By W.H.L., GPT-5.6 Sol, Claude Sonnet 5
Chapter 12 of the forthcoming book On Gradual AGI
Publication Candidate v1.04 · September 27, 2026
Abstract
Chapter 12 consolidates the Unified Gradual AGI Model as a realization-centered architecture without introducing a new higher-order theory or master equation. At its core is the mapping Rg(t)=μt(P(t), Av(t), Exg(t), Zd(t)), relating capability or Potential, Availability, group-specific Exposure, and domain-specific realization conditions to a realized state. Graduality, Ceiling–Floor–Slope, Optimization, Governance, Participation, Observation and Evidence, and Feedback and Recursion are placed relative to that core while preserving the boundaries that keep them non-identical. The chapter applies the architecture to Go, mathematics, and radiology; shows how a reader can move from observed evidence to warranted inference; preserves unequal claim statuses rather than flattening them into one verdict; states the model’s principal nonclaims and unresolved burdens; and closes by distinguishing Ongoing, Unfinished Transformation as a book-level methodological orientation from its narrower, domain- and horizon-relative empirical diagnostic. Unification is architectural, not algebraic.
Keywords
Gradual AGI; Unified Gradual AGI Model; realization; Potential; Availability; Exposure; domain-specific realization conditions; Multi-Plateau Framework; Ceiling–Floor–Slope; Optimization; Governance; Participation; epistemic discipline; feedback; recursion; AGI-inclusive humanity
Reader Guide
Architecture path. §§12.1–12.3 move from the model’s three design principles to the reader-level architecture and the boundaries that keep it coherent. Figure 12.1 is the compact visual synthesis; the Multi-Plateau Framework remains the logical architecture for typing, revising, and relating claims rather than a causal process inside realization.
Application path. §§12.4–12.5 apply the architecture across Go, mathematics, and radiology and then convert the chapter into a reader workflow: identify the object, separate observation from representation and inference, use the appropriate subsystem only where its conditions are met, and stop when the evidence does not support the stronger claim.
Status and limits path. §§12.6–12.9 preserve canonical, high-confidence, provisional, untested, measurement-limited, non-operable, and NotAdjudicated states; state what the model does not claim; carry forward open burdens without promoting them for closure; and return to Ongoing, Unfinished Transformation without turning it into an AGI-arrival forecast or a universal empirical verdict. Readers who want only the practical claim-evaluation workflow may read §§12.5–12.6.
12.1 From Components to a Model
Three design principles
Before describing the model, we state the three principles on which its architecture is built.
Open. The Unified Gradual AGI Model is offered as an open intellectual architecture. We have worked to make it original, but we do not treat its ideas, terminology, or architecture as proprietary to us. Anyone with something to contribute is welcome to join the discussion—to question the model, test it, extend it, or use it. Openness also applies to the model’s own workings: definitions, propositions, statuses, revisions, and retired constructs are recorded so that a reader can see what changed, when, and why.
Flexible. We treat the transformation associated with increasingly general AI as ongoing and unfinished for purposes of theory-building, and we do not claim to know in advance how it will unfold. The architecture is therefore designed to be updated, revised, or repaired. Its components are kept distinct so that one can change without dissolving the rest, and every claim carries an explicit status. When a claim fails, repair follows a defined path: locate the failure, identify what depends on it, revise the affected part, carry the consequences forward, and retest. Repair is not rescue; a revision may not merely shield a claim from testing.
Falsifiable. The architecture is designed to meet evidence. Its propositions and claims are meant to be stated with what would count against them, and each carries a status recording how far it has actually been tested. Some claims cannot yet be tested; we say so. Adverse evidence may force contraction (narrowing a claim’s scope or strength), revision, or retirement. The point is not to make every claim binary, but to make visible which claims survive, which contract, which remain unresolved or untested, and which require revision or retirement.
The three principles work together. Openness makes revision visible, flexibility permits it, and falsifiability disciplines it. Together they aim to make the model something that can be wrong in public and corrected in the open. With these principles in view, we turn to the model itself.
The Unified Gradual AGI Model was not built by beginning with a single equation and assigning every later phenomenon a place inside it. It developed in the opposite direction. Different analytical problems required different distinctions: capability had to be separated from realization; realization conditions had to remain heterogeneous rather than collapse into one scalar; synchronization required a bounded dynamical treatment; trajectory choice had to be distinguished from authority; governance from participation; participation from consequence; observation from the state observed; and temporal return from the realization mapping itself. The resulting architecture is unified not because these distinctions disappear, but because they can be placed in relation to one another without losing the boundaries that made them analytically useful.
Unification is architectural, not algebraic. The model is organized around a realization-centered architecture whose highest-level system object is AGI-inclusive humanity: humans, AI systems, institutions, infrastructure, and the wider social environment treated as parts of one evolving system. Its central bridge is realization—the relation between technical capability or potential and socially realized outcomes. Chapter 4 formalizes that bridge as
Here ‘inclusive’ names the analytic scope of the system object—co-presence and interaction among humans, AI systems, institutions, infrastructure, and environment—not a normative verdict about which social arrangements ought to be preferred.
Rg(t) = μt(P(t), Av(t), Exg(t), Zd(t))
where P(t) denotes capability or potential, Av(t) Availability, Exg(t) group-specific Exposure, and Zd(t) domain-specific realization conditions. These arguments are heterogeneous. They are not additive terms in a universal AGI score, and the mapping does not imply that every relevant process in the book becomes another argument of μt.
In this chapter, unity means relational placement without conceptual collapse. Optimization remains upstream of realization rather than becoming a fifth input. Governance cuts across realization through contestation, standing, intervention, observation, and institutional steering. Participation enters through contingent pathways: a Participatory Consequence may or may not become a realization change, and the human-side Subject–Act–Consequence apparatus sits within a reciprocal account of human and AI participation. Observation and Evidence constrain what can be warranted about the underlying process. Feedback and Recursion describe how realized outcomes can alter later conditions. Ceiling–Floor–Slope (CFS) characterizes one bounded family of synchronization dynamics rather than the whole architecture.
The same discipline applies to the Multi-Plateau Framework (MPF) introduced in Chapter 3. The MPF is the methodological-theoretical architecture through which claims about the Unified Model are organized. Its Plateaus are logical vantage points, not chronological stages through which AGI develops. A Plateau provides a relatively stable position from which a kind of claim can be defined, examined, formalized, tested, revised, or related to other claims without implying temporal succession, hierarchy, or a terminal state. Typed relations, versioning, and anti-rescue discipline help preserve those distinctions when claims change status or require repair. The MPF therefore organizes the theory; it is not another mechanism operating inside realization. Chapter 3 deliberately leaves the substantive realization mapping to Chapter 4.
That separation matters especially in a capstone chapter. A synthesis can easily become a form of conceptual compression in which distinctions painstakingly established earlier are treated as interchangeable merely because they belong to one theory. Chapter 12 takes the opposite approach. Its task is to show that the model becomes more coherent—not less—when its non-identities, stop rules, unequal evidentiary statuses, and unresolved boundaries remain visible. The chapter presents a reader-facing synthesis in the main text, while the exhaustive canonical register of definitions, propositions, formal relations, statuses, boundaries, and open items is retained in the separate companion Scholarly Appendix (Canonical Extraction Register), distributed with the scholarly version of the chapter.
The chapter first presents the architecture in one view and states the boundaries that keep its components from collapsing into one another. It then applies the model across three domains—Go, mathematics, and radiology—and shows how a reader can use the architecture to ask what has actually been observed or represented, what inference that evidence warrants, and which boundary would have to be crossed to say more. It then separates what the model presently claims from what it does not claim and from what remains open. Section 12.9 returns to the book-level idea of an Ongoing, Unfinished Transformation without turning that orientation into a forecast or a universal empirical verdict.
The result is a structured account of how increasingly general artificial intelligence can become consequential through heterogeneous realization conditions, actor choices, institutional contestation, participation, and historically conditioned change, together with the limits of what can be observed and warranted about that process. The purpose of the present chapter is to state that architecture as a whole without turning synthesis into theoretical expansion.
12.2 The Architecture in One View
With the design principles now stated in §12.1, this section presents the Unified Gradual AGI Model in the most compact reader-facing form available in the book. It does so in three complementary ways. First, an integrated architecture diagram gives the model as a whole at a glance. Second, a component table shows what each major part of the architecture does and where it sits relative to the realization mapping. Third, a set of one-line definitions provides a concise conceptual handle on the model’s key ideas. The aim is not to replace the fuller arguments developed in earlier chapters, but to make the architecture intelligible even to a reader encountering it here for the first time.
12.2.1 Integrated architecture
Figure 12.1 is the chapter’s reader-level synthesis. It is centered on realization rather than on chronology: capability, availability, exposure, and domain-specific realization conditions meet in the Chapter 4 mapping, while Optimization, Governance, Participation, Observation and Evidence, Feedback and Recursion, Graduality, and Ceiling–Floor–Slope occupy distinct positions around that center. The diagram also retains a compact MPF orientation band and the three design principles for reader navigation; it does not replace Figure 3.1, which remains the full representation of the Multi-Plateau Framework.
A first-time reader can read the figure from the center outward: begin with the Realization Mapping; then locate Optimization upstream, Governance across the mapping, CFS as the nested bounded submodel, Observation and Evidence as the epistemic boundary, Participation as a contingent pathway, and Feedback/Recursion as the temporal return.
Figure 12.1. Integrated Architecture of the Unified Gradual AGI Model. The figure presents the model as a realization-centered architecture. At its core is the mapping Rg(t)=μt(P(t), Av(t), Exg(t), Zd(t)), relating capability, availability, group-specific exposure, and domain-specific realization conditions to a realized state for group g. Optimization is upstream; Governance cross-cuts realization; Participation enters through contingent pathways rather than as a scalar input; CFS is a nested synchronization submodel; Observation and Evidence define the epistemic boundary; and Feedback/Recursion provides a temporal return path. The MPF band is an orientation aid only; Figure 3.1 remains the full methodological-theoretical architecture. The Typed relations and Cross-plateau operations panel reproduces Chapter 3 §3.5’s world/model relation types and operation-based cross-plateau mappings; it introduces no new Chapter 12 taxonomy. Within the CFS panel, ‘Potential ceiling’ is reader-facing shorthand for a domain-scoped capability ceiling; the displayed non-identity Ct ≠ P(t) governs, so the Ceiling is not identical to the realization mapping’s Potential term.
At the center of the substantive architecture, Chapter 4 formalizes realization as
Rg(t) = μt(P(t), Av(t), Exg(t), Zd(t))
Here P(t) denotes capability or Potential; Av(t) Availability; Exg(t) group-specific Exposure; Zd(t) domain-specific realization conditions; and Rg(t) the realized state for group g. The indices g and d are analytically distinct but not independent: groups are situated within domains, and domain-specific conditions can shape both group-specific Exposure and realized states. The arguments are heterogeneous, are not presumed additive, and do not form a universal AGI score.
The surrounding components are positioned by relation to that center. Ceiling–Floor–Slope (CFS) is a bounded synchronization submodel; Optimization evaluates candidate trajectories upstream of realization; Governance is cross-cutting rather than an argument of μt and operates through four functions—Execution, Observation, Steering, and Procedural Adjudication—while contestation and standing describe the governed relationships and claims; and Participation enters through contingent pathways that may alter already recognized realization conditions without becoming a fifth scalar argument of μt. Governance’s four functions answer what governance does; its four points of purchase—Access, Sequencing, Distribution, and Deployment Context—identify where governance can intervene. The two four-part taxonomies are distinct and need not map one-to-one.
