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

The Realization Architecture of Gradual AGI

Chapter 4 of the forthcoming book On Gradual AGI, based on Champaign Magazine’s Gradual AGI series installments

Publication Version: v1.2. September 9, 2026

Chapter 3 established the methodological architecture of the Unified Gradual AGI Model and described Gradual AGI as an Ongoing, Unfinished Transformation: an evolving transformation whose realization is not exhausted by a single event and whose relevant forms, relationships, and consequences remain open to further change. It also assigned the substantive realization architecture to this chapter.

Core question: How does AI capability become socially realized—and why does capability alone fail to determine that realization?

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

Here P(t) represents Potential, Av(t) Availability, Ex_g(t) group-specific Exposure, Z_d(t) domain-specific realization conditions, and μ_t the structured relation through which these heterogeneous conditions produce a realized state R_g(t). The arguments are not assumed scalar, additive, interchangeable, or individually sufficient.

The architecture admits a family of possible realization mappings. In any specified domain and application, μ_t denotes the particular realization relation being posited for that setting and time. The framework does not assume that mappings across domains share one functional form, nor that a specified μ_t is continuous, differentiable, unique, parametrically stable, or identifiable from observational data alone. In comparative applications, analysts should state explicitly whether the same realization relation is hypothesized across domains or whether domain-specific mappings are being compared.

The indices g and d are analytically distinct but not assumed independent. Groups are situated within domains, and domain-specific conditions can shape both group-specific Exposure and the realized states experienced by those groups.

4.1 From Potential to Realization

A system can become dramatically more capable without those capabilities being fully available, widely encountered, institutionally accepted, safely deployable, or socially consequential. Capability-centered accounts often imply that once a sufficiently capable system exists, the rest follows. Gradual AGI begins from a different premise: potential is not realization.

A capability may remain expensive to access, difficult to deploy, legally constrained, distrusted, poorly integrated into institutions, or deliberately withheld because its risks cannot yet be managed. Conversely, a technically modest capability may become highly consequential when it is inexpensive, ubiquitous, embedded in ordinary workflows, and connected to institutions able to act on its outputs.

The Gradual AGI series reached this problem cumulatively. Epistemic Extension rejected a single arrival event; Abundant Resources separated capability from the conditions that make it broadly usable; Synchronization showed that machine and institutional clocks can diverge; CFS formalized one family of those dynamics; Optimization rejected a single global objective; Governance made the potential-to-realization mapping governable; its empirical companion introduced Exposure as first-class; and Participation separated human Subject, Act, Consequence, and realized change.

Potential ≠ Availability ≠ Exposure ≠ Domain-specific realization conditions ≠ Realization

These five terms do not occupy the same formal position. The first four are heterogeneous arguments entering the mapping; the fifth is its realized output. Potential concerns what AI makes possible. Availability concerns whether that possibility has become practically obtainable. Exposure concerns how particular groups are situated relative to the technology’s reach and effects. Domain-specific realization conditions concern additional institutional, infrastructural, legal, organizational, safety, trust, procurement, and contextual states. Realization concerns what becomes instantiated in the world.

The non-identity symbols distinguish analytical objects; they do not assert statistical or causal independence. The objects can interact strongly, and a realized state can depend on conditions that are themselves mutually related, while the distinctions remain analytically useful.

Table 4.1 — Core Objects of the Realization Architecture

Object	Core question	What it is not
P(t) — Potential	What can AI make possible?	Availability or realized consequence
Av(t) — Availability	How practically obtainable has that capability become?	Adoption or Exposure
Exg(t) — Exposure	How is group g reached, affected, and positioned for recourse?	General Availability or welfare
Zd(t) — Domain conditions	What additional conditions materially structure realization in domain d?	A residual scalar or exhaustive checklist
μt — Realization mapping	How do heterogeneous conditions become consequential?	A fifth input or universal coefficient
Rg(t) — Realized state	What has actually become instantiated for the specified realization object and group?	Capability itself

Figure 4.1. Core realization architecture. The four curved inputs denote typed participation in the realization relation, not chronological passage or statistical independence. The output connection denotes the realized state for the specified realization object. The figure is non-additive, non-pipeline, and not a complete causal-edge inventory.

4.1.1 Contemporary Views of the Realization Gap

Frontier development

OpenAI’s September 6, 2026 report Research acceleration: The view inside OpenAI describes coding agents becoming embedded in frontier AI research. By mid-August, OpenAI reported 3.1 agent-workdays for every human workday in its research organization, alongside rising experiment rates and longer-horizon delegated tasks. The report explicitly cautions that such metrics should not be equated one-for-one with research progress: less automatable activities become relatively more important, compute can become a gating factor, difficult work continues to require human steering, and people still choose priorities and decide what to pursue, scale, pause, or deploy (OpenAI 2026a).

In the realization architecture, the 3.1 figure is evidence of increased practical availability and use of agentic research capacity inside the organization; it is not itself a measure of R_g(t) as overall research progress. OpenAI’s discussion of compute, human steering, safety restrictions, and prioritization illustrates why the mapping from increased agent availability to realized research progress remains conditional.

A second OpenAI publication from the same day approaches the frontier from the constraint side. In An Alien Mind, Chief Scientist Jakub Pachocki argues that rapid capability growth may continue while alignment, monitoring, confidence in safety, human control, and international coordination fail to keep pace. He explicitly anticipates voluntary slowdowns until shared safety bars exist (Pachocki 2026).

Neither OpenAI publication validates this architecture or the CFS recurrence. Together they illustrate a broader possibility: acceleration at one stage can relocate the limiting condition, and potential can outrun the conditions governing its realization.

Source note. The OpenAI and a16z materials used here are self-reported corporate or practitioner sources. They are treated as contemporaneous evidence about reported practices, constraints, and decision frames—not as independent validation of the architecture. Quantitative claims are limited to what the originating sources report unless independently corroborated. Independent corroboration can include third-party audits, replicated measurements, external datasets, or convergent evidence from sources not controlled by the reporting organization.

Commercial adoption

Joe Schmidt and Julian Marx’s July 2026 a16z framework distinguishes Lighthouse from Landgrab markets. In Lighthouse settings, strong product performance and favorable economics may still be insufficient because buyers seek credible precedent, social proof, and protection against institutional or career downside. In Landgrab settings, the problem is already understood, risk is more manageable, and demonstrated return can support rapid broad adoption. The August 13 episode of The a16z Show develops the same contrast in discussion among Elena Burger, Andy McCall, and Joe Schmidt (Schmidt and Marx 2026; a16z 2026).

