person may possess standing without acting. Another may participate without possessing formal standing. An institution may invite participation while retaining complete authority to disregard what participants say. Conversely, an affected party may acquire a legally consequential route of challenge without taking part in broader deliberation.

Standing ≠ Participation.

Standing concerns whether a claim can properly enter a governance process because the claimant bears a relevant relation to the governed realization. Participation concerns who actually acts, what kind of act is performed, and what consequence follows.

The relevant question is therefore not simply “Who is present?” It is: Whose realized condition is being altered, and what route exists for that affectedness to become institutionally consequential?

7.4.2 Exposure gives affectedness structure

Series #8 makes that question more precise by inserting group-specific exposure into the realization mapping:

Ex_g(t) = ⟨Rh_g(t), Su_g(t), Rc_g(t)⟩.

Reach asks whether the capability or its consequences contact group g. Susceptibility concerns how strongly that group is positioned to absorb the relevant effect. Recourse concerns the mechanisms through which an affected party can contest, correct, or obtain remedy after exposure.

Series #8 further decomposes recourse as:

Rc_g(t) = (St_g(t), Ep_g(t), Ck_g(t), Rl_g(t)),

where St denotes an available standing institution, Ep relevant epistemic capacity, Ck an independent check, and Rl sufficient reliability for recourse to be meaningful.

Two groups can face the same nominal technology, the same governing instrument, and even the same formal rule while occupying very different realized positions. One may rarely encounter the capability. Another may encounter it continuously. One may possess institutional resources that buffer adverse effects. Another may possess little capacity to detect, challenge, or remedy them.

That is why exposure cannot be replaced by a binary classification such as “affected/not affected.” Nor can recourse be reduced to the existence of an appeals form or complaint channel.

Formal channel ≠ effective recourse.

7.4.3 GP1 — coupling, intervention loci, and exposure

Exposure sharpens rather than replaces GP1. The inherited proposition begins from coupling: beneficial and damaging potential can coexist within the same capability substrate, while governance intervenes on the mapping from potential to realization through access, sequencing, distribution, and deployment context.

Its empirical burden has two distinguishable parts. The coupling claim would require a capability trajectory capable of showing whether damaging potential can saturate or separate while beneficial potential continues to increase. Series #8 did not assemble such a trajectory. It therefore reports the coupling component as untested, while illustrating coexistence in cases where the same general capability family generated both beneficial and damaging applications.

The completeness claim is asymmetrical. It cannot be confirmed merely by observing many interventions that fit the four points of purchase. But it can be challenged by finding an intervention on the governed mapping that cannot be assigned to any of them. Exposure presented precisely such a test. It did not produce a fifth point of purchase.

The appropriate status is:

• GP1 coupling: untested; coexistence illustrated.
• GP1 completeness: falsifier exercised, not triggered; finitely non-refuted.

Finite non-refutation is not proof of completeness.

7.4.4 From affectedness to variety

Once different populations can occupy different exposure states, a second governance problem becomes visible. A governing arrangement must first recognize enough of the disturbances confronting it before it can respond appropriately to them.

This is the basis of GP2. In its abstract theoretical form:

Rv_env ≥ Dv_env,

where Dv_env is the relevant disturbance variety presented by the environment and Rv_env the response variety available to address it. Series #8 formalizes GP2’s stronger dynamic claim as the possibility that disturbance variety grows faster than response variety:

dDv_env(t)/dt > dRv_env(t)/dt.

The claim is about variety. It is not a claim that more rules are automatically better, that stricter rules imply greater variety, or that an arrangement with fewer categories is necessarily weaker.

Variety ≠ Specificity ≠ Stringency.

An instrument can become narrower but more precise. It can cover fewer categories while attaching stronger obligations to them. It can expand categories without improving enforcement. Counting response types alone cannot tell us which of these changes occurred.

7.4.5 Recognized variety is not environmental variety

The deepest difficulty for GP2 is not counting responses. It is identifying the disturbance space against which those responses should be compared.

A governing document contains only disturbances that the governing arrangement has recognized sufficiently to encode. But affected populations may experience disturbances that never become formal categories. If so, the document reveals something about the governor’s representation of the environment rather than the environment itself.

Series #8 therefore distinguishes underlying environmental variety from the variety recognized by the governing arrangement and from the subset recoverable by the coding instrument. At minimum:

Dv_arr ≤ Dv_env.

Exposure makes that identification problem more serious. A single formally recognized disturbance may correspond to several materially different governance problems across groups with substantially different reach, susceptibility, and recourse.

Recognized disturbance variety ≠ environmental disturbance variety.

7.4.6 What the GP2 measurement actually found

Series #8 attempted the measurement rather than leaving GP2 at the conceptual level. It coded successive versions of governing instruments and compared recognized disturbance variety with recognized response variety. Before examining the results, three outcomes were specified: response variety could keep pace with or exceed growing disturbance variety; disturbance variety could outgrow response variety; or neither could change materially.