Observation and Evidence occupy a different logical position. They are the epistemic boundary governing what can be observed, represented, inferred, and assigned a claim status; they are not another realization input. Feedback and Recursion add the temporal return: prior realization, observation, representation, and response can alter later Potential, Availability, Exposure, domain conditions, institutions, objectives, and readiness. Because μt is a structured realization relation rather than an empirically identified invariant function, a claim that feedback changed μt itself carries a higher burden than a claim that later values of P(t), Av(t), Exg(t), or Zd(t) changed. Unless evidence can distinguish alteration of the relation from alteration of its arguments or domain conditions, the stronger μt-change claim should remain unadjudicated.
12.2.2 Components and organizing architecture
Figure 12.1 supplies the visual whole. Table 12.1 provides the complementary architectural reading: what question each major component answers, where it sits relative to μt, and which boundary prevents it from collapsing into neighboring concepts. The table is deliberately component-level rather than exhaustive; the item-level register remains in the separate companion Scholarly Appendix (Canonical Extraction Register).
Table 12.1. Components and Organizing Architecture of the Unified Gradual AGI Model
| Component | Question answered | Source | Placement relative to μt | Key boundary |
| MPF (organizing architecture) | What logical kind of claim is being made, and how may it be tested, revised, or related to other claims? | Ch. 3 | Outside the substantive realization relation; organizes claims about the model. | Plateaus are views, not stages; Plateau ≠ Facet; repair is not rescue. |
| Graduality | What counts as gradual realization, and when does principled revisability matter? | Ch. 2; restated Ch. 3 | Not an argument of μt; defines phenomenal membership and principled revisability. | Does not impose slowness or smoothness; phenomenal gradualness ≠ principled graduality. |
| Realization mapping | How does capability/potential become socially realized for a specified group under domain-specific conditions? | Ch. 4 | Central bridge: P(t), Av(t), Exg(t), Zd(t) are arguments of μt yielding Rg(t). | Potential ≠ Realization; heterogeneous arguments are not an additive scalar; notation ≠ empirical identification. |
| CFS / Synchronization | How can a domain-scoped capability state and realized state move relative to one another over time? | Ch. 5 | Nested bounded realization-dynamics submodel. | Ct ≠ P(t); Ft ≠ Rg(t); Slope is an observable, not an independent force. |
| Optimization | Which candidate trajectories are preferred under specified objectives, constraints, horizons, and uncertainty? | Ch. 6 | Upstream of realization; evaluates/selects trajectories and interventions. | Not an argument of μt; descriptive, formal-conditional, and normative claims remain distinct. |
| Governance | How are realization pathways contested, observed, steered, executed, and procedurally adjudicated, and whose claims have standing? | Ch. 7 | Cross-cutting, not an argument of μt; functions identify what governance does. | Points of purchase identify where it intervenes; functions ≠ points of purchase. Exposure remains distinct. |
| Participation | Who acts, in what form and relation, with what consequence, and when may that matter for realization? | Chs. 8–9 | Contingent pathways into already recognized conditions/routes; no universal coefficient. | Act ≠ Consequence ≠ realization change; human-side apparatus sits within reciprocal human–AI Participation. |
| Observation / Evidence | What was observable, how was it represented, and what inference/status does the evidence license? | Ch. 10 | Model-wide epistemic boundary governing claims about every component; distinct from Governance’s observation function and not an ordinary causal input. | Underlying State ≠ Observation ≠ Representation ≠ Inference ≠ Claim Status. |
| Feedback / Recursion | How can prior realization, observation, representation, and response alter later realization conditions? | Ch. 11 | Temporal return through later P, Av, Ex, Z and related institutions, objectives, and readiness; structural feedback to μ itself carries a higher burden. | Sequence ≠ feedback; feedback/recursion are not a fifth μt argument; recursive realization is not the same as recursive self-improvement. |
12.2.3 Key Ideas in One Line
The table locates the architecture. The following one-line summaries provide a faster conceptual index. They are reader-facing compressions rather than substitutes for the canonical definitions, propositions, and status statements preserved in the source chapters and the separate companion Scholarly Appendix (Canonical Extraction Register).
| Key idea | One-line meaning |
| AGI-inclusive humanity | The model’s highest-level system object: humans, AI systems, institutions, infrastructure, and the wider social environment, treated as one evolving system. The term originates in the series’ 2025 philosophical antecedent, First Principles of AGI-Inclusive Humanity, which framed humanity through nature-made and self-made dimensions and treated human-made artifacts, including AI, as extensions of the self-made side; that genealogy is retained, but the maturity ratio and “binary humanity” are not canonical model constructs. |
| Graduality | Realization can be gradual because it is not exhausted by one event; gradual does not mean slow, smooth, continuous, scalar, homogeneous, or uniformly incremental. |
| Depth and Width | Series #1’s orthogonal axes: Depth concerns epistemic reach—what kinds of knowledge, insight, and judgment become possible; Width concerns epistemic presence—how such contribution becomes persistent, accessible, and structurally embedded in society. In the mature model they remain genealogy rather than first-class variables of the realization mapping. |
| Potential and Realization | What AI has become capable of doing is not the same thing as what actually becomes socially realized. |
| Realization mapping | Rg(t)=μt(P(t), Av(t), Exg(t), Zd(t)) represents how heterogeneous capability and realization conditions combine to produce a realized state for group g. |
| Availability | Whether a capability is practically available for use, not merely technically possible. |
| Exposure | The group-specific way capability or its consequences reach people, position them to absorb effects, and leave them with meaningful recourse. |
| Domain-specific realization conditions Zd(t) | The domain-dependent conditions that help shape whether and how available capability becomes realized, without serving as a residual catch-all. |
| Multi-Plateau Framework (MPF) | The methodological-theoretical architecture that keeps definitions, representations, propositions, evidence, and system-level synthesis distinct while allowing them to be related, versioned, and revised. |
| Ceiling–Floor–Slope (CFS) | A bounded synchronization model for how a domain-scoped capability frontier and realized state can move on different clocks. |
| Optimization | The framework for comparing candidate trajectories under specified actors, objectives, constraints, horizons, and uncertainty without assuming one universal objective function. |
| Governance | The cross-cutting architecture through which realization is contested, executed, observed, steered, and procedurally adjudicated. |
| Participation | Intentional human action, represented through a human-side Subject–Act–Consequence apparatus within a broader reciprocal human–AI account, whose consequences may or may not become realization change. |
| Reciprocal participation | AI changes tasks, options, and environments within which humans act, while human action shapes the problems, uses, verification, governance, and conditions under which AI contributes. |
| Relational Grammar (RO, FP, AD) | A non-directional three-part description of realization orientation, foreclosure position, and actor dependence. |
| Responsive Floor | Tests whether pre-specified participatory differences are associated with differences in realization; a non-veto sequence is documented, but the aggregate verdict remains NotAdjudicated. |
| Observation and Evidence | The epistemic boundary separating what actually obtains from what is observed, represented, inferred, and ultimately warranted. |
| Epistemic repairability | The capacity to detect error, localize failure, revise the affected structure, propagate the consequences, and expose the repaired model to renewed testing. |
| Feedback | A prior realization, consequence, observation, or response becomes feedback only when it changes a condition relevant to later realization. |
| Recursion | Feedback becomes recursive when outputs of earlier realization re-enter and condition later realization through existing model components. |
| Recursive realization ≠ recursive self-improvement | Realized states may alter future realization conditions without implying autonomous recursive self-improvement. |
| No inferential promotion | Moving from capability to realization, act to consequence, association to causation, or another stronger claim requires additional evidence rather than being granted by the architecture. |
| Repair is not rescue | Revision after failure is legitimate only if the failure remains visible and the repaired claim is exposed to testing again. |
| Ongoing, Unfinished Transformation | At book level, a methodological orientation for theorizing an evolving transformation without presuming its final form; in a specified domain and horizon, ongoing and unfinished can become bounded process and closure-status assessments. |
| Open · Flexible · Falsifiable | The model is open in participation and provenance, flexible under revision and repair, and falsifiable under empirical contact. |
12.2.4 Notation and vocabulary
The overview uses the book’s settled notation in compact form; the separate companion Scholarly Appendix (Canonical Extraction Register) preserves the exhaustive canonical register.
| Notation | Meaning |
| Rg(t) | realized state for group g at time t |
| P(t) | Potential / capability |
| Av(t) | Availability |
| Exg(t) | group-specific Exposure |
| Zd(t) | domain-specific realization conditions |
| μt | the structured realization relation posited for the specified setting and time |
These objects are heterogeneous and are not additive terms in a universal score. The notation does not identify μt empirically as one universal function, and monotonicity is not assumed. A documented pathway is not automatically a counterfactually identified effect.
One inherited notation collision should be noted. Chapter 10 uses Stt for Underlying State, while Chapter 7’s recourse apparatus uses Stg(t) for an available standing institution. Chapter 12 avoids the latter shorthand in reader-facing prose. Where simplified reader-layer wording differs from local phrasing in a source chapter, the frozen architecture and canonical source definitions govern.
Indexing note. The target remains Rg(t), not Rg,d(t). The domain index d enters through Zd(t): the same group may be evaluated under different domain-specific realization conditions, and g and d are analytically distinct but not assumed independent. Chapter 12 does not introduce an additional domain index on the realized-state object.
Together, Figure 12.1, Table 12.1, and the one-line definitions provide the reader-level model in one view. Section 12.3 turns to the boundaries that keep this heterogeneous architecture coherent.
12.3 The Boundaries That Keep It One Model
The Unified Gradual AGI Model is unified by relation, not by collapse. Chapter 12 can place Potential, Availability, Exposure, domain conditions, Synchronization, Optimization, Governance, Participation, Observation/Evidence, and Feedback/Recursion inside one architecture only if the distinctions among them remain intact. The model therefore carries boundaries that are constitutive rather than decorative: they specify what may be connected, what must remain non-identical, what kinds of inference require additional evidence, and where the architecture must stop rather than absorb another concept into itself.
This is why the realization mapping does not function as a universal AGI equation. It supplies a central organizing relation, while the rest of the model is positioned around that relation according to type. Some objects are arguments of realization, some are upstream decision structures, some cut across realization, some are pathways through it, some are bounded submodels, and some govern what can be known about the process. The boundaries below are what keep those different logical roles from being flattened into one another.
12.3.1 Preserve non-identities: one architecture does not mean one variable
The first family of boundaries prevents conceptual compression. The model rejects a universal scalar for AGI, Participation, Governance, or realization. Local scalar representations may be useful when their scope is earned, but the Unified Model does not assume that capability, access, exposure, institutional conditions, participation, governance, and realized outcomes can be reduced to one commensurable number. Their heterogeneity is part of the architecture, not an inconvenience awaiting later simplification.
The same discipline applies to the status of claims. A definition is not a proposition, and a proposition is not a finding. A term may be well defined while the proposition that uses it remains untested; a proposition may be testable while the available evidence is non-operable or unreproduced; a finding in one bounded setting does not transform the definition itself into an empirical result. Keeping these logical types separate is one of the ways the Multi-Plateau Framework prevents the theory from promoting its own vocabulary into evidence.
At the substantive level, Potential is not Realization. A system can possess a capability without that capability being available, encountered, institutionally incorporated, acted upon, or realized for a particular group. Conversely, realized effects can feed back into later capability without erasing the analytical distinction between the two. The same non-identity recurs inside Participation. Readiness is not Action. A Participatory Act is not a Participatory Consequence, and a Participatory Consequence is not automatically a realization change. Participation is not identical to influence, and influence is not identical to causal effect.
These distinctions are not merely terminological. Each blocks an otherwise tempting inferential shortcut. Capability cannot be promoted directly into realized outcome; readiness cannot stand in for observed action; an act cannot be treated as consequential merely because it occurred; and a documented participatory pathway cannot be promoted into an identified causal effect without additional evidence. The architecture therefore preserves a chain of increasingly demanding claims instead of treating adjacency as entailment.
The same rule governs other inherited non-identities even where they are not part of the frozen thirteen-boundary list. In CFS, the Ceiling is not identical to Potential and the Floor is not identical to realized state. In the MPF, a Plateau is not a Facet. In Observation/Evidence, the Underlying State is not the Observation, the Observation is not its Representation, the Representation is not the Inference, and the Inference is not the Claim Status. These distinctions reinforce the same architectural principle: different analytical jobs require different objects.