Again, capability can already exist, the product can already work, and Availability can already be comparatively high. Realization can nevertheless depend on trust, procurement, perceived downside, social proof, and organizational willingness to act.

Knowledge assimilation

Mathematics now supplies both a frontier-capability event and a realization test. On September 8, 2026, OpenAI reported a proposed resolution of the Navier–Stokes Millennium Prize Problem produced by an internal AI system significantly more capable than its public models. OpenAI reports coordinating on the order of 10,000 agents for roughly 88 hours, followed by another 17 hours of Lean formalization and verification. Whether the proof ultimately receives broad mathematical and institutional acceptance remains open. Even so, the episode is evidence of a sharply advanced frontier in AI mathematical Potential: a Millennium-level research problem has become a plausible target of large-scale autonomous AI research rather than merely AI-assisted calculation or exposition (OpenAI 2026b; Clay Mathematics Institute 2026, 2018).

Terence Tao’s Mathematics in the age of AI asks what happens when AI can perform research-level mathematics and shifts attention from capability to the goals and values of mathematics. Proof generation may accelerate much faster than verification, exposition, digestion, community acceptance, and canonicalization. A proof can therefore exist and even be verified while another realization object—shared mathematical understanding—lags (Tao 2026a).

Tao’s September 2026 discussions sharpen the point. In Answers versus insight, he separates obtaining the answer to a problem from learning what the problem can teach; historically the two were tightly aligned because the route to the answer ran through the study of why existing approaches failed. AI can weaken that alignment (Tao 2026b). In his Navier–Stokes worked example, Tao argues that much of the mathematical value lies in discovering why successive plausible constructions fail, and that this process is compromised if the final answer is already available in advance. A technically solved problem can therefore add little of the mathematical value normally generated by the struggle to solve it (Tao 2026c).

This gives Chapter 4 a stronger realization principle: AI can realize the answer while bypassing the path through which humans would otherwise have realized understanding.

Figure 4.2. Three locations and forms of the realization gap. The annotations identify principal structures implicated in each case; they do not claim that any case is governed by a single variable or that the cases instantiate one common recurrence.

Taken together, these cases support two chapter-level claims. First, the realization gap is not one gap with one cause; it is a family of domain-specific separations between what AI makes possible and what AGI-inclusive humanity is presently able, willing, permitted, prepared, or able to absorb into durable practice and knowledge. Second, acceleration does not necessarily eliminate realization constraints: it can relocate bottlenecks, weaken historical complementarities, or foreclose alternative realization pathways.

4.2 Potential P(t)

4.2.1 Potential as Structured Possibility

Potential is the most upstream object in the realization architecture, but it is not a single measure of intelligence. P(t) denotes the structured capability or possibility space available at time t: what AI systems can in principle make possible under the relevant technical conditions. Systems can improve unevenly across reasoning, coding, science, perception, planning, tool use, communication, autonomy, and other capability families.

Unlike Availability and Exposure, whose bounded component structures are inherited from prior work, Potential is intentionally not assigned a canonical tuple here. Imposing one would create a new capability taxonomy the present architecture neither requires nor yet justifies. The asymmetry is therefore deliberate rather than unfinished formalization. This is a scope decision, not a claim that Potential is intrinsically unstructurable: a more explicit capability structure would require a sufficiently stable cross-domain taxonomy that is not established here.

P(t) ≠ R_g(t)

The September 2026 Navier–Stokes announcement makes this separation unusually concrete. The generation and formal checking of a proposed Millennium-problem solution can count as evidence that the mathematical capability frontier has moved even while human understanding, community acceptance, publication, and institutional recognition remain incomplete. Recognizing a rise in Potential therefore does not require pretending that every downstream realization object has already been achieved (OpenAI 2026b; Clay Mathematics Institute 2026, 2018).

4.2.2 Potential and the Other Realization Arguments

Potential is not Availability: an extraordinary capability can remain rare, expensive, or restricted. Potential is not Exposure: a capability can be mature and available while reaching different populations very unevenly. Nor should the surrounding realization environment be folded back into P(t). Energy, compute, infrastructure, law, institutions, procurement, safety requirements, and trust may determine whether a capability becomes consequential; their importance does not make them capability.

Rule: A condition that determines whether a capability can be realized is not thereby part of the capability itself.

Compute illustrates why causal role matters. Training or research compute can condition the future growth of P(t); inference or deployment compute can condition Availability; infrastructure requirements involving compute can operate within Z_d(t). The same physical resource can therefore participate in different parts of the architecture without becoming identical to any of them.

4.2.3 Expansion, Time, and Changing Significance

An increase in Potential is not necessarily useful in every domain. It can solve one scarcity while shifting the binding constraint toward verification, interpretation, judgment, infrastructure, or governance. Potential is also time-dependent: the same capability can matter far more later if price, access, regulation, interfaces, workflows, or institutional readiness change around it.

P(t₁) = P(t₂)  does not imply  R_g(t₁) = R_g(t₂)

4.2.4 Potential and Recursion

Potential and realization are analytically separate but dynamically connected. Some realized AI applications can contribute to future capability through AI-assisted coding, automated experimentation, scientific discovery, evaluation, or research agents.

P(t) → R_g(t) → P(t+1)

This is not an automatic self-improvement law. The OpenAI research-acceleration case illustrates both sides: AI can be realized as AI research activity that may contribute to future capability, while compute, human judgment, safety constraints, and organizational choices continue to shape the trajectory. Potential supplies possibility, not destiny.

4.3 Availability Av(t)

4.3.1 The Availability Tuple

Av(t) = ⟨Ubiquity, Accessibility, Affordability⟩

Ubiquity concerns how broadly a capability or service is present; Accessibility whether relevant actors can obtain or invoke it; Affordability whether practical use is economically feasible for the population, market, organization, or deployment scope specified by the analysis. These dimensions are separable but not assumed causally independent: wider distribution can reduce cost, affordability can expand distribution, and interface or infrastructure changes can jointly affect Accessibility and Ubiquity. They are not additive and need not move together.