The observed histories produced the third condition. Across the six version histories, neither recognized disturbance variety nor recognized response variety showed substantial growth under the documented coding protocol. The antecedent required by the growth-rate claim was therefore not exercised.

That result does not mean the two rates are equal in the environment. It does not establish that governance variety is sufficient. It does not falsify GP2. And it does not weakly support GP2 merely because the proposition sounds plausible in a rapidly changing technological environment.

GP2 is executable, measurement-limited, antecedent-unexercised, and not adjudicated.

This is more informative than simply writing “untested.” The measurement was performed; what failed was identification of the empirical relation required by the proposition.

7.4.7 Measurement changed the governance object

Running the variety instrument showed that a governing arrangement can change materially while recognized variety remains flat or moves in the opposite direction from other properties. Series #8 consequently introduces a vector-valued observable governance profile:

Gp_t = (Dv_t, Sp_t, Sg_t, Tr_t),

where Dv is recognized disturbance variety, Sp specificity, Sg stringency, and Tr revision transparency.

The empirical result is separability, not a ranking. No scalar governance index or dominance relation follows merely because some components increase.

A governance arrangement can become more specific and more stringent while recognizing fewer categories. Another can broaden recognized variety while becoming less transparent about revision. Which arrangement is “better” depends on evaluative criteria that this measurement does not supply.

7.4.8 Contestability is not the absence of failure

Standing, exposure, recourse, and variety also prevent a final conceptual shortcut. A contestable governance arrangement can still produce failure. An arrangement with meaningful standing and recourse does not guarantee that every harmful realization will be prevented. Conversely, the absence of an observed failure does not establish that affected parties possess meaningful contestability.

The relevant question is whether the arrangement can recognize differentiated consequences, admit affected claims, generate consequential response, and revise itself when existing categories or instruments prove inadequate.

GP1 and GP2 therefore remain deliberately incomplete. GP1’s intervention-locus component has survived a finite attempt at refutation while its coupling trajectory remains untested. GP2 has an executable discriminator, but available measurement does not identify the underlying environmental relation and did not exercise the growth antecedent. Neither result should be repaired by stronger prose.

7.5 Governance Dynamics in an Ongoing, Unfinished Transformation

The preceding sections describe governance structurally. GP3 and GP5 add time. GP3 asks whether an open-ended governance problem can reach a terminal solved state. GP5 asks how quickly governed outcomes are produced, detected, and corrected before any such state could be claimed. The propositions interact, but they should not be collapsed: one concerns non-terminality; the other concerns temporal decomposition.

7.5.1 GP3 — No terminal state

GP3 is deliberately titled “No terminal state (a corollary).” For an open-ended capability embedded in changing institutions, affected populations, and contested ends, no finite period of stability establishes that governance has been solved once and for all.

The claim follows primarily from the architecture already developed. The governed capability has no defined completion state; realization changes the environment; and procedural decision does not eliminate disagreement over ultimate ends. An arrangement may therefore be stable, durable, and effective for a period without becoming terminal in the stronger sense.

Stable ≠ Terminal.

This proposition has an unusual epistemic status. Contemporary churn is consistent with GP3 but does not prove it. Conversely, no single finite observation can falsify the strongest terminal-state claim. A governance arrangement that remained stable through decades of substantial capability, deployment, and institutional change would weaken the proposition’s practical importance, but finite duration alone cannot demonstrate finality.

GP3 status: analytically retained; terminal-state component not finitely adjudicable.

7.5.2 Recursion, revision, and transparency

GP3 belongs naturally within the Unified Model’s broader conception of Gradual AGI as an Ongoing, Unfinished Transformation. Governance can alter development incentives, permissions, exposure, recourse, deployment conditions, and institutional capacities; realized outcomes can in turn change political salience, liability, trust, and later governance.

Conceptually: GOV(t) → intervention at t → R_g(t) → changed conditions → GOV(t+1). This is feedback across time, not a fifth argument inside μ_t and not a new ontological object. Existing institutions and realization conditions simply enter new states.

Revision should therefore be interpreted rather than counted. A rule may change because capability changed, a previously unrecognized disturbance became visible, a new affected population became exposed, an instrument proved overly broad or narrow, or the institution learned from experience. The absence of revision may indicate stability, or it may indicate failure to detect, failure to act, or lack of authority.

Revision does not imply prior failure; absence of revision does not imply success.

Revision transparency asks whether later observers can reconstruct what changed and why. It is distinct from revision frequency and remains one component of the governance profile Gp_t = (Dv_t, Sp_t, Sg_t, Tr_t). A regime can revise frequently while documenting little, or revise rarely while preserving a clear evidentiary and procedural trail. Governance churn therefore has no intrinsic evaluative direction.