12.3.2 Preserve placement: neighboring structures do not become extra inputs
A second family of boundaries governs placement relative to the realization mapping. Optimization remains outside μ. It compares candidate trajectories under specified actors, objectives, constraints, horizons, and uncertainty; it does not become another causal argument merely because optimization can alter what actors choose to do. This keeps evaluation and selection conceptually distinct from the mapping that describes realization.
Governance is likewise cross-cutting rather than a fifth input or a terminal settlement function. Governance can execute, observe, steer, and procedurally adjudicate; it can intervene at Access, Sequencing, Distribution, and Deployment Context; and its outcomes can change later realization conditions. None of this turns Governance into a single scalar quantity added to μ. Its architectural role is to contest and condition realization across multiple points rather than to close the process with a final governing state.
Exposure requires a particularly careful boundary because it sits inside the realization mapping while Governance operates around it. Group-specific Exposure is an argument of realization. It remains distinct from Governance’s four points of purchase and is not a fifth governance point. Reach, Susceptibility, and Recourse describe how a group is situated relative to capability and its consequences; governance institutions may affect those conditions, but that interaction does not make the two constructs identical.
The Multi-Plateau Framework is also kept outside the substantive process. Its Plateaus are logical views, not a pipeline through which AGI moves. P1 through P5 classify different kinds of theoretical work: what exists, what terms mean, how objects and relations are represented or measured, what is proposed or known, and how components relate across the larger system. They do not describe five temporal stages of transformation. The MPF organizes claims about the model; it is not a mechanism inside realization.
This placement discipline is what permits the Unified Model to remain architecturally rich without becoming algebraically indiscriminate. A component can matter greatly while remaining outside μ. A cross-cutting structure can alter arguments of μ without becoming one of them. A methodological framework can organize every claim in the model without becoming a causal process. Unity is therefore achieved through typed relations among components rather than by giving every important idea the same mathematical status.
12.3.3 Preserve epistemic discipline: revision cannot erase evidentiary history
The third family of boundaries protects the model against self-sealing revision. The architecture is intentionally flexible, but flexibility is constrained by provenance. Retired 2025 constructs may not be silently restored simply because a later problem makes them convenient again. If an earlier concept is reintroduced, the change must be explicit: its earlier retirement remains part of the genealogy, the reason for restoration must be stated, and the revised object must be treated as a new version rather than as if the intervening history never occurred.
The same rule applies to evidentiary status. Provisional, untested, non-operable, unidentified, or unreproduced results may not be silently promoted for the sake of producing a cleaner synthesis. Chapter 12 is therefore not licensed to convert a High-confidence candidate into a canonical principle merely because the book is ending, or to convert pathway evidence into causal identification because a domain illustration would read more neatly that way. The strongest warranted relationship should be stated, and the inference should stop there.
This discipline is central to the three design principles stated in §12.1. Openness makes the record visible; flexibility permits revision; falsifiability requires the model to preserve the possibility of failure. Epistemic repairability joins them. A failed or weakened claim may be localized, revised, and retested, but repair does not erase the historical fact that an earlier version failed its test. A model that can always reinterpret adverse evidence without carrying forward the cost of the failure would be flexible in appearance but unfalsifiable in practice.
For the same reason, Chapter 12 does not assign the Unified Model one global validation score. Components have different evidentiary histories, observation opportunities, formal burdens, and empirical contact. Some claims are supported; some remain finitely non-refuted; some are untested, unresolved, non-operable, unidentified, or unreproduced. Synthesis preserves that unevenness rather than averaging it away.
12.3.4 Preserve non-terminality: the model has no final settlement state
A final boundary concerns terminality. The Unified Model does not assume a terminal equilibrium or a final achieved AGI state. This follows from more than a preference for open-ended language. The architecture allows capability, Availability, Exposure, domain conditions, objectives, institutions, participation, observation, and realized outcomes to change over time and to alter one another through feedback. A state that looks settled under one domain, horizon, or institutional arrangement can therefore become provisional under another.
Governance does not remove this openness by supplying a final settlement function, and the MPF does not remove it by supplying a final stage. Likewise, ‘Ongoing, Unfinished Transformation’ is not a promise that every domain changes forever. At the book level it is an architectural orientation that refuses to assume closure in advance; at the domain level, ongoing and unfinished can become bounded claims about process and closure status that require evidence. The absence of a universal endpoint therefore does not mean that local processes cannot stabilize, plateau, terminate, or become analytically closed for a specified purpose.
Taken together, the thirteen frozen boundaries define the negative space of the Unified Gradual AGI Model. They say what the model will not collapse, what it will not promote, what it will not treat as an extra input, and what it will not declare finished merely for the sake of theoretical closure. Their function is constructive: by preventing false equivalences, they allow distinct components to interact without losing their identity. That is what makes the architecture one model rather than either a single oversized equation or a loose collection of adjacent concepts.
Frozen boundary register. The thirteen boundaries carried forward into Chapter 12 are:
1. No universal scalar for AGI, Participation, Governance, or realization. (§12.3.1; §12.7.1)
2. Definition is not proposition; proposition is not finding. (§12.3.1; §12.6)
3. Potential is not realization. (§12.3.1; §§12.4–12.5)
4. Readiness is not action. (§12.3.1; Chapters 8–9)
5. Participatory Act is not Participatory Consequence, and Participatory Consequence is not realization change. (§12.3.1; §12.6.3)
6. Participation is not influence, and influence is not causal effect. (§12.3.1; §12.6.3)
7. Optimization remains outside μ. (§12.3.2)
8. Governance is cross-cutting; it is not a terminal settlement function. (§12.3.2; §12.7)
9. Exposure is an argument of realization and remains distinct from Governance’s four points of purchase; it is not a fifth governance point. (§12.3.2)
10. Plateaus are not a pipeline. (§12.3.2)
11. No terminal equilibrium or final achieved AGI state is built into the architecture. (§12.3.4; §12.9)
12. No silent restoration of retired 2025 constructs. (§12.3.3)
13. No silent promotion of provisional or unreproduced results. (§12.3.3; §12.6)
The next sections can therefore apply the architecture without renegotiating its basic types. The empirical question is not whether every domain exhibits every component in the same form. It is whether the model’s distinctions and relations remain useful when brought into contact with different domains, different observation opportunities, and different evidentiary burdens.
12.4 The Model Applied: Three Domains
The architecture becomes useful only when it can organize materially different domains without forcing them into the same trajectory. This section therefore applies the Unified Gradual AGI Model to three deliberately heterogeneous cases: Go as a mature post-superhuman comparator, mathematics as an emerging open-knowledge transition, and radiology as an emerging regulated-profession transition. The comparison is not a maturity ranking and not a forecast. It asks the same architectural questions of each domain while preserving their different objectives, institutions, verification practices, participation structures, and evidentiary burdens.
The exercise follows the realization mapping rather than a single outcome metric. For each domain, the analysis asks what capability event changed Potential; what is practically Available; how Exposure is distributed; which domain-specific realization conditions matter; how Governance and Participation shape the pathway; whether later responses constitute feedback; what evidence actually supports the resulting claims; and whether any terminal configuration is justified. A domain can be advanced on one dimension and unresolved on another. That asymmetry is expected rather than treated as measurement failure.
12.4.1 Why these three domains
The three cases create a useful contrast because they place similar questions under different institutional conditions. Go has fixed rules, a stable objective, dense repeated trials, rapid feedback, and roughly a decade of post-AlphaGo adaptation. Mathematics has open-ended problem selection, multiple standards of significance, formal and informal verification, attribution and priority, pedagogy, mentorship, and institutional recognition. Radiology places AI inside a regulated clinical profession in which task performance must be separated from deployment, liability, workflow, professional responsibility, and patient-facing consequences.
Their temporal positions also differ. Go is observed retrospectively after a decisive capability event. Mathematics is being observed while capability, disclosed adoption, verification practice, contestation, and institutional response are changing at once. Radiology currently offers strong bounded-task and reader-assistance evidence but only prospective evidence about profession-level realization. These temporal labels describe observation position and evidentiary horizon, not a maturity ranking. Go’s longer post-superhuman record is used only as a possibility proof that machine superiority in a central technical task can coexist with persistent human institutions and role reconfiguration; it is not a template or forecast for mathematics or radiology. This temporal heterogeneity is analytically useful: it prevents the Unified Model from mistaking one domain’s history for a universal sequence.
12.4.2 Go: a mature post-superhuman comparator
Go supplies the clearest mature comparison because the capability event is historically distinct. AlphaGo’s 2016 victory over Lee Sedol demonstrated superhuman performance in the domain’s central technical task (Google DeepMind 2026). Yet that capability event did not terminate professional Go. A decade later, professional competition, teaching, commentary, and institutional play remain active. What changed was the organization of expertise around machine evaluation, training, interpretation, and strategy.
The longitudinal evidence is consistent with that distinction. Choi et al. (2025a), using 749,190 professional moves, report improved move quality after the introduction of AI-powered Go, greater alignment with superior AI solutions, and fewer and smaller errors, with heterogeneous effects across age and skill. Choi et al. (2025b) report AI-catalyzed knowledge gains alongside increased concentration and reduced diversity in move selection, while Shin et al. (2023) report increased novelty in human decisions after the advent of superhuman AI. These results do not imply that all human roles improve or that AI influence is uniformly beneficial; they show that machine superiority in move selection can coexist with continued human participation and altered expert practice.
The full realization architecture helps explain why the domain did not collapse into a simple replacement story. Potential changed sharply when superhuman play became available. Availability subsequently expanded through training and analysis tools. Exposure became routine for players, coaches, commentators, and audiences through AI-mediated evaluation, while the effects varied by role, skill, and institutional position. Domain-specific realization conditions—fixed rules, rapid evaluative feedback, dense repeated trials, established competition structures, and a clear win-loss objective—made AI output unusually easy to compare and incorporate. Those conditions are not shared by mathematics or medicine.
Participation also changed rather than disappearing. The relevant Role Plasticity question is no longer whether humans can outperform the machine at move selection, but how players learn, interpret, teach, explore, compete, and construct meaning after machine superiority. The feedback pathway is correspondingly concrete: AI evaluation and training alter human learning and practice; changed practice then becomes part of the next round of AI-mediated play and interpretation. Go therefore supplies a mature example of transformation without terminality: a decisive capability threshold changed the domain without finishing it.
12.4.3 Mathematics: an emerging open-knowledge transition
Mathematics is the fastest-moving and least settled of the three cases. The capability side is no longer hypothetical: frontier systems can generate, formalize, and verify advanced mathematical work, and highly visible 2026 episodes have pushed questions of proof, attribution, verification, authorship, and institutional authority into the foreground. OpenAI’s September 8 release of a claimed Navier–Stokes solution with a Lean formalization is useful here as a capability-and-evidence illustration, not as a shortcut to settled mathematical realization (OpenAI 2026a). As of September 26, the Clay Mathematics Institute continued to list Navier–Stokes among its active Millennium Prize Problems (Clay Mathematics Institute 2026). A company release, a formal artifact, community verification, priority, attribution, and eventual institutional acceptance are therefore analytically distinct states.
Because this case is being observed in real time, the 2026 episodes are used as dated snapshots of changing realization conditions rather than as settled evidence of long-run mathematical or institutional effects.
The adoption trajectory is now measurable as well. Epoch AI’s September 18 analysis, authored by Tara Abrishami, reports in its published data table that the share of arXiv mathematics preprints explicitly acknowledging AI use rose from 3.89 percent in April 2026 to 25.25 percent in August. In August, 1,453 of 5,754 preprints acknowledged AI use, and 337—5.86 percent of all mathematics preprints that month—credited AI with a substantial research contribution. These figures measure disclosed use, not total prevalence, and they do not by themselves establish contribution quality, causal influence, or Exposure. Adoption is evidence about realized use; it is not a substitute for the model’s other arguments.