Av(t) ≠ U + A + F

4.3.2 Availability, Deployment, and Exposure

High Availability does not imply adoption. Users can distrust a system; regulators can restrict it; institutions can lack workflows for incorporating it; procurement can move slowly. Nor is Availability the same as Exposure. A system can be cheap and ubiquitous while one occupation uses it constantly and another scarcely encounters it; conversely, a proprietary system used by a small number of powerful institutions can indirectly expose a large population.

Av(t) ↛ R_g(t)

4.3.3 Availability and the Resource Stack

Availability must not become a container for every resource needed to build or operate AI. Energy, compute, semiconductors, networks, data centers, supply chains, labor, and institutions can all matter. The Availability construct asks a narrower question: how ubiquitous, accessible, and affordable is the relevant capability? A resource belongs in Availability only insofar as its role can be represented without material loss through those dimensions. The Availability/resource-stack boundary is therefore narrowed, not closed.

4.3.4 Abundance and New Scarcities

Greater Availability can relocate scarcity. If code generation becomes abundant, review and integration may become scarce; if experiments become abundant, judgment and evaluation may become more important; if proofs become abundant, verification and canonicalization can become limiting; if content becomes abundant, attention, trust, and provenance can become scarcer. Abundance does not eliminate scarcity. It changes what is scarce.

4.3.5 Time and Feedback

Availability can rise through falling prices, better interfaces, wider distribution, easier APIs, and improved hardware, or contract through withdrawal, restriction, price increases, infrastructure failure, or security controls. Through the realized changes it helps produce, Availability can influence later Exposure, domain conditions, and even future Potential. This feedback is mediated through realized changes and does not require a new direct edge from Availability to every later condition.

4.4 Exposure Ex_g(t)

Ex_g(t) = ⟨Reach_g(t), Susceptibility_g(t), Recourse_g(t)⟩

Availability tells us whether a capability is practically obtainable. It does not tell us who is actually reached, how consequential that contact is, or what ability affected groups possess to respond. Exposure is therefore group-specific and structured rather than scalar.

4.4.1 Reach

Reach asks whether and how an AI capability, deployment, decision, or consequence extends to a group. Direct use is only one form. A person can be reached because an employer uses AI in hiring, an insurer uses it in evaluation, a platform uses it in ranking, or a government uses it in administration. Narrow direct Availability can therefore coexist with extensive indirect Reach. Reach concerns the existence and extent of contact or mediated reach; weak Reach means limited contact, not low consequence conditional on contact.

4.4.2 Susceptibility

Equal Reach does not imply equal consequence. Susceptibility concerns the conditional significance of that contact once Reach exists: the degree and manner in which a group is positioned to experience consequences. It can concern benefit as well as burden; complementary skills and institutional support can make one group especially positioned to benefit while another faces displacement, dependency, error, or loss of bargaining power.

4.4.3 Recourse

Recourse concerns the capacity to recognize, contest, redirect, correct, exit, or obtain remedy for an AI-mediated decision or consequence. A nominal appeal channel is not necessarily effective Recourse. People may need to know AI was involved, identify the responsible actor, possess standing to challenge the decision, reach an institution with authority to act, and obtain a consequential response. Recourse is represented inside Exposure when viewed as a property of the affected group’s position. The institutional arrangements that create or remove Recourse may simultaneously appear in Z_d(t) or Governance when analyzed as features of the realization environment.

4.4.4 Non-Scalar and Group-Specific Exposure

Ex_g(t) ≠ Reach_g(t) + Susceptibility_g(t) + Recourse_g(t)

Reach can rise while Susceptibility falls; Recourse can improve without changing Reach. Two groups can encounter the same system with very different configurations. Exposure is not a welfare score; it identifies a relation between technological realization and particular groups.

4.4.5 Exposure, Consequence, and Observation

Ex_g(t) ↛ R_g(t)

A worker can encounter an AI system without being displaced; a patient can be evaluated by AI without receiving different treatment; a community can be exposed to an infrastructure proposal without the project ultimately being built. What analysts observe about Exposure is also not identical to Exposure itself. Public records can miss informal exposure, complaint counts can reflect awareness or reporting mechanisms as much as burden, and visible controversy does not establish representativeness. The empirical measurement problem therefore belongs to the Observation/Evidence layer. Conversely, a group-indexed realized change ordinarily presupposes some direct or indirect exposure pathway, but that pathway need not involve voluntary use or direct interaction.

4.4.6 Exposure Over Time

Exposure can change without capability changing as institutions expand deployment, regulators create remedies, professional groups learn to avoid or complement systems, and work structures change. Realized outcomes can in turn reshape later Exposure. This recursion is mediated through realized change and does not require a new direct causal edge for every interaction.

4.5 Domain-Specific Realization Conditions Z_d(t)

Potential, Availability, and Exposure still do not exhaust realization. A capable, available system can reach a relevant group and nevertheless remain unrealized because a hospital lacks an approved workflow, a buyer cannot complete procurement, a regulator prohibits a use, a data-center project lacks grid capacity, or a laboratory does not consider its monitoring sufficiently reliable. These typed domain-specific realization conditions are represented by Z_d(t).

4.5.1 A Typed Structure, Not a Residual

Unlike Availability and Exposure, Z_d(t) is not assigned a canonical tuple. Clinical medicine, mathematics, education, military systems, public administration, infrastructure, enterprise software, and consumer products need not share one contextual decomposition. But Z_d(t) cannot become a miscellaneous residual. A proposed condition should be relevant to realization in the specified domain, distinguishable from Potential, Availability, and Exposure, and necessary to preserve information that would otherwise be materially lost.

4.5.2 Application Rule for Candidate Domain Conditions

1. What is the realization object? The condition must matter to a specified realized state rather than to “AI adoption” in the abstract.
2. Why is the condition domain-specific? Its relevance should arise from the institutional, material, legal, professional, or organizational setting.
3. Why does it not belong more naturally in P(t), Av(t), or Ex_g(t)? Classification should follow causal role rather than terminology or physical identity.
4. What evidence would show that the condition matters? The analysis should identify observable contrasts, documented decisions, thresholds, timing changes, or other evidence capable of supporting or challenging the proposed role.

This protocol keeps the open structure operationally disciplined without turning recurring condition families into a universal checklist. Parsimony is a further requirement: prefer a simpler representation when it preserves the relevant empirical distinctions. A candidate Z_d(t) condition should not be added merely because it can explain an observed outcome after the fact; where feasible, candidate conditions and competing mapping behaviors should be specified before outcome analysis. The burden of proof lies with the more complex representation to show that added structure improves prediction, mechanism discrimination, or explanatory resolution.