7.5.3 GP5 — Production, detection, and correction

GP5 turns from institutional finality to temporal performance. The original paper framed a trade-off between decision speed and the detection-and-correction capacity that may accompany broader participation or functional separation. More clearance points can slow a decision, while more independent sensing can sometimes improve the chance that a failure is found and acted upon.

That claim is conditional rather than axiomatic. More participation does not always add clearance points; more institutional separation does not guarantee useful observation; and some harms are so irreversible that post hoc correction is a poor governing objective. This chapter therefore retains GP5 only with its correctability boundary visible.

Series #8 clarifies the internal structure through three stages:

Production ≠ Detection ≠ Correction.

Production is the rate at which realized outcomes requiring governance attention are generated. Detection is the rate at which those outcomes become identified. Correction is the rate at which identified outcomes are acted upon. Only detection and correction receive symbols in Series #8—r_det and r_cor—because the empirical companion does not measure the production rate.

The decomposition is not merely verbal. A failure can be visible without being corrected; an affected party can detect an incident without authority or recourse; a regulator can receive a report without an applicable remedy; and an organization can contain a single incident without revising the conditions that produced it. Selected cases therefore support empirical separability of the stages, not general rate estimates.

GP5 structural decomposition: qualitatively supported. GP5 rate claim: untested.

7.5.4 Formal and substantive detection; movable bottlenecks

Empirical contact also showed that detection itself is heterogeneous. Series #8 distinguishes formal verification from substantive comprehension:

r_det = (r_det^f, r_det^s).

The motivating mathematics cases showed that an artifact can be generated and formally verified faster than a community can understand, interpret, or independently reconstruct it. The governance analogue is broader: an audit can complete without the auditor understanding the causal pathway; a compliance check can verify specified fields without establishing that the system behaves safely in context; a threshold violation can be detected before its meaning is understood.

Formal Detection ≠ Substantive Detection.

The distinction does not establish two universal detection rates and supplies no aggregation rule. Its value is diagnostic. If production accelerates while detection remains fixed, unnoticed outcomes accumulate. If detection accelerates while correction remains fixed, known but unresolved problems accumulate. If formal detection scales while substantive understanding does not, verified but poorly understood artifacts accumulate.

This is a movable-bottleneck problem. Improving one stage can expose the next. A regulator that issues rules quickly may still detect slowly; an organization that detects rapidly may correct selectively; a system that corrects individual incidents may still fail to revise the categories through which later disturbances are recognized.

7.5.5 Irreversibility and the limit of correction

Correction-centered governance applies only where consequential correction remains possible. If a realized consequence cannot be undone, meaningfully compensated, or prevented from propagating once triggered, increasing r_cor after realization has limited value.

The governance burden then moves upstream toward access restrictions, sequencing, attenuation, precautionary thresholds, or deployment-context controls. This is not a rescue clause for GP5; it is a boundary condition inherited from the source framework. It also echoes Chapter 6: revisability is valuable only while relevant options remain open.

Other temporal processes certainly exist—legislative schedules, litigation, election cycles, deployment cycles, standard-setting, research progress, and public mobilization. Chapter 7 does not convert each into an additional formal clock. The canonical GP5 decomposition remains production, detection, and correction, with formal/substantive detection as an empirical refinement.

7.5.6 Institutional memory and direction without terminality

Repeated revision creates a further risk: governance can learn, but it can also forget. Policies are overwritten, personnel turn over, rationales disappear, and version histories preserve text without preserving the reasons that gave a rule its meaning. Revision transparency therefore supports both external accountability and institutional memory.

Non-terminality nevertheless does not imply institutional relativism. Series #7 retains detectability, contestability, and correctability as dimensions that can be investigated without assuming a universal governance-success scalar. They can orient inquiry—Can failures be seen? Can claims enter? Can correction still matter?—without producing a final institutional blueprint.

The limit of those dimensions appears when actors disagree over what counts as an error in the first place. That is not a rate problem. It is the problem of contested ends, to which §7.6 turns.

7.6 Contested Ends, Adjudication, and Provenance

Governance can recognize disturbances, observe consequences, and respond over time without resolving disagreement over ultimate ends. This is the point at which Governance most clearly separates from Optimization. Sometimes the dispute is not about how best to achieve an agreed objective; the objectives themselves differ, and no common scalar is available to dissolve the disagreement.

7.6.1 GP6 — the adjudication asymmetry

GP6 distinguishes procedural closure from settlement of ultimate ends. Courts, regulators, appeals bodies, contracts, and internal review processes can determine authority, standing, valid procedure, or remedy. Those decisions can bind even when parties continue to reject the substantive objective behind them.

Decision Closure ≠ Normative Settlement.

The distinction assigns a limited but important burden to Procedural Adjudication. It can determine whether an actor was entitled to decide, whether required process was followed, whether a challenge had standing, or what remedy attaches to a violation. It does not thereby demonstrate that the underlying end is universally correct.