Availability is therefore changing, but it remains uneven. Access to frontier systems, formal-verification tools, compute, interfaces, and institutional support differs across researchers and organizations. SAIR’s September Open Math Model initiative is best treated as a prospective response aimed at open-weight scientific models and tooling; it should not be counted as present Availability before the proposed infrastructure exists. Exposure is likewise group-specific: researchers can differ in direct access, dependence on AI-mediated work, vulnerability to changing standards, and ability to contest or correct outputs and attribution decisions.
The domain-specific realization conditions are correspondingly rich. Verification, journals, conferences, priority, attribution, pedagogy, mentorship, prestige, formal proof systems, and expert communities mediate whether an AI-generated result becomes mathematical realization. Governance and Participation therefore become unusually visible. The Leiden Declaration, Mathathon’s verification response, laboratory disclosure practices, and new open-model initiatives are examples of institutional contestation and adaptation. They matter not because any one response settles the field, but because they can change later conditions of access, verification, recognition, and acceptable participation.
Mathematics is also a strong prospective test of reciprocal Participation. Human roles may shift among proving, selecting, verifying, interpreting, integrating, teaching, and adjudicating; AI roles may shift among instrument, generator, verifier, collaborator-like actor, and object of contestation. Chapter 9 does not license importing a global verdict into this domain: Role Plasticity was Untested in Frame One, Conditional Non-Fungibility had discriminant evidence but unresolved validation, and the aggregate Responsive Floor remained NotAdjudicated. Mathematics instead supplies a new longitudinal setting in which those propositions can be tested under different conditions. The appropriate Relational Grammar is the non-directional profile (RO, FP, AD), not a presumed sequence.
The strongest feedback claim should remain modest. AI-generated work has already prompted verification, contestation, disclosure rules, event redesign, and infrastructure proposals. Those are documented responses. A recursive claim requires the further step: evidence that the changed institutional or technical conditions re-enter later mathematical realization and alter what happens next. That longer loop is emerging, not yet established as a universal pattern. Mathematics therefore illustrates both the power and the discipline of the Unified Model: capability can accelerate faster than the institutions that validate, distribute, contest, and absorb it.
12.4.4 Radiology: an emerging regulated-profession transition
Radiology provides a complementary test because the realization object is embedded in clinical responsibility. The 2026 RADAR study, published in Science, reports expert-level performance on a bounded abdominal-CT task set. The model was trained on more than 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-wise image-text pairs, and it was evaluated across 18 anatomical structures and 146 imaging findings. In a reader study, assistance from RADAR increased the diagnostic sensitivity of 26 radiologists by about 10 percent.
That is strong task-level and assistance evidence, but it is not profession-level realization. Since publication, the project has also made training and inference code and pretrained checkpoints publicly available for research use (Alibaba DAMO Academy 2026). This expands one dimension of Availability: researchers can inspect, reproduce, adapt, and run the system more directly than when capability was represented principally by the published result. Yet the project’s own public materials continue to state that RADAR is intended for research purposes and that further improvements and prospective clinical studies are required before direct clinical deployment. Potential has therefore moved substantially and research Availability has widened, while clinical Availability, Exposure, and realized institutional use remain more bounded. The distinction matters because greater technical accessibility does not by itself constitute clinical realization.
Domain-specific realization conditions in radiology include prospective validation, clinical workflow, accountability, liability, regulation, hospital procurement, interoperability, patient safety, and professional sign-off. These conditions are not residual obstacles external to the model; they are part of the domain-specific pathway through which capability may or may not become realized. Exposure likewise depends on where AI enters workflow and who bears consequences or possesses recourse, not simply on whether a model exists.
The reader study is especially important for Participation because it directly observes human–AI assistance rather than inferring replacement from benchmark performance. It supports a complementarity pathway in a bounded setting, while longer-run Role Plasticity, Foreclosure Position, Actor Dependence, occupational adaptation, and Participation-Sensitive Realization remain prospective. A radiologist may shift from direct detection toward integration, adjudication, supervision, communication, or exception handling; whether those shifts occur at scale is an empirical question, not a conclusion licensed by present task performance.
The feedback pathway is likewise prospective. AI assistance may alter workflow, oversight, training, procurement, and role allocation, which could then condition later deployment and system design. But sequence alone is not feedback. A stronger claim requires evidence that those responses actually changed later realization conditions and affected subsequent practice. Radiology therefore provides the clearest reminder among the three cases that expert-level capability and professional transformation are different objects.
12.4.5 Table 12.2 — Three-domain application grid
Table 12.2 compresses the three applications into a common set of questions. It does not convert them into a common scale.
| Dimension | Go | Mathematics | Radiology |
| Temporal position | Mature post-superhuman comparator. | Emerging knowledge-production transition with observed disclosed-adoption trajectory. | Emerging regulated-profession transition. |
| Capability / Potential | Superhuman play demonstrated by AlphaGo. | Frontier systems generate, formalize, and verify advanced work; 2026 high-profile results stress verification and attribution. | RADAR reports expert-level performance on bounded abdominal-CT tasks. |
| Availability | AI analysis and training tools are routinely accessible within professional practice and commentary. | Frontier access and formal tools are expanding but uneven; SAIR open-model infrastructure is prospective, not current Availability. | Research code and pretrained checkpoints publicly available for research use; direct clinical deployment not established. |
| Exposure / Governance | Players and institutions encounter AI evaluation through training, commentary, and competition infrastructure; effects vary by role. | Differential access, disclosure, verification, attribution, and recourse; active contestation and institutional response. | Exposure depends on clinical workflow, regulation, liability, deployment, and patient-facing accountability. |
| Domain-specific realization conditions | Fixed rules, stable objective, dense trials, rapid feedback, established competition institutions. | Verification, peer review, journals, priority, attribution, pedagogy, formal proof systems, institutional recognition. | Prospective validation, workflow, procurement, interoperability, liability, regulation, professional sign-off. These are treated as Zd(t) conditions, not as external friction. |
| Participation / human role | Role Plasticity observed after machine superiority as an empirical domain pattern; this does not adjudicate Chapter 9 Proposition B, which remains Untested in Frame One. | Prospective test of Role Plasticity (Proposition B: Untested in Frame One), Conditional Non-Fungibility (discrimination evidence; validation unresolved), and Participation-Sensitive Realization / Responsive Floor (aggregate NotAdjudicated) across proving, selecting, verifying, interpreting, and teaching. | Reader-assistance complementarity observed; profession-level Role Plasticity and occupational adaptation remain prospective. Chapter 9 Proposition B remains Untested and the aggregate Responsive Floor remains NotAdjudicated. |
| Feedback / institutional response | AI evaluation and training alter human learning and practice; changed practice enters later AI-mediated activity. | Generated work and adoption prompt verification, contestation, disclosure, event redesign, and infrastructure proposals; repeated recursive return remains emerging. | Possible pathway from assistance to workflow, oversight, training, and role adaptation; longer feedback loops remain prospective. |
| Current evidentiary status | Mature comparator with independent longitudinal studies. | Pre-AI baseline + observed 2026 disclosed-adoption growth + documented institutional adaptation; long-run role effects remain emerging. | Peer-reviewed task and reader-assistance evidence; profession-level effects remain prospective. |
| Terminality diagnostic | No terminal domain state: professional Go persists in altered form after superhuman play. | No stable terminal configuration established for the current domain and observation horizon. | No terminal occupational or institutional configuration established. |
Note. Statuses in Table 12.2 are categorical, not ordinal. Cells are not scores, should not be counted, and should not be compared numerically across domains. The table is an application grid, not a maturity ranking. Adoption is not Exposure; task performance is not profession-level realization; and sequence is not feedback unless a changed condition is demonstrated.
12.4.6 What the comparison shows—and what it does not
Across all three domains, the most stable finding is not a common rate of change but a common need for separation. Capability events matter, yet they do not determine realization by themselves. Availability can expand without uniform adoption. Adoption can increase without equal Exposure. Governance can alter access, standards, and recourse without becoming a fifth realization input. Participation can remain consequential after machine superiority without guaranteeing a particular outcome. Feedback can condition later rounds without implying autonomous recursive self-improvement.
The domain comparison also shows why the Unified Model should resist a universal scalar. Go looks mature because the post-superhuman period is long and densely observed, but its objective structure is unusually simple relative to mathematics and radiology. Mathematics has rapidly rising disclosed AI use and intense institutional adaptation, but its long-run role effects remain emerging. Radiology has strong bounded-task and assistance evidence, but clinical deployment and profession-level effects remain prospective. None of those descriptions is a higher or lower score. They are different configurations of evidence and realization.
The comparison therefore does not license a forecast from Go to mathematics or radiology. Go establishes a possibility: machine superiority in a central technical task can coexist with persistent human institutions and substantial role reconfiguration. Mathematics and radiology test whether related patterns emerge under different verification regimes, accountability structures, and stakes. Their results may converge, diverge, or force the model to contract. That possibility is part of the architecture rather than a threat to it.
The practical payoff is a disciplined sequence of questions. When a new AI capability appears, the model asks not only whether it is impressive, but whether it is available, who is exposed, which domain conditions mediate realization, who participates and with what consequence, what institutions contest or steer the process, what later conditions change, and what the evidence actually licenses us to say. Section 12.5 turns that sequence into an explicit reader workflow for evaluating claims about AGI.
12.5 Using the Model: Questions for Any Claim About AGI
The Unified Model is meant to change how claims about AGI are examined, not merely how the theory is diagrammed. A useful application begins by asking less dramatic questions than ‘Has AGI arrived?’ or ‘What will happen next?’ It asks what kind of object is being discussed, what was actually observed, what relationship the evidence supports, and which boundary would have to be crossed to justify a stronger claim. The same discipline applies whether the claim concerns a benchmark result, an adoption statistic, a governance proposal, a participation pathway, or a purported recursive improvement loop.
This section therefore does not add another protocol to the architecture. It brings together procedures already established in Chapters 3, 4, 5, 10, and 11 and shows which one governs which analytical task. The Multi-Plateau Framework identifies logical type and version. Chapter 4 governs claims about realization and the admission of domain conditions. Chapter 5 supplies a stop rule when CFS cannot be operationalized. Chapter 10 governs the move from observation to inference. Chapter 11 raises the burden for feedback, recursion, and recursive self-improvement. Repair is invoked only after adverse contact, not as a substitute for it.
12.5.1 First question: What object is the claim about?
A claim should first be assigned to the object it is actually about. Capability or Potential is not realization. Availability is not Exposure. Governance is not Participation. An observed act is not its consequence. A consequence is not automatically a change in realization. A temporal sequence is not feedback. These separations are not terminological niceties; they determine what evidence can bear on the claim and what stronger conclusions remain out of reach.
Chapter 3’s minimum application mode offers a compact first pass. The analyst asks: (1) what logical role is being performed; (2) what information would be lost if only that role were retained; (3) what the principal relations are and whether a mistyped relation would change the inference; (4) what would count as adverse contact; and (5) which version was tested and what dependent conclusions must be revisited if the object changes. The point is not to label every sentence with a Plateau. It is to prevent a definition, a measurement, a proposition, an observation, and a system-level interpretation from inheriting one another’s status merely because they concern the same topic.
For claims about social realization, Chapter 4 adds a second discipline. The realization object must be specified before an outcome is interpreted. If a candidate domain condition is proposed, the analyst should ask why it is domain-specific, why it does not belong more naturally in Potential, Availability, or Exposure, and what evidence could show that it matters. The burden lies with added complexity: a condition that cannot be distinguished empirically, is supported only after the outcome is known, or adds no explanatory resolution should be merged, omitted, or left unclaimed rather than preserved inside the domain-condition term.