A proposed domain condition fails admission when it cannot be distinguished empirically from another argument, is supported only post hoc, generates no discriminating pathway or prediction, or adds no explanatory resolution relative to a simpler nested representation. In such cases it should be merged, omitted, or left unclaimed rather than retained inside Z_d(t). A domain representation is provisionally adequate when the included conditions pass these admission criteria, materially outperform or clarify a simpler representation, and remaining unexplained variation does not justify further structure under the same evidentiary and parsimony standards.

Table 4.2 — Domain Conditions and Mapping Behavior: Clinical Medical AI Example

Candidate domain state in Zd(t)	Why it belongs in Zd(t)	Possible behavior through μt
Regulatory authorization	Legal status governing whether clinical use is permitted	Absolute deployment gate, conditional authorization, or restricted-use pathway
Hospital IT interoperability	Institutional/technical capacity to integrate the system into clinical workflow	Integration delay or rate constraint
Clinical safety validation	Domain-specific assurance state	Threshold or conditional gate
Clinician trust	Professional acceptance relevant to actual use	Soft adoption threshold or gradual rate effect
Procurement approval	Organizational authorization to acquire or deploy	Discrete institutional decision or delay
Liability allocation	Legal/institutional assignment of responsibility	May inhibit, redirect, or condition deployment

Note. The table illustrates how a domain can contain several relevant Z_d(t) states while μ_t specifies how each becomes consequential. The entries are neither exhaustive components of Z_d(t) nor universal medical-AI requirements. It is an illustration of the application method, not a required medical-AI checklist.

The same clinical example can be read across the full architecture: P(t) might denote a validated diagnostic capability; Av(t) its practical obtainability at feasible cost and infrastructure; Ex_g(t) the Reach, Susceptibility, and Recourse of clinicians or patients; Z_d(t) authorization, interoperability, liability, and workflow conditions; μ_t the posited combination of approval gates, integration delays, and adoption responses; and R_g(t) the actual incorporation of the system into clinical decision practice. Each element would require its own evidence rather than being inferred from deployment alone.

4.5.3 Boundaries with Potential, Availability, and Exposure

Institutional approval is not machine capability; procurement policy is not reasoning performance; legal permission is not intelligence. Likewise, not everything that affects practical use belongs in Availability. The correct classification depends on causal role. Exposure remains different again: it describes a group’s relation to AI, while Z_d(t) describes additional properties of the environment through which realization occurs.

4.5.4 Conditions Are Not Gates

Key distinction: Z_d(t) describes relevant states of the domain; μ_t describes how those states participate in realization.

A safety evaluation is a condition; a rule that prohibits deployment below a safety level is a gate. Trust may be a condition; whether adoption accelerates above a trust threshold is a property of the mapping. Law may be a condition; whether absence of authorization constitutes a veto or delay is a property of μ_t. Not every condition is therefore a gate.

4.5.5 Dynamics, Governance, and Participation

Domain conditions can inhibit or enable realization and evolve over time. Governance can alter them through law, standards, institutional design, and resource allocation. Participation can sometimes alter them through evidence-gated pathways such as consultation, protest, professional deliberation, or collective action. Neither Governance nor Participation should therefore be inserted into Z_d(t) as a scalar coefficient.

4.5.6 Why Z_d(t) Remains Open

It would be easy to impose a fixed vector of domain conditions and create the appearance of completeness. The Unified Model does not do so. A fixed vector would imply evidence that the same dimensions are necessary or sufficient across domains. That evidence does not presently exist. The open scope of Z_d(t) is therefore an explicit architectural commitment rather than unfinished notation.

4.6 The Realization Mapping μ_t: Gates, Rates, and Heterogeneity

The four arguments tell us what conditions are present. They do not tell us what those conditions do. μ_t is not another input; it is the structured relation through which heterogeneous inputs become consequential.

More precisely, μ_t denotes a family of admissible realization relations rather than one prespecified function. Different domains may instantiate different functional forms—or relations better represented through discrete rules, institutional decisions, threshold structures, delays, or qualitative mechanisms. The architecture therefore does not assume that μ_t is uniquely determined, continuously differentiable, parametrically stable, or recoverable from observational data alone.

4.6.1 The Mapping Is Not Addition

R_g(t) ≠ P(t) + Av(t) + Ex_g(t) + Z_d(t)

The arguments have different types and no general common unit. Extreme capability may not overcome legal prohibition; wide Availability may not overcome institutional refusal; favorable institutional conditions cannot realize a capability that does not exist. μ_t must therefore allow non-substitutability and interaction.

4.6.2 Gates, Thresholds, Rates, and Delays

A gate is not a new variable; it is one way μ_t can make an existing condition consequential. For this chapter, a gate is a mapping relation under which failure to satisfy a specified condition prevents realization along the specified pathway. A threshold specifies a crossing value; a delay alters timing; a rate effect alters speed or intensity without necessarily blocking realization. Regulatory approval, safety confidence, or institutional authorization can therefore behave as gates in some settings, while other conditions may alter rates, produce delays, redirect activity, interact with other conditions, or prove empirically weak. For example: regulatory authorization can operate as a gate; adoption can accelerate after a trust threshold is crossed; improved Availability can alter the rate of adoption without determining whether adoption occurs; and procurement or certification can impose a delay even after other conditions are satisfied.

4.6.3 Heterogeneity Across Actors, Groups, and Time

Different actors can respond differently to the same conditions. One organization adopts at a level of evidence another finds insufficient; one jurisdiction permits a use another prohibits; one profession accepts a system another resists. The mapping itself can change as institutions learn, social proof accumulates, legal uncertainty is resolved, or safety techniques become inadequate or improve.

μ_t1 ≠ μ_t2

Local scalarity does not imply global scalarity. A narrow realization problem may admit a meaningful scalar threshold, coefficient, rate, or response curve without implying that the heterogeneous architecture as a whole is reducible to one scalar realization law.

4.6.4 Bottleneck Relocation versus Path Foreclosure

Bottleneck relocation occurs when acceleration relaxes one constraint and makes another relatively more important while leaving the relevant opportunity set substantially intact.

Path foreclosure occurs when realization along one trajectory changes the opportunity set itself, reducing or eliminating realizations that could have emerged from an alternative route.