7.6.2 Direct empirical contact: the Anthropic–Department of War dispute

Series #8 found its strongest direct support for GP6 in the 2026 dispute between Anthropic and the U.S. Department of War. The case has since moved beyond the March preliminary-injunction posture used in the original empirical companion.

On August 27, 2026, the U.S. District Court for the Northern District of California granted Anthropic summary judgment on its First Amendment and Due Process claims and on its Administrative Procedure Act challenge to the Hegseth Directive and supply-chain-risk designation. The court found the challenged measures unlawful while stating that the Department remained free to select the AI vendor of its choice (U.S. District Court, N.D. Cal. 2026).

That boundary is analytically clean. The court could decide whether government action complied with constitutional, statutory, and procedural constraints. It could vacate unlawful measures and supply binding legal closure. It did not decide the underlying substantive question of which AI system the Department should use, what safety restrictions should govern military or surveillance applications, or how competing national-security and safety objectives should be balanced.

The later summary-judgment posture therefore strengthens the original mechanism-level observation without broadening its scope. Procedure received authoritative settlement; the contested end remained a matter for political, contractual, and institutional choice.

GP6: direct case support, not prevalence evidence.

One unusually clear dispute establishes that the configuration can occur. It does not establish how often it occurs across AI governance, nor does it validate every part of the broader framework.

7.6.3 Authority, contestation, and collective choice

The case exposes a broader distinction. Authority to decide is not the same as agreement that the decision is substantively correct. Treating the two as equivalent produces either procedural inflation—assuming lawful decision proves normative correctness—or procedural nihilism—assuming continuing disagreement makes binding decision impossible. The framework adopts neither.

Nor does visible contestation automatically generate collective choice. Actors may publish opposing statements, lobby governments, litigate, resign, organize coalitions, or compete through markets. These activities can change incentives and political conditions, but they do not automatically create a rule that converts disagreement into an authoritative common selection.

Contestation ≠ Collective Choice ≠ Normative Settlement.

Series #8’s open-weights episode illustrated that gap: positions over openness, export controls, distillation, testing, and related policy were public and rapidly contested, yet no single forum converted those positions into a binding settlement over the underlying direction. Contestation was observable; authoritative collective choice remained distributed.

7.6.4 Exit, Entry, and the Authority to Choose a Trajectory

September 2026 supplied a useful paired illustration of the same problem at the frontier-laboratory level. Jacob Coxon resigned from Anthropic and publicly challenged the adequacy of frontier laboratories as private sites for choosing potentially civilization-scale trajectories. The following day, OpenAI appointed Paul Christiano to the OpenAI Foundation Board, its Safety and Security Committee, and a non-voting observer role on the OpenAI Group PBC Board (OpenAI 2026a).

Christiano’s own statement makes the contrast analytically useful. He wrote that recent capability progress and continuing alignment difficulty had led him to assign meaningful near-term probability to catastrophic loss of control and that the industry, including OpenAI, was not on track to reduce that risk to a level he regarded as acceptable. Yet he chose to enter OpenAI’s governance structure because he believed stronger internal oversight could still reduce the danger (Christiano 2026).

No evidence considered here establishes that Christiano’s appointment was a response to Coxon or that the two events were coordinated. Their relevance is structural, not causal. Coxon represents exit plus external contestation; Christiano represents entry plus internal governance.

Preferred trajectory ≠ authoritatively selected trajectory.

The pairing makes the Chapter 6→7 transition concrete. Optimization can identify or defend a trajectory under stated objectives, constraints, horizons, and uncertainty assumptions. Governance becomes necessary when that preference must acquire standing, board-level authority, enforcement capacity, or the ability to constrain execution.

Christiano’s appointment therefore changes the institutional location of a safety-oriented evaluative perspective. It gives that perspective a route into steering and oversight rather than leaving it solely in public argument. But the move does not resolve Coxon’s deeper challenge: whether private-company governance is itself an adequate locus of authority for decisions whose consequences may extend well beyond the firm.

The contrast also sharpens the distinction between function and independence. Internal safety representation can increase functional differentiation without creating an independently situated realization-level observer. OpenAI can describe Christiano as an independent voice while, after appointment, he becomes formally embedded in the institution. That does not negate independence of judgment; it simply shows why organizational position, intellectual provenance, and observation-pathway independence are different variables.

For that reason, the paired episode remains illustrative. It does not validate GP4, GP6, or GP7 and does not support a prevalence claim.

7.6.5 GP7 — provenance and the clean discriminator

Observation is shaped by the path through which observers acquire methods, categories, incentives, information, and institutional position. Provenance therefore extends beyond employment to shared training, professional culture, common data sources, methodological inheritance, funding, organizational ancestry, or another pathway capable of producing correlated observational structure.

Organizational independence ≠ provenance independence.

GP7 asks whether shared provenance can produce correlated blind spots. Let F denote the latent population of relevant failures and O_p(F) the failures recovered through an observation process of provenance p. The empirical burden is not merely to show that two observers disagree. It is to show that shared provenance is associated with correlated omissions relative to a sufficiently independent benchmark.