12.5.2 Second question: What was actually observed or represented?
Once the object is identified, the next task is to separate the underlying state from the observation and the representation used to describe it. A benchmark score is a representation of observations produced under an evaluation design. A usage percentage is a measurement under a particular disclosure or logging regime. A governance proposal is a document specifying possible future action, not evidence that the proposed mechanism has been exercised. A model-development chronology is a sequence of recorded events, not yet evidence of a feedback loop.
Chapter 10’s five-term separation is therefore operative throughout Chapter 12: Underlying State is not Observation; Observation is not Representation; Representation is not Inference; and Inference is not Claim Status. Each transition introduces a possible failure point. Observation opportunity may be inadequate. A proxy may not survive a version change. A representation may hide heterogeneity. An inference may exceed what the design identifies. A claim status may therefore remain unresolved even when the underlying event is real and well documented.
This separation also blocks a common shortcut in AGI discourse: treating visibility as prevalence. Public demonstrations, benchmark records, disclosed AI use, reported incidents, or visible institutional responses are observations made under particular regimes. They can be important without being complete. Where observation opportunity is limited, non-observation cannot be promoted into absence.
12.5.3 Third question: What inference does the evidence warrant?
Chapter 10 provides the evidentiary language for this step. A descriptive observation establishes that an event, record, measurement, or state was observed within a specified boundary. Pathway evidence adds a documented sequence or institutional or technical connection. Provenance attribution links an artifact or result to a source or process. Association establishes systematic co-variation. Causal contribution requires evidence that a factor made a difference within a jointly caused process. An identified causal effect requires an explicit counterfactual contrast under stated assumptions.
These claim types are not a ladder of prestige. A carefully bounded descriptive observation may be exactly the right answer to a descriptive question. The discipline is to state the strongest relationship the evidence supports and stop there. Provenance does not by itself establish causal contribution. Association does not establish causation. Pathway evidence can document consequential linkage without quantifying a unique effect. Lack of identification does not mean that no contribution occurred; it means that the stronger causal claim has not been earned.
For realization claims, Chapter 4 adds an empirical application protocol: specify the domain and realization object; operationalize the relevant Potential, Availability, Exposure, and domain-condition states where feasible; specify a small set of competing realization mechanisms; identify observations capable of discriminating among them; distinguish pathway evidence from causal or counterfactual identification; compare the richer architecture with a simpler baseline; and retain only complexity that improves prediction, mechanism discrimination, or explanatory resolution. No single study must identify every element. What matters is that measured, assumed, untested, and unresolved parts remain visible.
12.5.4 Fourth question: Which boundary must be crossed to claim more?
A stronger claim is legitimate only when the additional evidentiary burden corresponding to the next boundary has been met. Capability evidence must cross Availability, Exposure, domain conditions, and a defensible realization relation before it supports a claim about social realization. Adoption evidence must cross group-specific Exposure before it supports claims about who is reached, affected, or able to seek recourse. A governance design must be exercised before it becomes evidence of governance action, and exercised action must alter a later condition before it becomes evidence of feedback. A participatory act must reach consequence and then realization before it licenses a claim about Participation-Sensitive Realization.
The same rule limits formal apparatus. CFS is applicable only when a Ceiling, Floor, longitudinal measurement, and plausible coupling can be specified without material ambiguity. Otherwise the correct verdict is ‘CFS not presently applicable.’ The stop rule is not a weakness of the model. It prevents a bounded submodel from becoming a universal explanation by adding variables or regimes simply to force every domain to fit.
Where adverse contact does occur, Chapter 3’s repair sequence governs what happens next: exposure to test; failure detected; failure localized; dependencies identified; component revised; consequences propagated; revised claim retested. The historical failure remains in the record, and the repaired object does not inherit the earlier object’s validation. This is how flexibility remains compatible with falsifiability.
12.5.5 Four worked examples
The following four examples remain reader-facing illustrations of the architecture. Examples 1 and 3 use dated 2026 cases because they expose the model’s boundaries especially clearly; they are illustrations of how to reason from evidence, not load-bearing adjudications of the broader frontier claims.
1. Frontier scientific capability results: mathematics and theoretical physics. A striking scientific output first establishes a claim about what a system did under a specified observation and verification regime; it does not by itself establish field-level realization, conceptual novelty, or a new scientific principle. Mathematics provides one version of this problem through formal proof, verification, and attribution. A September 2026 theoretical-physics case provides another. Anthropic reported that Claude Science computed the six-particle amplitude in planar N=4 super-Yang–Mills theory at nine loops, extending the prior eight-loop result of Dixon and Liu (2023), and Lance Dixon independently validated the result (Anthropic 2026). Dixon emphasized that Claude executed an exceptionally fragile but inherited calculational program; in his view, the deeper scientific threshold would be AI producing new physical principles or insights before humans. Matt von Hippel likewise interpreted the episode as evidence that an assumed computational barrier was lower than expected, not as the discovery of a new physical method, both in the Anthropic guest post and in his own follow-up (Anthropic 2026; von Hippel 2026), while Woit (2026) similarly distinguished the complexity of the calculation from conceptual novelty. In the Unified Model, the result therefore bears on Potential, realized scientific work, Participation, and Observation/Evidence, while the separate questions of conceptual novelty, community acceptance, and—outside a formal toy theory—empirical physical realization retain their own burdens.
2. An adoption statistic. Suppose a survey, telemetry system, or disclosure analysis reports rapid growth in the use of an AI system within a profession. That is evidence about use under a defined observation regime. Depending on the measure, it may inform Availability or realized use; it is not automatically Exposure, because Exposure also concerns group-specific Reach, Susceptibility, and Recourse. Nor does adoption alone establish benefit, substitution, productivity, causal contribution, or profession-level transformation. To say more, the analyst must specify who adopted what, for which task, under which institutional conditions, and what changed because of that adoption. A high adoption rate may coexist with narrow authority—decision rights concentrated in a limited set of actors or institutions—weak consequence, or highly unequal exposure.
3. A governance architecture under contestation. OpenAI’s September 21 proposal, Building standards for the next phase of AI, calls for internationally shared technical standards for frontier AI and automated AI research, including common measurements of RSI-relevant progress, thresholds for human review of automated research, and shared incident-classification and reporting protocols (OpenAI 2026b). The proposal explicitly says those standards would not themselves be licenses, mandatory prerelease review, or model-approval requirements; national governments would decide whether and how to give them legal force. On September 22, OpenAI separately published Priorities and principles for effective third party assessments, calling for independent assessors to receive sufficiently deep access across training, evaluation, and deployment to challenge laboratory assumptions and reach their own conclusions about safeguards (OpenAI 2026c). The proposed assessment scope includes safety cases, critical safeguards, frontier capability evaluations—including AI self-improvement—and investigation of serious misalignment incidents, together with requirements concerning assessor independence, conflicts of interest, methodology, access, actionable findings, and publication.
The two proposals instantiate different but connected parts of the Unified Model. Shared standards concern Observation, Steering, and Procedural Adjudication across institutions; independent assessment inserts an additional observational and adjudicative layer between developer claims and wider governance. This is especially important where safety claims concern systems whose relevant behavior, safeguards, or internal deployment conditions cannot be fully assessed from public outputs alone. The architecture is therefore moving beyond a generic call for external oversight toward specified objects of assessment, access conditions, institutional roles, and evidentiary burdens.
Contestation remains part of the governance object rather than noise around it. Sam Altman argued that people outside frontier laboratories should have a real say and that standards should help prevent concentration of power (Altman 2026); immediate public discussion questioned who should lead such a regime, what would bind laboratories, and whether frontier developers could credibly design mechanisms intended partly to hold themselves accountable (Hacker News 2026). A contemporaneous Reuters/Ipsos survey found that 73 percent of U.S. adults were concerned that AI companies were not doing enough to prevent serious harm and that most respondents favored a major federal role in setting safety standards; this measures broader public sentiment rather than reaction to the OpenAI proposal itself (Lange 2026). Analytically, the episode therefore combines Observation, Steering, Procedural Adjudication, standing, contestation, and potential feedback into later realization conditions. It is stronger evidence of an emerging governance architecture than a generic policy statement, but it remains evidence of design, specification, and institutional positioning—not yet evidence that the proposed assessment regime has been executed at sufficient scope or that it has changed realization.
4. A recursive-self-improvement claim. Suppose AI is used repeatedly in the development of later AI systems. The first task is to separate sequence, feedback, adaptation, recursive re-entry, recursion, and RSI. Repeated improvement is not automatically feedback; feedback is not automatically recursion; recursion is not automatically self-improvement. Strong evidence for RSI would require repeated AI participation in identifying improvement opportunities, designing or implementing changes, validating results, and reusing improved capability in another cycle, with human direction no longer supplying the key technical decisions at every step. The nine-loop physics episode sharpens the boundary. Claude reportedly carried a difficult frontier calculation for days with very little scientific supervision. Subsequent provenance also matters: Matt von Hippel reported that Song He’s group at the Chinese Academy of Sciences had independently obtained most of the result through a more human-directed pathway using GPT-6 assistance, and Lance Dixon later reported that the group had also computed the nine-loop symbol; Anthropic’s account links the Song He, Jirong Jing, and Xiang Li work as a concurrent nine-loop result (Anthropic 2026). The episode therefore exhibits two differently organized human–AI research pathways approaching the same frontier rather than a uniquely machine-only crossing. That distinction changes the interpretation of the episode without changing the RSI boundary: neither pathway improved a successor AI system and then reused that improved capability in a further self-improvement cycle. The episode is evidence of substantial AI-mediated scientific production and, in the Claude pathway, unusually high execution autonomy; it is not evidence of RSI.
12.5.6 A compact reader workflow
For ordinary use, the full architecture can be compressed into four questions:
1. What object or component is the claim about? Identify the logical type and architectural location before interpreting the result.
2. What was actually observed or represented? State the observation regime, measurement, representation, version, population, domain, and time boundary that matter to the claim.
3. What inference does the evidence warrant? Use the narrowest relationship that captures the evidence: description, pathway, provenance, association, causal contribution, or identified effect.
4. Which boundary must be crossed to say more? Name the additional observation, mechanism, comparator, exercise, or identification burden required for the stronger claim; if it is unavailable, stop at the current status.
These four questions are intentionally simpler than the full ledger, but they preserve the model’s central discipline. They keep capability distinct from realization, evidence distinct from inference, design distinct from exercise, and sequence distinct from feedback. They also make explicit where a claim must stop when the required burden has not been met.
The Unified Model therefore does not ask readers to accept a master answer to every AGI claim. It gives them a structured way to determine what kind of claim is being made, what evidence bears on it, which distinctions must remain visible, and what would have to happen before a stronger conclusion becomes warranted. That is the practical meaning of an architecture designed to be open, flexible, and falsifiable.
12.6 What the Model Claims, and With What Status
A unified model can become misleading if synthesis makes all of its claims appear equally settled. The Unified Gradual AGI Model therefore carries forward the statuses earned in the source chapters rather than replacing them with a single chapter-level verdict. Definitions remain definitions; synthetic principles retain their admission tier; empirical propositions retain the result of actual contact; and unresolved or unexercised claims remain unresolved or unexercised. Chapter 12 does not promote a claim merely because it is useful to the architecture.
This section states the model’s principal positive claims in that inherited form. It is not a new proposition set. The identifiers and statuses below come from the frozen Architecture, the Multi-Plateau Framework (MPF), and the relevant component chapters. Where a claim is canonical, that does not automatically mean that it is an empirically identified law; where a claim is provisional, that does not mean that it has been refuted.
12.6.1 Claims fixed at the architectural level
Phenomenal gradualness — Frozen architectural definition. Given an individuation rule fixing the transformation under study, realization is gradual when it is not exhausted by one event. The criterion does not require slow, smooth, continuous, scalar, homogeneous, or small-step change. This is the book’s settled answer to the recurring mistake that gradual must mean slow.