OpenAI’s research-acceleration case primarily illustrates the first possibility: faster coding or experimentation can make compute, evaluation, human judgment, or safety relatively more constraining. Tao’s mathematics examples raise the stronger second possibility: obtaining an answer early can reduce the opportunity to generate techniques, insights, questions, and learning that would have emerged through unsuccessful exploration. The distinction is analytical rather than evaluative; whether preservation or foreclosure of an alternative path is desirable belongs partly to Optimization. Prospectively, a bottleneck-relocation claim predicts that alternative realization opportunities remain available while another constraint becomes binding. A foreclosure claim is stronger: it requires evidence that taking one realization path changes the availability or probability of outcomes that would otherwise have remained reachable through another path. If the alternative pathway remains materially available after the first path is taken, a foreclosure claim is not supported; the evidence is more consistent with bottleneck relocation, substitution, or another non-foreclosure mechanism.

A field-level version of bottleneck relocation was articulated just before the September episode. Alonso Castillo-Ramírez argues that AI may end handcrafted proof production as the dominant mode of frontier mathematical research, not by ending mathematics but by relocating scarcity. If proofs become abundant, the harder constraints shift toward verification, understanding, judgment, mathematical taste, exposition, question selection, and the organization of rapidly expanding results. In Chapter 4 terms, this is primarily a relocation claim: the division of mathematical labor changes while the broader opportunity set need not contract (Castillo-Ramírez 2026).

Weinreich advances the stronger objection. In The crisis of AI-generated mathematics, he argues that automated proof production can decouple proof output from human understanding, overwhelm reviewing and attribution systems, distort professional incentives, and redirect which questions mathematicians choose to pursue (Weinreich 2026). In Chapter 4 terms, this is a path-foreclosure hypothesis at field scale: acceleration may alter the opportunity set for human learning, method formation, and research autonomy. The present episode does not establish that such foreclosure has occurred. It establishes the concern, the proposed mechanism, and the need to distinguish it empirically from ordinary bottleneck relocation.

4.6.5 CFS as a Bounded Realization-Dynamics Family

Ceiling–Floor–Slope is the principal formal realization-dynamics family inherited by this chapter. It represents settings in which advancing capability interacts with slower-moving or heterogeneous realization floors, thresholds, lags, and synchronization regimes. But CFS is one possible instantiation of realization dynamics inside μ_t, not the definition of μ_t itself. Some realization processes may be dominated by legal vetoes, discrete institutional decisions, bargaining, coordination, or path dependence. CFS remains valuable because it is bounded and falsifiable. Analysts should use CFS when a domain supplies interpretable ceiling and floor constructs, longitudinal variation, a meaningful gap between them, and a testable hypothesis about how relative rates, thresholds, or delays govern realization. Where those conditions cannot be operationalized, the realization problem should not be forced into CFS.

4.6.6 Contemporary Illustrations Are Not Validation

Pachocki illustrates capability versus safety confidence; OpenAI research acceleration illustrates accelerated stages versus shifting bottlenecks; a16z illustrates usable products versus institutional adoption; Tao illustrates answers versus understanding and discovery; and Castillo-Ramírez illustrates the possibility that abundant proof production relocates scarcity toward understanding, judgment, and exposition. The September 2026 Navier–Stokes episode adds a particularly compact live example: frontier mathematical capability can advance sharply while candidate proof, formal verification, human understanding, community acceptance, and institutional recognition remain distinct realization objects. Weinreich adds a stronger field-scale resistance and foreclosure hypothesis. None of these thereby instantiate the same recurrence. They illustrate the architecture; they do not validate one submodel.

4.6.7 Governance, Participation, and Optimization

Governance can alter Z_d(t) and μ_t through law, standards, institutional design, resource allocation, and authoritative decisions. Participation can enter realization pathways, but the Participation work requires a stronger discipline:

Act ↛ Consequence ↛ RealizationChange

Participation is therefore not a universal coefficient. Optimization occupies a different position again. μ_t asks how realization occurs under given conditions; Optimization asks which trajectories should be preferred under particular objectives and tradeoffs. The second cannot be answered merely by accelerating the first.

4.6.8 Identification, Non-Monotonicity, and Scope

Writing μ_t does not mean it has been empirically identified. Multiple conditions may change together, counterfactuals may be absent, institutional reasons may remain undocumented, and apparent thresholds may be artifacts of sparse data. The empirical discipline should therefore distinguish pathway observed from counterfactual identified. The mapping is also not assumed monotonic: more Potential, Availability, or Exposure need not produce more of every realized outcome. A domain-specific μ_t claim should therefore state which observations would distinguish it from plausible competing relations and which evidence would merely document that realization occurred.

4.7 Realized State, Consequence, and Recursion

4.7.1 Realization Requires an Object

A claim that “AI has been realized” is incomplete. Realized as what—a product, a workplace practice, a verified theorem, an institutional rule, a scientific capability? A proof and mathematical understanding are different realization objects. Illustratively, R^proof(t) may increase while R^{understanding}(t) lags. This notation is explanatory only and is not an addition to the frozen architecture. A realization object should be specified before interpreting the outcome and bounded tightly enough that its change can be assessed independently; defining the object after observing the result would make the claim too easy to fit.

That distinction became unusually concrete in September 2026. OpenAI reported that an internal model significantly more capable than GPT-6 Astra produced a proposed resolution of the Navier–Stokes Millennium Prize Problem using on the order of 10,000 concurrent agents. OpenAI reports roughly 88 hours from launch to the candidate solution and another 17 hours of Lean formalization and verification by Astra, and says the proof establishes statements C and D of the official Clay formulation. Yet as of September 8, 2026, the Clay Mathematics Institute still listed Navier–Stokes among the unsolved Millennium problems. Its rules require qualifying publication, a waiting period of at least two years, and general acceptance by the global mathematics community before CMI considers a proposed solution (OpenAI 2026b; Clay Mathematics Institute 2026, 2018).

The analytical point does not depend on adjudicating the separate disputes that emerged around the episode. A frontier capability event can immediately divide into different realization objects: a candidate proof, a formally checked proof, a humanly understood proof, a publishable exposition, a community-accepted theorem, and an institutionally recognized resolution. Questions of attribution, priority, or provenance may remain unsettled as well, but they are secondary to the present chapter’s core distinction between capability and realized mathematical knowledge.