The benchmark requirement matters because otherwise the test is circular. If the benchmark failure population is generated through the same observation pathway being evaluated, failures that pathway systematically misses cannot appear in the benchmark. Apparent agreement may then reflect shared omission rather than convergent accuracy.

7.6.6 Why GP7 is presently gated

The current case record does not provide the independent benchmark population needed for a discriminating GP7 test. Protected reporting may expose problems that internal evaluation did not surface; independent researchers may detect vulnerabilities absent from developer reports; affected users may reveal realized failures invisible in predeployment testing; third-party audits may identify different deficiencies. Those examples exercise the mechanism but do not establish the required covariance.

GP7 status: specified, presently gated/non-operable, untested.

An executable study would need at least five elements: a stated provenance dimension; two or more observation processes; a bounded failure domain; a rule for matching comparable failures; and a benchmark-construction process sufficiently independent of the provenance under test. The last requirement means that “independent” must always be specified relationally: independent with respect to which data source, funding relationship, professional network, methodology, or organizational dependency?

Candidate mechanisms include protected reporting, randomized external inspection, independent incident repositories, and third-party audit. None guarantees provenance independence. An incident repository populated only through developer disclosure inherits developer observability; an external audit staffed through the same narrow technical network may reproduce methodological blind spots; a community reporting channel may have different provenance but insufficient technical access.

Their importance is narrower: the gating condition is institutional rather than logically impossible. Governance can, in principle, build observation pathways that make stronger tests possible.

7.6.7 Provenance contestation and reflexivity

The Coxon episode also illustrates why provenance evidence must not become ad hominem inference. Some public commentary questioned whether his warning reflected independent insider judgment or coordinated advocacy. The existence of that dispute is observable; the truth of the coordination allegation is a separate claim.

Provenance contested ≠ provenance dependence established.

The disciplined sequence is: identify provenance → identify a plausible dependency → specify its predicted observational consequence → test against sufficiently independent evidence. Employment, funding, equity, co-authorship, or professional network can be relevant provenance information without proving bias, and bias—if established—would not by itself make an underlying factual claim false.

The same rule applies reflexively to this research program. Series #8 disclosed overlapping relationships involving Anthropic and used role separation, source triangulation, and review controls. Those practices can reduce identifiable dependencies. They are not independent replication and therefore do not count as evidence for GP7.

7.6.8 Decision closure, normative settlement, and epistemic closure

GP6 and GP7 together expose three forms of closure that should not be confused:

Decision closure: an institution authoritatively decides what happens next. Normative settlement: the underlying disagreement over ends is resolved. Epistemic closure: the relevant facts and failures have been sufficiently observed.

Decision Closure ≠ Normative Settlement ≠ Epistemic Closure.

A court may provide decision closure while normative disagreement persists. A community may reach broad agreement about an end while lacking evidence about realized effects. An observation process may establish a failure with high confidence while no institution possesses authority to correct it.

The proposition statuses therefore remain deliberately asymmetric: GP6 has direct case-level support; GP7 has a specified discriminator but no adequate benchmark. Section 7.7 turns from those propositions to what the framework became after empirical contact.

7.7 Governance Under Empirical Contact

The Governance framework was not designed to survive empirical contact by interpreting every result as confirmation. Table 7.1, placed near the chapter’s outset, records deliberately unequal statuses: direct case support, finite non-refutation, non-adjudication, non-finite adjudicability, and non-operability under current evidence. Those labels are not grades and should not be collapsed into supported versus unsupported.

7.7.1 What measurement changed

The most important empirical contribution was framework correction rather than proposition confirmation.

First, the affected-party gap identified in Series #7 became an explicit group-specific exposure state without creating a fifth point of purchase. Second, detection split into formal verification and substantive comprehension. Third, governance measurement became vector-valued through the four-component governance profile, separating recognized disturbance variety, specificity, stringency, and revision transparency. Fourth, the European AI Office case clarified that statutory access, realization-level observation, and provenance independence are different institutional achievements.

These are substantive revisions. They do more theoretical work than a sequence of weak confirmation claims would have done because they make the framework more discriminating about what is being observed and what remains unresolved.

7.7.2 Separability before ordering; cases before prevalence

Measurement can establish that governance dimensions differ before it establishes how they should be ordered. Variety, specificity, stringency, and revision transparency can move independently. A framework may become more specific while recognizing fewer disturbance categories, or more stringent while becoming less transparent about revision.

The empirical result is separability, not dominance. The framework therefore refuses to infer that the governance profile at time t+1 dominates the profile at time t without a separately justified evaluative rule.

The same discipline applies to cases. A case in which procedural adjudication leaves ultimate ends unresolved establishes that the GP6 configuration can occur; it does not estimate its prevalence. A case in which formal and substantive detection separate establishes empirical separability; it does not estimate the size of the gap across AI governance. Several developer-centered cases do not establish that all frontier governance shares the same structure.