Principled graduality — Frozen architectural definition. Information-responsive iteration plus preserved revisability is a decision/process principle. It is not entailed by phenomenal gradualness. The model therefore distinguishes a descriptive claim about how a transformation is individuated from a normative or procedural commitment to revisable action.
Potential–Realization Separation (SYN-003) — Canonical (cross-corpus established). What has become possible and what has actually become realized are distinct objects and must be modeled separately. This is the central positive claim behind the realization architecture and the reason capability results cannot stand in for social realization.
No Inferential Promotion (SYN-002) — Canonical (derived). Later-level claims require their own support. Capability does not imply realization; Subject does not imply Act; Act does not imply Consequence; Consequence does not imply realization change; participation does not imply influence; detection does not imply correction; and association does not imply causation. The principle is canonical by valid composition from already-established boundaries rather than by pretending to be a separate empirical law.
12.6.2 Synthetic principles that remain active but not fully promoted
Here SPAT refers to five inherited admission checks: source anchoring; recurrence or constituent sufficiency; non-redundancy/discriminability; status-inheritance discipline; and a stated defeat condition or scope boundary. The full displayed walk-throughs still open for SYN-001, SYN-004, SYN-005, and SYN-006 remain prospective work rather than content supplied by the current Canonical Extraction Register. If completed, they should be recorded in a future revision of the separate Scholarly Appendix (Canonical Extraction Register) or subsequent series work, rather than retrofitted into this chapter as though already adjudicated.
Local Scalarity / Global Non-Scalarity (SYN-001) — High-confidence candidate. Bounded actors may use local scalar objectives, but no canonical universal scalar has been justified for the Unified Gradual AGI system as a whole. Its present status is deliberately below canonical because the full displayed Synthetic Principle Admission Test (SPAT) walk-through remains open.
Indexed Optimality (SYN-004) — High-confidence candidate. An optimum is relative to actor, objective, constraints, scope, and horizon unless a broader evaluative rule is separately justified. The model therefore permits local optimization without inferring a unique civilization-scale objective. Full displayed SPAT treatment remains open.
Direction Without Terminality (SYN-005) — High-confidence candidate. Persistent normative direction and continuing governance are compatible with the absence of a uniquely specifiable terminal optimum or solved state. The claim permits direction without converting the model into a theory of inevitable convergence. Full displayed SPAT treatment remains open.
Procedural Settlement ≠ Substantive Settlement (SYN-006) — High-confidence candidate. Institutions may settle procedure, authority, or entitlement while leaving contested ultimate ends unresolved. A decision can therefore close a process without proving that the underlying normative dispute has been substantively settled. Full displayed SPAT treatment remains open.
Recursion Without Ontological Proliferation (SYN-008) — High-confidence candidate. Feedback can change the states of existing objects and relations without requiring a new fundamental category for every feedback product. This is the architectural reason Feedback and Recursion return into existing realization conditions rather than becoming a fifth argument of the realization mapping.
Genealogy note. SYN-007, Governance Can Be Recursive, is not an active independent principle. It is retained genealogically as a domain-specific corollary merged into SYN-008.
Realization Processes Are Reflexive (SYN-009) — High-confidence candidate. Realization can modify conditions relevant to later realization, including institutions, exposure, readiness, objectives, observation, participation, and resource or availability conditions. The claim is about possible endogenous alteration of conditions, not a claim that every temporal sequence is feedback.
Adaptive Disequilibrium (SYN-010) — Provisional. Gradual AGI may move through provisional configurations disrupted by capability, institutional, technological, or participatory change, followed by adaptation without convergence to a final equilibrium. Its defeat condition is stated but has not been exercised. Chapter 12 therefore preserves its provisional status rather than treating the capstone synthesis as promotion.
12.6.3 Participation claims after empirical contact
Participation provides the clearest example of why claim status must remain heterogeneous. Chapter 8 defined Propositions A–C conceptually; Chapter 9 then subjected them to a bounded empirical frame. The result was not a single verdict for Participation as a whole.
Proposition A — Conditional Non-Fungibility — Discrimination/nonredundancy evidence observed; validation unresolved. In the executed coding, same or similar Act Forms preserved different Realization Orientation, Foreclosure Position, and Actor Dependence states, conditional on faithful application of the instrument. But intersubjective reproducibility cannot be reconstructed from the surviving archive. The discrimination result therefore survives while instrument validation remains unresolved pending prospective replication.
Proposition B — Participatory Role Plasticity — Untested in Frame One. The required human-task automation antecedent was not instantiated in the frozen empirical frame. This is not a null result and not evidence against Role Plasticity. It is an unexercised empirical burden that requires a design in which human and AI functions can actually be observed redistributing over time.
Proposition C / Responsive Floor — Documented non-veto association sequence observed; aggregate Responsive Floor NotAdjudicated. Chapter 9 documents a non-veto pathway in which participation enters a consequential sequence, but the aggregate comparator remains sparse and heterogeneous. The evidence therefore supports a documented association sequence without identifying a general causal effect or licensing an aggregate Responsive Floor classification.
The same empirical chapter strongly exercises the separation Act ≠ Consequence ≠ realization change. Bounded Null, information-only, procedure-only, unresolved, and realization-positive cases all occur. Participatory Readiness remains measurement-limited, and no universal Participation scalar or direct realization-mapping coefficient is identified.
12.6.4 What these statuses mean for the Unified Model
The status architecture is itself part of the model’s discipline. Canonical means that an object or principle has earned its place at the stated logical level and scope; it does not erase the distinction between a definition, a derived rule, and an empirically identified effect. A high-confidence candidate has strong anchoring and discriminability but retains at least one open promotion condition. Provisional means that scope, redundancy, falsifiability, or empirical exercise remains unresolved. Untested means the required antecedent or design has not been exercised; NotAdjudicated means the available evidence is insufficient for the requested aggregate verdict.
This also explains why Chapter 12 does not promote the remaining chapter-level Governance propositions, CFS regimes, Participation tests, or feedback classifications into universal model claims. They retain their local statuses and applicability conditions. A chapter-level finding can support the architecture without becoming a law of the whole system.
The resulting Unified Gradual AGI Model is therefore positive without pretending to be complete. It claims, canonically, that capability and realization must be separated and that stronger inferences require additional evidence. Its frozen architecture also places optimization, governance, participation, observation, and feedback in distinct roles. As high-confidence candidates, it further holds that no canonical universal scalar has been justified for the Unified Gradual AGI system as a whole, and that realized conditions can alter later conditions without requiring a new ontology at every turn. It also records, in the same architecture, which claims are canonical, which remain high-confidence candidates, which are provisional, and which empirical propositions are still untested or not adjudicated.
That mixture is intentional. The model’s openness and flexibility would be empty if every unresolved claim were silently promoted at the capstone stage. Its falsifiability would be empty if failed, untested, or unadjudicated claims disappeared during synthesis. The next section therefore states the inverse boundary: what the Unified Gradual AGI Model does not claim.
12.7 What the Model Does Not Claim
A model can be distorted as easily by what readers infer from it as by what its authors state. The Unified Gradual AGI Model therefore carries explicit nonclaims alongside its positive claims. These are not disclaimers added after the fact. They are part of the architecture: they mark the boundaries beyond which the present definitions, mappings, propositions, and evidence do not license further inference.
The central discipline is simple: the model represents a structured realization process without claiming to predict a single date, threshold, path, optimum, or terminal state for AGI. It separates descriptive, formal-conditional, empirical, and normative work rather than allowing one to inherit the authority of another.
12.7.1 No single number, threshold, date, or universal path
No universal scalar. The model does not assign one number to AGI, realization, Participation, Governance, or the state of AGI-inclusive humanity. Local scalar measures can be useful within bounded tasks, actors, or domains, but they do not become a canonical civilization-scale score.
No universal threshold or arrival date. The realization mapping does not define a single point at which AGI must be declared to have arrived. Gates and thresholds can exist within particular domains or submodels, but they are possible behaviors of a realization relation, not universal model components. The architecture therefore does not supply a calendar forecast or one decisive crossing that settles the system as a whole.
No universal realization path. The model does not posit Potential → Availability → Exposure → domain conditions → Realization as a fixed pipeline. Those objects are analytically distinct, their relations can differ across groups, domains, and time, and later realized states can alter conditions relevant to subsequent realization.
12.7.2 No automatic movement from capability to realization
Capability does not automatically become realization. An increase in Potential creates possibility; it does not entail a corresponding increase in realized state. Availability, Exposure, domain-specific conditions, governance, participation, and other realization-relevant structures can delay, redirect, constrain, enable, or differentiate what follows.
More capability does not guarantee more realization. The architecture does not assume monotonicity. A stronger capability can coexist with uneven adoption, institutional resistance, changing bottlenecks, path dependence, or the foreclosure of other valued realization pathways.
The realization mapping is not an empirically identified universal function. The notation Rg(t) = μt(P(t), Av(t), Exg(t), Zd(t)) specifies a structure of inquiry. It does not, by notation alone, identify one stable functional form across domains or time, establish additivity, or supply counterfactual causal effects. Formal representation is not empirical identification.
12.7.3 No inevitable endpoint, equilibrium, or disappearing human role
No terminal equilibrium or final achieved AGI state. The frozen architecture rejects a built-in terminal condition. A domain can reach a temporary, local, or analysis-relative closure without establishing that AGI-inclusive humanity has entered a final solved state. Ongoing adaptation, contestation, revision, and feedback remain analytically possible.
No claim of inevitable convergence. Feedback can amplify, damp, stabilize, redirect, or delay. Recursion is a temporal relation, not an inevitability engine. SYN-010 Adaptive Disequilibrium remains provisional because its stated defeat condition—a domain reaching an absorbing equilibrium in which relevant later disturbances no longer alter realized state—has not been exercised. Treating the principle as settled before that condition is tested would itself violate the model’s falsifiability discipline.
No automatic disappearance of humans. Reciprocal Participation allows human and AI roles to change, redistribute, and become mutually conditioning. It does not predict that human subjects, judgment, verification, governance, or authority must vanish. Proposition B on Role Plasticity remains Untested in Frame One, so the framework does not convert contemporary automation into a general law of human replacement.
Recursive realization is not recursive self-improvement. Earlier realization can alter later capability or other realization conditions without establishing autonomous RSI. Stronger RSI claims require repeated evidence that successive improvement cycles increasingly depend on AI-generated technical decisions rather than humans supplying the decisive selection, validation, and integration steps.
12.7.4 No automatic normative prescription
Realization is not desirability. The realization mapping describes how capability and realization conditions become socially realized; it does not tell society which realized outcome should be preferred. Optimization evaluates candidate trajectories only under specified actors, objectives, constraints, horizons, and assumptions.
Gradual does not mean slow, and gradualness does not create a duty to slow down. Phenomenal gradualness is a descriptive criterion: realization is gradual when the relevant transformation is not exhausted by one event. It does not require slow, smooth, continuous, scalar, homogeneous, or small-step change. Principled Graduality adds information-responsive iteration and preserved revisability, but even that does not license an automatic moral command to preserve every option or proceed at a particular pace.
The is–ought firewall therefore remains intact. Description states what is happening. Formal analysis states what follows under stated assumptions. Normative evaluation states what ought to count in choosing among alternatives. Evidence can inform normative judgment, but neither a mathematical representation nor the empirical fact of gradual change can convert a value commitment into an empirical conclusion.
12.7.5 Why the nonclaims matter
These limits do not weaken the Unified Gradual AGI Model. They define its usable scope. A framework that claimed a universal score, a universal arrival threshold, an inevitable path, a final equilibrium, or a built-in normative answer would be simpler to summarize but harder to defend. The present architecture instead keeps heterogeneous objects heterogeneous and requires stronger claims to earn stronger evidence.