4.7.2 Consequence, Scale, Partiality, and Reversibility

A consequence need not imply a broader realization change. A recommendation may not be adopted; a pilot can produce local gains without scaling; a protest can delay a decision without changing the final outcome. Realization can be local without being systemic, partial rather than binary, and reversible through withdrawal, policy change, loss of trust, or organizational reversal.

4.7.3 Realized Outcomes and Feedback

R_g(t) → {Av(t+1), Ex_g(t+1), Z_d(t+1)}

The schematic is illustrative, not a new inventory of direct causal edges. Successful deployments can expand Availability; failures can reduce trust; regulation can improve Recourse; organizations can reorganize; markets can reprice skills. The realized state becomes part of the history inherited by future realization.

4.7.4 Recursion into Future Potential

P(t) → R_g(t) → P(t+1)

In some domains realized AI activity can contribute to future capability, with AI-assisted research as the clearest contemporary example. This does not imply runaway self-improvement: the loop remains conditioned by compute, human judgment, institutional decisions, safety, and other realization conditions. The relevant time scale is empirical and domain-specific: research cycles, procurement cycles, regulatory cycles, and institutional learning can operate at very different rates. Chapter 4 does not posit one universal temporal law; Synchronization/CFS provides one formal apparatus where differential rates and delays can be operationalized. Observation frequency should therefore be chosen relative to the expected time scale of the process: sampling too coarsely can conceal thresholds or delays, while sampling too frequently can mistake transient variation for structural change. Where relevant variables operate on systematically different time scales, the empirical specification should state those scales explicitly.

4.7.5 Bottlenecks, Complementarity, and Scarcity

Realization can change scarcity. Faster experimentation can expose judgment as scarce; more content can expose attention as scarce; more proof generation can expose verification and canonicalization as scarce. Where outputs historically arrived together—solution, technique, training, explanation, further questions—selectively accelerating one can also weaken their complementarity.

4.7.6 Tao, Path Dependence, and Foreclosure

Tao’s September 2026 discussions make this distinction concrete. In Answers versus insight, he separates obtaining the answer to a mathematical problem from learning what the problem can teach. His September 3 Navier–Stokes worked example then gives that concern a specific form: much of the mathematical value of a hard problem lies in successive attempts that fail, because each obstruction can motivate a new construction, method, or insight. This learning depends on the investigator not already possessing the successful answer; foreknowledge can contaminate the instructive dead ends (Tao 2026b, 2026c).

Chapter 4 inference: Faster realization of one object can reduce the opportunity to realize other objects that would have emerged from the slower path.

This is path dependence rather than mere bottleneck relocation. The route through which realization occurs can influence what additional knowledge, techniques, skills, or opportunities remain realizable afterward. Foreclosure is not automatically undesirable; the analytical point is that realization changes the future opportunity set, while Optimization evaluates whether that change is desirable.

4.7.7 Optimization and Participation in Recursive Realization

If the objective is simply to maximize solved problems, rapid autonomous proof production can look highly successful. If the objective includes understanding, technique formation, training, and cumulative human knowledge, the same trajectory may be less desirable. Participation can likewise redirect recursive trajectories through worker action, professional deliberation, public opposition, user feedback, or community contestation, but Act still does not entail Consequence or RealizationChange.

The wider mathematical response shows that even once proof production becomes technically possible, the desired realization trajectory remains plural. Castillo-Ramírez anticipates a transformed division of labor in which proofs become abundant while mathematicians concentrate more heavily on reviewing, discerning, digesting, explaining, organizing results, and choosing which questions matter (Castillo-Ramírez 2026). Weinreich argues instead for total opposition to AI use in mathematics and calls for coordinated action by individuals, departments, journals, and institutions (Weinreich 2026). The IMU-endorsed Leiden Declaration advocates governed accommodation: disclosure, human responsibility, attribution, peer review, appropriate formal verification, and community control, while allowing that preservation of mathematical values may justify delay in obtaining results (Leiden Declaration 2026). Tao’s position is different again: AI forces mathematics to state more clearly which goals—answers, understanding, technique, training, or other values—it is trying to optimize. These positions concern what kind of mathematical realization should be pursued, not whether the Navier–Stokes capability event occurred.

4.7.8 Observation, Feedback, and Ongoing Transformation

Realized states remain distinct from what analysts can observe. Evidence can be delayed, selective, incomplete, or systematically biased toward visible events. Feedback can amplify, stabilize, constrain, or redirect; a feedback loop is not an inevitability engine. As Chapter 3 established, Gradual AGI is an Ongoing, Unfinished Transformation: what has already been realized alters what can be realized next. Section 4.9.5 makes this distinction operational by separating an underlying realized state from the evidence-conditioned estimate available to an analyst.

4.8 What the Realization Mapping Does Not Claim

The equation identifies a structure of inquiry; it does not predetermine the empirical answer. The architecture makes the following non-claims explicit.

4.8.1 No Universal Scalar

Local measurement may be useful, but the framework does not assume a universal number for Potential, Availability, Exposure, domain conditions, or Realization.

4.8.2 No Additivity

The arguments have different types and need not be substitutable; the realization equation is not a sum.

4.8.3 No Guaranteed Monotonicity

More capability does not necessarily mean more realization; faster output can even foreclose another valued path.

4.8.4 No Guarantee That Capability Becomes Realized

An increase in P(t) creates possibility but does not entail a corresponding increase in R_g(t).

4.8.5 No Universal Set of Gates

Gates are possible behaviors of μ_t, not universal model components.

4.8.6 No Universal CFS Recurrence

CFS remains a bounded, falsifiable submodel. Its value depends on preserving that boundary.

4.8.7 No Exhaustive Z_d(t)

Recurring condition families do not form a universal vector.

4.8.8 No Automatic Causal Identification

Formal representation is not empirical identification; pathway observed is not counterfactual identified.

4.8.9 No Claim That Realization Equals Desirability

The mapping explains realization; Optimization evaluates possible trajectories.

4.8.10 No Universal Realization Path

The architecture does not posit P→Av→Ex→Z→R as a fixed sequence. The objects are distinct and their relations dynamic and recursive.

4.9 How the Unified Model’s Frameworks and Cross-Cutting Structures Enter

The realization architecture does not stand apart from the Unified Model, but neither should every prior framework become another argument of μ_t. The relevant structures have different logical roles.