Case-level evidence ≠ prevalence evidence.

This limitation is especially important in a rapidly changing institutional environment. The European AI Office already shows how a post-publication institutional change can weaken one mechanism behind a proposition while leaving other shortfalls intact. The cases are therefore best read as tests of discriminability, operability, and mechanism.

7.7.3 Instruments, blind spots, and the GP7 meta-hypothesis

GP2 revealed a second measurement limit. Governance instruments expose the disturbances they recognize, not the whole disturbance environment. A regulation, standard, or safety framework contains categories that have already passed through institutional recognition. Coding can recover changes in that representation without revealing what remains outside it.

Exposure intensifies the problem because one formally recognized disturbance may correspond to materially different realization conditions across groups with different reach, susceptibility, or recourse. Empirical governance research therefore faces a two-stage observation problem: observe governance, and where possible observe what governance fails to observe.

GP7’s missing independent benchmark establishes non-operability of the present test. Whether the same absence also signals an institutional deficit in the observation environment is a further hypothesis generated by that limitation, not a finding of GP7. The absence of a benchmark tells us what stronger research would require; it does not prove that existing observers already share blind spots.

7.7.4 Negative results, reproducibility, and archival fragility

Several propositions did not receive the evidence their strongest formulations require. GP1 lacks the necessary coupling trajectory; GP2’s growth antecedent was not exercised; GP3 cannot be finitely settled in its strongest form; GP5 lacks general rate estimates; and GP7 lacks an adequate benchmark. Those are results about the present evidence boundary, not invitations to substitute suggestive anecdotes.

Governance evidence is also unusually vulnerable to archival loss. Policies are revised in place, web pages are overwritten, and earlier model cards or safety frameworks become difficult to recover. Longitudinal claims about variety, specificity, stringency, or revision transparency depend on reconstructing prior institutional states.

Archival persistence is therefore part of the observational environment required for reproducible governance research. This chapter does not elevate preservation into a fifth governance function; it records it as a condition for studying change reliably.

7.7.5 What empirical contact licenses

A smaller set of conclusions survives more strongly. Potential and realization must remain distinct; affected populations require explicit exposure treatment; governance dimensions can be separable; production, detection, and correction should not be collapsed; formal verification can separate from substantive understanding; institutional access is not realization-level observation; and procedural closure can coexist with unresolved ends.

The framework became more discriminating after contact with evidence.

The remaining task is therefore not to strengthen every proposition. It is to state exactly what Governance can and cannot claim.

7.8 What Governance Can—and Cannot—Claim

The framework is diagnostic and conditional before it is prescriptive. Its value depends as much on what it refuses to infer as on the distinctions it introduces.

7.8.1 Core nonclaims

No universal governance-success scalar follows from Gp_t, Ex_g(t), the four functions, or the four points of purchase. A bounded decision may justify its own weights, thresholds, or admissibility rules, but those choices belong to that specified evaluative problem rather than to Governance as such.

More governance is not automatically better. Additional review can expose failures and also create delay; stronger restrictions can prevent harmful realization and also foreclose beneficial realization; wider participation can reveal otherwise invisible interests without guaranteeing influence. The relevant evaluation depends on objectives, affected populations, alternatives, uncertainty, reversibility, and opportunity cost.

Several recurring non-identities should therefore be read as boundaries rather than as slogans: Formal Participation ≠ Effective Contestability; Detection ≠ Correction; Institutional Existence ≠ Effective Observation; Stability ≠ Terminality; Procedural Settlement ≠ Substantive Settlement. Each marks a different point at which formal structure can be mistaken for realized governance performance.

The European AI Office illustrates the point. Statutory powers, model access, and authority to request corrective measures are real institutional changes. They should be credited as such. They do not by themselves establish realization-level observation across affected populations or provenance independence. Partial functional improvement is not completed governance.

The empirical companion also supports no prevalence estimate. Case and instrument evidence can discriminate mechanisms without establishing their population frequency. GP6’s direct support remains one dispute; GP4’s probe remains one institutional trajectory; GP5’s structural separation remains qualitative rather than a rate estimate.

7.8.2 Governance is not a complete theory of legitimacy

The framework can identify authority, standing, contestability, recourse, observation, procedural validity, and the distribution of governance functions. It does not derive a universal principle specifying when authority is morally legitimate.

Questions of democratic authorization, constitutional legitimacy, consent, rights, sovereignty, representation, distributive justice, and political obligation require additional normative argument. This is not a defect to be repaired by silently importing one theory of legitimacy. Black’s account of polycentric regulation, discussed in §7.1.2, is useful precisely because it treats legitimacy and accountability as relational and contested without claiming a single grand solution.