The result is a model that can say something substantive without pretending to say everything. It can represent realization, distinguish its components, trace pathways, compare trajectories, record contestation and participation, impose epistemic burdens, and follow feedback through time. It cannot, without additional evidence or argument, turn those tools into a date for AGI, a universal score of progress, a forecast of human disappearance, proof of autonomous recursive self-improvement, or a moral instruction for society.
That boundary leads directly to the next question: which parts of the model remain open, what evidence would change them, and where the architecture is still deliberately incomplete.
12.8 What Remains Open and What Would Change It
The Unified Gradual AGI Model is not complete in the sense of having every boundary calibrated, every proposition exercised, or every realization mechanism identified. That incompleteness is deliberate and recorded. An open item can mark an unresolved scope, a missing empirical antecedent, an uncalibrated construct, a non-operable test, an unidentified causal mechanism, or a proposition whose defeat condition has not yet been exercised. Those states are not interchangeable, and Chapter 12 does not convert them into one generic category of uncertainty.
The governing repair rule remains the one established earlier in the book: adverse contact should localize the problem, revise only what the evidence requires, propagate the consequences to dependent claims, preserve the historical record, and expose the revised object to renewed scrutiny. Repair can contract the model as well as extend it. A distinction that adds no discriminating value should be merged or omitted rather than protected by additional machinery.
12.8.1 Architecture-level items that remain open
The exact scope of domain-specific realization conditions remains open. The model retains Zd(t) as domain-specific realization conditions without claiming an exhaustive universal list. A proposed condition earns a place only when it matters to a specified realization object, is distinguishable from Potential, Availability, and Exposure, preserves information that would otherwise be lost, and survives comparison with a simpler representation. If a candidate condition is merely post-hoc, empirically indistinguishable from another argument, or adds no explanatory or predictive resolution, the case for retaining it weakens; it would ordinarily be merged, omitted, or left unclaimed unless further evidence establishes discriminating value.
The Availability / supporting-resource-stack boundary remains narrowed but not closed. Availability remains a first-class realization argument, while the broader resource stack remains analytically distinct and some resource conditions may instead belong in Zd(t). The boundary should be settled by causal role rather than by convenience or physical identity. Evidence showing that a proposed resource variable behaves consistently as practical obtainability, rather than as a domain condition or another construct, would sharpen this boundary.
Several architecture items are deliberately non-blocking. A domain-general transformation-individuation rule I is not claimed; source-relative individuation is sufficient at present. Standing derivability—whether the Standing facet can be reconstructed from Plateau plus other retained metadata without analytically relevant information loss—remains an open MPF facet question. The exact synchronization-coupling symbol need not be frozen prematurely, and no canonical book term has yet been fixed for a single ‘AGI maturity’ or ‘realized AGI condition.’ These are real open items, but none licenses a new substantive construct merely to achieve terminological closure.
Synthetic-principle promotion remains incomplete. SYN-001, SYN-004, SYN-005, and SYN-006 remain High-confidence candidates because their full displayed SPAT walk-throughs are still open. They should change status only by satisfying the inherited admission criteria, not because the book has reached its final chapter. SYN-010 Adaptive Disequilibrium remains Provisional; an absorbing domain equilibrium in which relevant later disturbances no longer alter the realized state would defeat or narrow its present formulation.
One item should not be reopened: the post-catalog #9–#10 Graduality extension check is already Resolved / Pass. It found no definitional amendment required. That is a compatibility result, not a claim that Graduality has been empirically validated across every installment.
12.8.2 Realization and synchronization: identification, parsimony, and longitudinal burden
The realization relation is structurally specified but not universally identified. Future work must distinguish which gates bind, which thresholds are causal, which delays are structural, and which competing realization mechanisms survive counterfactual analysis. A richer time-varying mapping is not warranted if a simpler stable or nested representation organizes and predicts the specified domain equally well. If competing mechanisms cannot be distinguished by any feasible observation, the mapping claim is too unconstrained for that application.
Exposure must continue to earn first-class status. Reach, Susceptibility, and Recourse are retained because they preserve analytically distinct group-level conditions. If they add no explanatory or discriminating value across the domains in which Exposure is invoked, first-class Exposure would be redundant. Conversely, a future governance intervention that cannot be classified under Access, Sequencing, Distribution, or Deployment Context would trigger the existing completeness falsifier for the four points of purchase.
CFS remains bounded by four explicit open burdens. First, the joint long-run behavior of Ceiling growth, coupling, and any residual gap remains unidentified; the limiting coupling rate alone does not determine whether a permanent residual exists or what such a residual would mean. Second, the predicted overtrust regimes have not produced a clean multi-period Regime 4 or 5 trace in the principal empirical domains. Third, endogenous gate oscillation is not solved by baseline CFS. Fourth, the displayed gap-dependent power-law form is illustrative rather than uniquely identified; alternative monotonic functional forms remain viable until denser longitudinal evidence discriminates among them.
The CFS stop rule therefore remains active. A realization gap does not confirm CFS. If Ceiling, Floor, longitudinal measurement, or plausible coupling cannot be specified without material ambiguity, the correct result remains: CFS not presently applicable.
12.8.3 Optimization and Governance: unresolved choice and authority conditions
Optimization retains several open formal and empirical conditions. Long-horizon direction under genuinely plural objectives remains an open normative/formal problem. Common Pareto analysis requires enough shared outcome structure to be meaningful, and standard Nash Bargaining Solution use requires defensible utility assumptions. Dynamic inconsistency remains only partly formalized. The Cohesion/Dynamism construction remains proof of concept rather than a universal aggregator.
Absorption Capacity remains structurally useful but uncalibrated. The retained sign relation does not provide a universal unit or threshold. A future operationalization would require a domain-specific exposure/capacity variable defined before outcome observation, repeated comparable disturbances, and an outcome measure capable of distinguishing greater absorption from merely lower exposure. Nor should Absorption Capacity be identified with the CFS Floor, societal readiness, or Participation readiness.
Barrier crossing remains unresolved. Rapid capability growth or a dramatic month does not establish a critical transition. The existing annual series cannot estimate the longitudinal indicators needed to adjudicate a metastable barrier-crossing claim. A future claim requires transition evidence rather than narrative intensity. As Chapter 6 §6.7.7 records, the initial content-licensing prediction remains failed; the revised prediction remains live and must be tested prospectively rather than used to rewrite the original result.
Governance remains empirically uneven. Susceptibility is named inside Exposure but remains unmeasured and uncalibrated. GP2 is executable but its growth antecedent was not exercised; GP7 is presently non-operable because an adequate independent benchmark is missing. GP1 lacks the necessary coupling trajectory, GP5 lacks general rate estimates, and the strongest form of GP3 cannot be finitely settled. These are evidence boundaries, not invitations to substitute contemporary examples for the missing design.
The Governance framework also makes no prevalence claim. Establishing how common particular governance configurations are would require a different sampling design. Likewise, the framework does not become a complete theory of political legitimacy merely because it can represent authority, standing, recourse, contestability, and procedural adjudication.
12.8.4 Participation: what prospective evidence must still do
Conditional Non-Fungibility requires prospective independent replication. The executed coding produced discrimination/nonredundancy evidence for RO, FP, and AD conditional on faithful application, but intersubjective reproducibility cannot be reconstructed from the surviving archive. The result can be strengthened only by a new design that preserves coder provenance and permits independent reproduction.
Role Plasticity remains Untested in Frame One. The required automation antecedent was absent. A future adjudication must observe a task sequence in which human and AI functions can actually redistribute over time. The reciprocal human-AI form is likewise prospective because Frame One did not independently code the full reciprocal configuration.
The aggregate Responsive Floor remains NotAdjudicated. The documented non-veto sequence does not by itself supply the comparator structure needed for an aggregate floor classification. Stronger evidence requires sufficiently comparable cases, pre-specified realization conditions, and a design capable of separating a documented pathway from causal contribution. Participatory Readiness remains measurement-limited, Contrast B remains unexercised, causal contribution remains unidentified, and temporal institutionalization remains prospective.
These burdens can produce different outcomes. A future study may support a proposition, leave it untested, show the instrument to be non-operable, contract a causal claim to pathway evidence, or demonstrate that a proposed scalar or direct realization coefficient remains unwarranted. The model should preserve those differences rather than force them into a binary pass/fail result.
12.8.5 Feedback, recursion, and terminality: stronger claims require stronger temporal evidence
Sequence remains insufficient for feedback. A credible feedback claim must identify an initiating realization, observation, or consequence; a response mechanism; a condition changed through that response; and a subsequent realization or action under the changed condition. Recursive re-entry requires the changed condition to enter another realization cycle, and repeated recursion requires more than one return.
Structural feedback carries a higher burden than ordinary feedback. A change in Potential, Availability, Exposure, or domain-specific conditions should be presumed sufficient unless evidence shows that the conversion relationship itself changed. Because μt is not identified as an invariant empirical function, “the relation changed” is not operationally licensed merely by observing different outcomes at different times. A structural-feedback claim requires otherwise comparable conditions plus evidence that altered institutional, technical, or procedural relationships changed the realization relation in a way not adequately represented as changed values of P(t), Av(t), Exg(t), or Zd(t). Where that distinction cannot be made, the μt-change claim remains a regulative caution rather than an adjudicated finding.
Recursive self-improvement remains a narrower empirical claim. AI-assisted AI research can constitute capability feedback without establishing autonomous RSI. Stronger evidence would require repeated improvement cycles in which AI systems materially contribute to identifying, designing or implementing, validating, and reusing capability improvements, with human direction no longer supplying the key technical decisions at every step.
RSI stop rule. Do not classify a sequence as RSI unless repeated improvement cycles show AI materially contributing to identifying improvement opportunities, designing or implementing changes, validating results, and reusing the improved capability in a subsequent improvement cycle, while human direction no longer supplies the key technical decisions at every step. If any of those conditions is unobserved or materially unresolved, stop at the strongest supported lower classification—capability feedback, recursive re-entry, or repeated recursion.
The Ongoing / Unfinished diagnostic is also defeasible at the domain level. A domain is ongoing insofar as consequential outcomes continue to alter later realization conditions, and unfinished insofar as no empirically justified terminal configuration has been established for the specified domain and observation horizon. A sufficiently stable domain configuration in which further AI-mediated change no longer materially altered the realization conditions under examination would change that local diagnosis. It would not, by itself, establish a terminal state for AGI-inclusive humanity as a whole.
12.8.6 What would count as model change
The model should change when evidence defeats a distinction, identifies a simpler adequate representation, shows that a proposed mechanism is redundant, exercises an untested antecedent, makes a previously non-operable comparison possible, identifies a causal relation at the required burden, or demonstrates a stable boundary where the current theory predicts continued change. Different findings would require different repairs. Some would recalibrate a parameter or operationalization; some would contract a proposition; some would merge categories; some would retire a construct; and some could require revising the realization architecture itself.
Architecture-level defeat requires more than the failure of one local proposition. Revision of the architecture itself would be required if repeated empirical application defeated one or more load-bearing separations rather than merely refining a component. Candidate defeat conditions include: (1) robust cross-domain evidence that separating Potential from Realization adds no discriminating or explanatory value over capability alone; (2) repeated evidence that Availability, Exposure, and domain-specific realization conditions add no stable classificatory or explanatory value once capability is specified; or (3) systematic empirical collapse of the placement distinctions among Governance, Participation, Observation/Evidence, and Feedback/Recursion such that their roles cannot be maintained without ad hoc reclassification. These are architecture-level burdens: if exercised, they would require redesign of the organizing relation, not merely a status change inside one subsystem.
Worked repair already appears in the book’s own record. Chapter 6 preserves an initial content-licensing prediction as failed rather than rewriting it into success, while the present chapter revised its interpretation of the nine-loop physics episode after concurrent Song He-group provenance became available: the new evidence contracted the uniqueness of the pathway without changing the RSI stop-rule classification. In both cases, revision preserves the earlier evidentiary history and changes only what the new evidence warrants.