4.9.1 Synchronization and CFS

Synchronization describes the possibility that machine and human/institutional clocks differ. CFS formalizes one family of those dynamics. CFS is one possible formalization of realization dynamics, not the realization mapping itself.

4.9.2 Optimization

Optimization asks which trajectories should be preferred. Its objectives are plural and need not reduce to one global scalar, so it remains upstream of the realization mapping rather than entering as an argument.

4.9.3 Governance

Governance can alter both Z_d(t) and μ_t. It changes not only conditions but, in some settings, the rules by which conditions become consequential—for example by converting a discretionary practice into a legal gate, changing appeal rights, or redefining authorization thresholds. In that sense the mapping can be partly endogenous to governance decisions. Governance is therefore cross-cutting with respect to realization rather than a terminal stage after it.

4.9.4 Participation

Participation contributes a Subject→Act→Consequence architecture while preserving the empirical discipline Act ↛ Consequence ↛ RealizationChange. It enters through typed, evidence-gated pathways rather than as a universal coefficient.

4.9.5 Observation/Evidence

Observation/Evidence is a cross-cutting epistemic layer. Its function is not to cause realization; it constrains what can responsibly be inferred about P, Av, Ex, Z, μ, and R. The underlying world and the evidence available about it are not identical.

R̂_g(t) ≠ R_g(t)

Here R̂_g(t) denotes an analyst’s evidence-conditioned estimate of an underlying realized state R_g(t). The notation is pedagogical shorthand for the observation problem, not a new state variable used elsewhere in the frozen architecture. Differences can arise through incomplete observation, measurement error, selection effects, reporting asymmetries, missing counterfactuals, or disagreement about the realization object itself.

Table 4.3 — Evidence Status for Realization Claims

Claim	Evidence that may support it	What remains unidentified
Deployment occurred	Product, contractual, or institutional records	Actual intensity or effective use
Group was reached	Administrative, survey, usage, or process evidence	Full susceptibility or consequence
Participatory pathway occurred	Dated act → response → decision evidence	Counterfactual effect absent the act
Realized state changed	Outcome or institutional-state evidence	Mechanism unless separately established
Mechanism identified	Design, natural experiment, matched comparison, credible counterfactual, or equivalent evidence	Generalizability beyond the studied setting

The Observation/Evidence layer therefore does more than warn that measurement is imperfect. It disciplines the epistemic status assigned to different realization claims.

4.9.6 Feedback and Recursion

Recursion is not another imported framework. It is a structural property of realization: R_g(t) can reshape conditions governing later realization. Feedback is mediated through realized change and does not require a new direct edge between every pair of variables.

Figure 4.3. Integration without reduction. Optimization, Governance, and Participation intersect the realization architecture through different logical roles; Observation/Evidence is epistemic; Feedback/Recursion is structural. CFS formalizes some μ_t dynamics but is not μ_t itself. The figure is explanatory and does not supersede the frozen diagram edge inventory.

4.9.7 Integration Without Reduction

Synchronization is not Governance; Governance is not Participation; Participation is not Exposure; Exposure is not Observation; CFS is not μ_t; Recursion is not another argument. The Unified Model gains explanatory power by preserving these distinctions while making their interfaces explicit. Its unity lies in relationships among distinct structures rather than one master scalar or one universal recurrence.

4.10 Falsifiers, Open Questions, and Handoff

The realization architecture is useful only if it can be wrong. It therefore ends not with another mechanism but with its failure conditions and unresolved questions.

4.10.1 Falsifiers

• Capability sufficiency: The architecture would be weakened if realized states across sufficiently diverse domains could be predicted from Potential alone, with Availability, Exposure, domain conditions, and mapping variation adding little explanatory value.
• Scalar or additive reduction: If a stable scalar or additive representation repeatedly preserved the relevant distinctions and predicted heterogeneous realization with little information loss, the present structured non-scalar treatment would be unnecessarily complex.
• Parsimony failure: If a simpler representation that merges or omits one of the architecture’s distinctions performs equivalently on the specified empirical task, the additional distinction has not earned its place in that application.
• Mapping complexity or unconstrained elasticity: In a specified domain, if a simpler nested mapping—including a stable or constant relation—organizes and predicts realization as well as the richer time-varying account, the additional mapping complexity is not warranted. Conversely, if competing μ_t mechanisms cannot be distinguished by any feasible observation, the mapping claim is too unconstrained for that application.
• CFS failure: A realization gap alone does not confirm CFS. The relevant question is whether the process actually displays the Ceiling–Floor–Slope structure being claimed.
• Exposure redundancy: If Reach, Susceptibility, and Recourse added no explanatory value across the domains where they are invoked, first-class Exposure would be unnecessary.
• Weak recursion: A recursion claim fails in a specified domain when repeated opportunities for realized change show no detectable alteration in the pre-specified later conditions, beyond the study’s measurement resolution and baseline expectations. In that application the recursive claim should be narrowed or omitted.
• Path-independence: If alternative routes to the same endpoint systematically produced equivalent downstream knowledge, institutions, skills, and opportunity sets, the path-dependence claim would be weak.

4.10.2 Open Questions

• Scope of Z_d(t): Its conceptual role is clear; its exhaustive internal scope is not. Additional structure is sufficiently justified only when it passes the admission rules, survives comparison with a simpler representation, and supports observations capable of discriminating its proposed role.
• Availability/resource-stack boundary: The current status remains narrowed, not closed. Classification follows causal role.
• Identification of μ_t: Which gates actually bind? Which thresholds are causal? Which delays are structural? Which mechanisms survive counterfactual analysis? Future empirical applications should pre-specify, where feasible, which observations would discriminate among competing μ_t mechanisms rather than merely documenting that realization occurred.
• Acceleration and the object of value: When does acceleration merely relocate a bottleneck, when does it weaken a complementarity, and when does it foreclose another realization pathway?
• Observation and underlying state: High visibility is not prevalence; low complaint counts are not necessarily low burden; observed Participation is not causal influence; observed adoption is not necessarily effective realization.

4.10.3 Empirical Application Protocol

1.  Specify the domain and realization object before interpreting the outcome.

2.  Operationalize the relevant Potential, Availability, Exposure, and domain-condition states with pre-declared indicators where feasible.