The chapter should therefore mark evaluative language carefully. “Adequate realization-level observation,” for example, is a functional criterion relative to the governed object and the proposition being tested; it is not a claim that the observing institution is legitimate in the broader political-philosophical sense. Likewise, “greater contestability” describes more effective capacity for challenge, not an unconditional claim that every additional opportunity for challenge improves every decision.

7.8.3 Boundaries with adjacent frameworks

Governance interacts with the rest of the Unified Model without absorbing it. It is not a fifth input to R_g(t) = μ_t(P(t), Av(t), Ex_g(t), Z_d(t)); it can alter those conditions and the pathways around them without becoming one undifferentiated coefficient. Governance does not absorb Optimization: authority to choose does not establish evaluative correctness. Governance does not absorb Participation: institutions can create standing, channels, and recourse without establishing who actually acts or what follows.

Synchronization remains distinct as well. A governance intervention may change a Ceiling, Floor, delay, or the conditions under which a CFS gap is meaningful, but institutional improvement is not equivalent to faster gap closure. Realization likewise remains the process by which heterogeneous conditions become consequential; Governance acts on that process without replacing it.

Table 7.2. Boundaries: What Governance Does and Does Not Own

Framework	Governing question	What it owns	What Governance must not absorb
Realization	How does potential become consequential under heterogeneous conditions?	P(t), Av(t), Ex_g(t), Z_d(t), and the mapping μ_t	Governance can modify realization pathways but is not another universal input to μ_t.
Synchronization	How does a bounded capability–realization gap evolve?	Ceiling, Floor, Slope, gap dynamics, delay, and applicability conditions	Governance must not reinterpret every institutional problem as a CFS gap or every governance improvement as faster gap closure.
Optimization	Which candidate trajectory should be preferred under specified evaluative conditions?	Objectives, constraints, horizons, uncertainty, candidate trajectories, admissibility, selection rules	Governance does not supply a universal objective or convert authority into evaluative correctness.
Governance	Who can shape, observe, steer, contest, adjudicate, and revise realization pathways?	Governed object, intervention loci, functions, standing, contestability, response structure, procedural adjudication, provenance-sensitive observation	Governance does not claim a universal success scalar, final institutional form, or complete theory of legitimacy.
Participation	Who participates, through what acts, and with what consequences?	Participatory Subjects, Acts, relational grammar, recognition, uptake, and Participatory Consequences	Governance can create participatory structures but cannot infer actual participation or its consequences from their existence.

7.8.4 The chapter’s positive claim

Governance matters because realization is neither automatic nor institutionally neutral. Capabilities become consequential through pathways that can be opened, narrowed, sequenced, observed, contested, constrained, redirected, corrected, or foreclosed. Those pathways contain authority, information asymmetries, unequal exposure, and disputes about ends.

The four points of purchase and four functions provide a vocabulary for locating those processes without pretending that one institutional design controls the whole pathway. Standing, exposure, and recourse identify whose realized condition may differ; the proposition ledger identifies which claims have actually survived empirical contact and at what status.

Governance organizes the authority and contestation through which realization pathways are shaped over time.

That claim is stronger than treating governance as a downstream regulatory afterthought and weaker than claiming governance determines realization. It is the level of claim the architecture and evidence support.

7.9 From Governance to Participation

Chapter 7 began where Optimization stopped: when objectives and constraints must become authoritative across heterogeneous actors. It ends at a second boundary. Governance can create structures of standing, observation, contestability, recourse, steering, and adjudication; it cannot infer the acts that enter those structures or the consequences that follow.

7.9.1 Institutional possibility versus realized act

Standing is a property of the governance relation. Participation concerns what an actor actually does within, around, against, or outside that relation. A worker may possess formal standing and remain silent; another may lack recognized standing and organize publicly. A community may consult, litigate, protest, refuse, negotiate, or co-create. Those acts are not fungible merely because they all count as participation.

Institutional possibility ≠ participatory act.

7.9.2 Recognition, uptake, and consequence

An institutional process may recognize an act without taking it up in a way that changes subsequent action. A regulator can receive a complaint, a board can hear a dissenting member, or a government can hold a consultation without altering the decision.

Recognition ≠ Uptake.

Governance asks whether a structure existed through which the act could enter. Participation asks what act entered, how it was structured, whether it was recognized or taken up, and what consequence followed. Contestability becomes empirically visible through the combination of channel, act, and response.

7.9.3 Exit, entry, and the move to Participation

The Coxon–Christiano contrast provides a final bridge. Governance can describe board authority, safety oversight, organizational decision rights, public regulation, and external coordination. Participation asks why one researcher exited and contested publicly while another entered an institutional governance role, how those acts differ relationally, and what consequences each produced.

Participation can also change governance itself: resignation can trigger scrutiny, litigation can redefine authority, a public campaign can create a new forum, and a board appointment can alter internal steering. Participation is therefore not simply something that happens inside Governance; it can recursively change the environment into which later acts occur.