What should not change the model is equally important. A new headline case, an apparently dramatic month, a more powerful system, a successful demonstration, or a fashionable label does not by itself license promotion. Nor does an unresolved result justify adding a variable solely to preserve the preferred story. The burden remains discriminating evidence, versioned claims, explicit status, and repair without rescue.
The architecture is therefore open in a precise sense: it specifies where change is permitted, what evidence would force change, and which questions are still not answered. That prepares the final section for a more limited conclusion. The book can characterize Gradual AGI as an Ongoing, Unfinished Transformation without turning that orientation into a claim that every domain is permanently unsettled or that the transformation has one predetermined end.
12.9 An Ongoing, Unfinished Transformation
The Unified Gradual AGI Model ends where its architecture requires it to end: not with a declaration that AGI has arrived, not with a date by which it must arrive, and not with a terminal state into which the model expects society to converge. Its final claim is narrower. Capability, realization, governance, participation, evidence, and feedback must be analyzed as distinct but connected parts of an evolving human–AI system whose realized conditions can become conditions of what happens next.
12.9.1 Two meanings that must remain distinct
Architectural orientation. Chapter 3 uses Ongoing, Unfinished Transformation as methodological shorthand for theory-building. It describes an evolving transformation whose realization is not exhausted by one event, is not presumed complete at a particular capability threshold, and whose relevant forms, relationships, and consequences remain open to further change. In that register, the phrase is not an empirical claim about the pace or historical state of any particular transformation.
Domain-level diagnostic. Chapter 11 gives the same phrase a narrower empirical meaning. A specified domain is ongoing insofar as consequential outcomes continue to alter later realization conditions. It is unfinished insofar as no empirically justified terminal configuration has been established for that domain and observation horizon. Ongoing is therefore a process claim; unfinished is a closure-status claim.
Those two registers should not be collapsed. The Chapter 3 orientation explains why the model remains revisable. The Chapter 11 diagnostic explains what evidence would justify describing a particular domain as dynamically ongoing or not yet terminal. A domain can stabilize without falsifying the architectural orientation, and the architectural orientation does not entitle the book to pronounce every domain permanently unsettled.
12.9.2 What the unified architecture adds
The contribution of the Unified Gradual AGI Model is not a new master equation layered above the preceding chapters. Unification is architectural, not algebraic. The realization mapping remains the central bridge between what AI makes possible and what becomes socially realized. Graduality identifies the relevant form of transformation without equating gradual with slow. CFS supplies one bounded family of synchronization dynamics. Optimization asks which trajectories are preferred under specified objectives and constraints. Governance locates authority, contestation, observation, steering, and adjudication around realization. Participation distinguishes subjects, acts, and consequences. Observation and Evidence constrain what the model is entitled to infer. Feedback and Recursion show how prior realization can condition later realization.
The resulting system is deliberately heterogeneous. No single scalar replaces these objects. No subsystem absorbs the others. No forward arrow is automatic. Capability does not entail realization; institutional possibility does not entail participatory action; action does not entail consequence; consequence does not entail realization change; sequence does not entail feedback; feedback does not entail recursion; and recursion does not entail recursive self-improvement.
What binds the model together is therefore not reduction to one quantity but disciplined relation among different kinds of claim. The MPF keeps definitions, formal representations, propositions, empirical findings, and synthetic principles from being silently promoted across logical levels. The status ledger keeps canonical, high-confidence, provisional, untested, measurement-limited, non-operable, and NotAdjudicated results visibly distinct. The open-item ledger keeps incompleteness from being hidden at the moment of synthesis.
12.9.3 The model’s final posture
Taken together, the chapters support a model of AGI-inclusive humanity in which humans, AI systems, institutions, infrastructure, and the wider social environment co-evolve through uneven realization. Capability advances matter, but they enter domains with different availability, exposure, institutional conditions, optimization problems, governance arrangements, participatory configurations, observational limits, and feedback histories. The same capability can therefore become realized differently across groups, places, institutions, and time.
That heterogeneity is not a defect to be removed. It is the object the theory is trying to preserve. A framework that erased it would be easier to summarize but less capable of distinguishing possibility from realization, authority from preference, participation from influence, observation from inference, and historical sequence from feedback.
The book therefore closes with a conditional rather than terminal formulation. Where realized consequences continue to alter later realization conditions, the transformation is ongoing. Where no empirically justified terminal configuration has been established for the specified domain and observation horizon, it is unfinished. Where either condition ceases to hold, the corresponding diagnostic should change.
At the level of the book as a theory-building project, Ongoing, Unfinished Transformation remains the more general orientation: the architecture is open to new evidence, flexible enough to permit local repair, and falsifiable enough to require contraction or retirement when claims fail. Openness makes revision visible; flexibility permits it; falsifiability disciplines it.
The Unified Gradual AGI Model is therefore best understood not as a prediction of one future state, but as a structured way to ask what is becoming possible, what is becoming realized, for whom, under what conditions, through which institutions and acts, on what evidence, and with what consequences for what comes next.
That is the sense in which Gradual AGI is treated here as an Ongoing, Unfinished Transformation: not as a claim that history has one inevitable direction, but as a commitment to keep capability, realization, human participation, institutional change, and epistemic discipline inside the same revisable model.
Chapter References
Book and Unified Gradual AGI Sources
W.H.L. & Claude. 2025. “First Principles of AGI-Inclusive Humanity.” Champaign Magazine, June 17, 2025. https://champaignmagazine.com/2025/06/17/first-principles-of-agi-inclusive-humanity/
W.H.L. & ChatGPT. 2026. “Gradual AGI as Epistemic Extension.” Gradual AGI Series #1. Champaign Magazine, January 19, 2026. https://champaignmagazine.com/2026/01/19/gradual-agi-as-epistemic-extension/
W.H.L. & GPT-5.5. 2026. “Gradual AGI as Abundant Resources.” Gradual AGI Series #2. Champaign Magazine, June 11, 2026. https://champaignmagazine.com/2026/06/11/gradual-agi-as-abundant-resources/
W.H.L. & GPT-5.5. 2026. “Gradual AGI as Synchronization for Transformative Adoption.” Gradual AGI Series #3. Champaign Magazine, June 29, 2026. https://champaignmagazine.com/2026/06/29/gradual-agi-as-synchronization-for-transformative-adoption/
W.H.L. & Claude (Sonnet 5). 2026. “Ceiling, Floor, and Slope: A Falsifiable Dynamical Model of Synchronization for Gradual AGI.” Gradual AGI Series #4. Champaign Magazine, July 4, 2026. https://champaignmagazine.com/2026/07/04/ceiling-floor-and-slope-a-falsifiable-dynamical-model-of-synchronization-for-gradual-agi/
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. https://champaignmagazine.com/2026/07/22/gradual-agi-as-optimization-a-conceptual-framework/
W.H.L. & Claude (Sonnet 5, Opus 5). 2026. “Gradual AGI as Optimization: Formal Models and Empirical Tests.” Gradual AGI Series #6. Champaign Magazine, July 24, 2026. https://champaignmagazine.com/2026/07/24/gradual-agi-as-optimization-formal-models-and-empirical-tests/
W.H.L. & Claude (Opus 5). 2026. “Gradual AGI as Contestation: A Framework for Governance.” Gradual AGI Series #7. Champaign Magazine, August 3, 2026. https://champaignmagazine.com/2026/08/03/gradual-agi-as-contestation-a-framework-for-governance/
W.H.L., Claude Opus 5, & GPT-5.6 Sol. 2026. “Gradual AGI as Contestation: Measuring Governance Under Empirical Contact.” Gradual AGI Series #8. Champaign Magazine, August 11, 2026. https://champaignmagazine.com/2026/08/11/gradual-agi-as-contestation-measuring-governance-under-empirical-contact/
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. https://champaignmagazine.com/2026/08/27/gradual-agi-as-participation-subjects-acts-and-consequences/
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. https://champaignmagazine.com/2026/08/30/gradual-agi-as-participation-3m-taxonomy-relational-grammar-and-responsive-floor-test/
Contemporary External Sources Used in Chapter 12
Abrishami, T. (2026, September 18). In August, 25% of math preprints acknowledged AI use, up from 4% in April. Epoch AI. https://epoch.ai/data-insights/math-preprints-disclosed-ai-use
Alibaba DAMO Academy. (2026). RADAR: An expert-level generalist AI for abdominal CT diagnosis [Code and model repository]. https://github.com/alibaba-damo-academy/damo-radar
Altman, S. (2026, September 22). Post on AI standards, stakeholder voice, and concentration of power [Post]. X. https://x.com/sama/status/2102414347364335917
Anthropic. (2026, September 25). Yes, Claude can do Nine Loops. Guest post by Matt von Hippel, with an addendum by Lance J. Dixon. https://www.anthropic.com/research/yes-claude-can-do-nine-loops
Choi, S., Kang, H., Kim, N., & Kim, J. (2025a). How does artificial intelligence improve human decision-making? Evidence from the AI-powered Go program. Strategic Management Journal. https://doi.org/10.1002/smj.3694
Choi, S., Kang, H., Kim, N., & Kim, J. (2025b). The dual edges of AI: Advancing knowledge while reducing diversity. PNAS Nexus, 4(5), pgaf138. https://doi.org/10.1093/pnasnexus/pgaf138
Clay Mathematics Institute. (2026). The Millennium Prize Problems. https://www.claymath.org/millennium-problems/
Dixon, L. J., & Liu, Y.-T. (2023). An eight loop amplitude via antipodal duality. Journal of High Energy Physics, 2023, 98. https://doi.org/10.1007/JHEP09(2023)098
Google DeepMind. (2026, March 10). From games to biology and beyond: 10 years of AlphaGo’s impact. https://deepmind.google/blog/10-years-of-alphago/
Hacker News. (2026, September 23). Building standards for the next phase of AI [Discussion thread]. https://news.ycombinator.com/item?id=49802825
Lange, J. (2026, September 22). Three out of four Americans say AI firms not doing enough to prevent disaster, Reuters/Ipsos poll finds. Reuters. https://www.reuters.com/world/three-out-four-americans-say-ai-firms-not-doing-enough-prevent-disaster-2026-09-22/
Leiden Declaration on Artificial Intelligence and Mathematics. (2026, June 2). DOI 10.5281/zenodo.20302944. https://leidendeclaration.ai/
Mathathon Organizers. (2026, September 10). A response from Mathathon. Proofs and Prompts. https://proofsandprompts.com/2026/09/10/a-response-from-mathathon/
OpenAI. (2026a, September 8). On the Navier–Stokes Millennium Prize Problem. https://openai.com/index/navier-stokes-solution/
OpenAI. (2026b, September 21). Building standards for the next phase of AI. https://openai.com/index/building-standards-next-phase-ai/
OpenAI. (2026c, September 22). Priorities and principles for effective third party assessments. https://openai.com/index/priorities-principles-third-party-assessments/
SAIR Foundation. (2026). Building open models for the mathematical community. https://sair.foundation/open-math-model/
Shin, M., Kim, J., van Opheusden, B., & Griffiths, T. L. (2023). Superhuman artificial intelligence can improve human decision-making by increasing novelty. Proceedings of the National Academy of Sciences, 120(12), e2214840120. https://doi.org/10.1073/pnas.2214840120
Tao, T. (2026, September 18). SAIR’s Open Math Model initiative. What’s new. https://terrytao.wordpress.com/2026/09/18/sairs-open-math-model-initiative/
von Hippel, M. (2026, September 25). It Got to My Field. 4 gravitons. https://4gravitons.com/2026/09/25/it-got-to-my-field/
Woit, P. (2026, September 25). Some HEP-TH News. Not Even Wrong. https://www.math.columbia.edu/~woit/wordpress/?p=15881
Zhang, Q., Zhang, J., Cao, W., Lu, Z., Chang, W., et al. (2026). An expert-level generalist AI for abdominal CT diagnosis. Science, 393(6817). https://doi.org/10.1126/science.aec6129
Leave a Reply