3.  Specify a small set of competing candidate realization mechanisms rather than one post-hoc μt story.

4.  Identify observations capable of discriminating among those mechanisms.

5.  Distinguish descriptive pathway evidence from causal or counterfactual identification.

6.  Compare the richer architecture with a simpler baseline or nested representation—for example, one that absorbs a candidate domain condition into Availability, assumes a common or time-stable realization relation, omits a proposed interaction, or otherwise uses fewer analytically distinct terms.

7.  Retain only the complexity that improves prediction, mechanism discrimination, or explanatory resolution.

No single study must identify every element. The requirement is narrower: empirical applications should make clear which parts of the architecture are measured, which are assumed, which remain untested, and what evidence could force simplification or revision.

4.10.4 The Chapter’s Main Claim

The chapter’s distinctive claim is not merely that capability can fail to determine social consequence. It is that realization claims should separate Potential, Availability, group-specific Exposure, domain-specific conditions, and the realization relation—and should retain those distinctions only when they produce empirical or explanatory payoff. AI capability does not become socially consequential through a single automatic transmission; realization depends on heterogeneous, changing, and partly governable relations among these analytically distinct objects.

Potential remains upstream, but upstream is not sufficient. A theory that examines only P(t) can describe an expanding technological frontier; it cannot by itself describe AGI-inclusive humanity. The question therefore changes from “How capable has AI become?” to “What has that capability actually become in the world, for whom, under what conditions, through which pathways, and with what consequences for what becomes possible next?”

4.10.5 Handoff

Chapter 4 has separated Potential from Availability, Availability from Exposure, Exposure from domain-specific conditions, and all four from the realized state they help produce. It has defined the role of μ_t without pretending that one mechanism governs every domain; placed CFS inside the broader architecture without universalizing it; shown how Optimization, Governance, and Participation intersect realization without becoming additional arguments; and made feedback, observation, uncertainty, and open scope part of the model rather than afterthoughts.

The later chapters inherit a more demanding question: How does each part of the Unified Model alter what AGI-inclusive humanity is able, willing, permitted, prepared, or ultimately able to realize?

Gradual AGI is not the name of a completed transition. As Chapter 3 established, it concerns an Ongoing, Unfinished Transformation. Chapter 4 adds the realization architecture through which that unfinished transformation can be studied.

Chapter References

Book and Unified Gradual AGI Sources

W.H.L., GPT-5.6 Sol, & Claude Sonnet 5. 2026. “A Multi-Plateau Framework of the Unified Gradual AGI Model.” Chapter 3 of On Gradual AGI. Champaign Magazine, September 6, 2026. Link

W.H.L. & ChatGPT. 2026. “Gradual AGI as Epistemic Extension.” Gradual AGI Series #1. Champaign Magazine, January 19, 2026. Link

W.H.L. & GPT-5.5. 2026. “Gradual AGI as Abundant Resources.” Gradual AGI Series #2. Champaign Magazine, June 11, 2026. Link

W.H.L. & GPT-5.5. 2026. “Gradual AGI as Synchronization for Transformative Adoption.” Gradual AGI Series #3. Champaign Magazine, June 29, 2026. Link

W.H.L. & Claude. 2026. “Ceiling, Floor, and Slope: A Falsifiable Dynamical Model of Synchronization for Gradual AGI.” Gradual AGI Series #4. Champaign Magazine, July 4, 2026. Link

W.H.L., Claude Sonnet 5, & GPT-5.5. 2026. “Gradual AGI as Optimization: A Conceptual Framework.” Gradual AGI Series #5. Champaign Magazine, July 22, 2026. Link

W.H.L. & Claude. 2026. “Gradual AGI as Optimization: Formal Models and Empirical Tests.” Gradual AGI Series #6. Champaign Magazine, July 24, 2026. Link

W.H.L. & Claude. 2026. “Gradual AGI as Contestation: A Framework for Governance.” Gradual AGI Series #7. Champaign Magazine, August 3, 2026. Link

W.H.L., Claude, & GPT-5.6 Sol. 2026. “Gradual AGI as Contestation: Measuring Governance Under Empirical Contact.” Gradual AGI Series #8. Champaign Magazine, August 11, 2026. Link

W.H.L., GPT-5.6 Sol, & Claude Sonnet 5. 2026. “Gradual AGI as Participation: Subjects, Acts, and Consequences.” Gradual AGI Series #9. Champaign Magazine, August 27, 2026. Link

W.H.L., GPT-5.6 Sol, & Claude Sonnet 5. 2026. “Gradual AGI as Participation: 3M Taxonomy, Relational Grammar, and Responsive Floor Test.” Gradual AGI Series #10. Champaign Magazine, August 30, 2026. Link

Contemporary External Sources

a16z. 2026. “The Two Ways to Sell AI: Lighthouse or Landgrab?” The a16z Show, August 13, 2026. Elena Burger with Andy McCall and Joe Schmidt. Link

Castillo-Ramírez, Alonso. 2026. “The End of an Era in Mathematical Research.” Proofs and Prompts, August 12, 2026. Link

Clay Mathematics Institute. 2018. “Rules for the Millennium Prize Problems.” Revised September 26, 2018. Link

Clay Mathematics Institute. 2026. “Navier–Stokes Equation.” Accessed September 8, 2026. Link

Leiden Declaration on Artificial Intelligence and Mathematics. 2026. June 2, 2026. DOI: 10.5281/zenodo.20302944. Link

OpenAI. 2026a. “Research acceleration: The view inside OpenAI.” September 6, 2026. Link

OpenAI. 2026b. “On the Navier–Stokes Millennium Prize Problem.” September 8, 2026. Link

Pachocki, Jakub. 2026. “An Alien Mind.” OpenAI, September 6, 2026. Link

Schmidt, Joe, & Julian Marx. 2026. “Lighthouse or Landgrab? How to Pick Your AI Sales Strategy.” Andreessen Horowitz, July 27, 2026. Link

Tao, Terence. 2026a. “Mathematics in the age of AI.” arXiv:2608.16753. Link

Tao, Terence. 2026b. “Answers versus insight.” Mathstodon, September 3, 2026. Link

Tao, Terence. 2026c. “Navier–Stokes as a worked example.” Mathstodon, September 3, 2026. Link

Weinreich, Max. 2026. “The crisis of AI-generated mathematics.” arXiv:2608.02859. Submitted August 3, 2026; revised August 24, 2026. Link