7.9.4 Chapter conclusion

Governance in an Ongoing, Unfinished Transformation is neither a final institutional blueprint nor a scalar measure of control. It is an architecture of authority, observation, contestation, steering, adjudication, and revision around realization pathways.

Its propositions survive empirical contact unevenly, as Table 7.1 records. The framework also changed under measurement: exposure became explicit; formal and substantive detection separated; governance measurement became vector-valued; and institutional access was distinguished more sharply from realization-level observation and provenance independence.

The resulting framework does not tell us that more governance is always better, supply a universal scalar, settle contested ends, or infer actual participation from participatory institutions. It provides a disciplined account of where authority and contestation enter the movement from potential to realized consequence.

Governance identifies the structures through which realization can be contested. Participation asks who enters those structures, what they do there, and what follows. Chapter 8 begins there.

References and Contemporary Source Notes

Black, Julia. 2008. “Constructing and Contesting Legitimacy and Accountability in Polycentric Regulatory Regimes.” Regulation & Governance 2(2): 137–164. https://doi.org/10.1111/j.1748-5991.2008.00034.x.

Christiano, Paul. 2026. “Personal Statement on Joining the OpenAI Board.” September 9, 2026. https://paulfchristiano.substack.com/p/personal-statement-on-joining-the.

Dafoe, Allan. 2018. AI Governance: A Research Agenda. Centre for the Governance of AI, Future of Humanity Institute, University of Oxford. August 27, 2018. https://www.governance.ai/research-paper/agenda.

European Commission. 2026. “AI Act.” Shaping Europe’s Digital Future. Updated August 2026. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai.

European Union. 2026. Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated version of July 27, 2026. EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng.

Fung, Archon. 2006. “Varieties of Participation in Complex Governance.” Public Administration Review 66(s1): 66–75. https://doi.org/10.1111/j.1540-6210.2006.00667.x.

National Institute of Standards and Technology (NIST). 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1.

OpenAI. 2024. “An Update on Our Safety & Security Practices.” September 16, 2024. https://openai.com/index/update-on-safety-and-security-practices/.

OpenAI. 2026a. “Paul Christiano Joins OpenAI Foundation Board.” September 9, 2026. https://openai.com/index/paul-christiano-joins-openai-foundation-board/.

OpenAI. 2026b. “Our Structure.” Accessed September 11, 2026. https://openai.com/our-structure/.

Raji, Inioluwa Deborah, Andrew Smart, Rebecca N. White, Margaret Mitchell, Timnit Gebru, Ben Hutchinson, Jamila Smith-Loud, Daniel Theron, and Parker Barnes. 2020. “Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing.” Proceedings of FAT* ’20, 33–44. https://doi.org/10.1145/3351095.3372873.

Taeihagh, Araz. 2021. “Governance of Artificial Intelligence.” Policy and Society 40(2): 137–157. https://doi.org/10.1080/14494035.2021.1928377.

U.S. District Court for the Northern District of California. 2026. Anthropic PBC v. U.S. Department of War et al., Case No. 26-cv-01996-RFL, Order on Cross Motions for Summary Judgment, Filing 250, August 27, 2026. https://docs.justia.com/cases/federal/district-courts/california/candce/3:2026cv01996/465515/250.

W.H.L. and Claude (Opus 5). 2026a. “Gradual AGI as Contestation: A Framework for Governance.” Gradual AGI Series #7. August 3, 2026. Current live version v1.9. Champaign Magazine. https://champaignmagazine.com/2026/08/03/gradual-agi-as-contestation-a-framework-for-governance/.

W.H.L., Claude (Opus 5), and GPT-5.6 Sol. 2026b. “Gradual AGI as Contestation: Measuring Governance Under Empirical Contact.” Gradual AGI Series #8. August 11, 2026. Current live version v1.4.6. Champaign Magazine. https://champaignmagazine.com/2026/08/11/gradual-agi-as-contestation-measuring-governance-under-empirical-contact/.

W.H.L., Claude (Sonnet 5, Opus 5). 2026c. “Gradual AGI as Optimization: Formal Models and Empirical Tests.” Gradual AGI Series #6. Originally published July 24, 2026; current live version v1.4, July 26, 2026. Champaign Magazine. https://champaignmagazine.com/2026/07/24/gradual-agi-as-optimization-formal-models-and-empirical-tests/.

W.H.L., GPT-5.6 Sol, and Claude Sonnet 5. 2026d. “Gradual AGI as Participation: Subjects, Acts, and Consequences.” Gradual AGI Series #9. Publication Version v1.0, August 27, 2026. Champaign Magazine. https://champaignmagazine.com/2026/08/27/gradual-agi-as-participation-subjects-acts-and-consequences/.

Zeff, Maxwell. 2026. “The AI Researcher Who Just Quit Anthropic Says It’s ‘Crunch Time for Humanity.’” WIRED. September 9, 2026. https://www.wired.com/story/anthropic-researcher-quits-jacob-coxon-ai-fears-humanity/.