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

An Open Architecture for an Ongoing, Unfinished Transformation

Chapter 3 of the forthcoming book On Gradual AGI, based on Champaign Magazine’s Gradual AGI series installments

Publication Version: v1.0. September 6, 2026

Abstract

This chapter develops the Multi-Plateau Framework (MPF) as the methodological-theoretical architecture of the Unified Gradual AGI Model. Starting from Gradual AGI as an Ongoing, Unfinished Transformation, it asks how a theory can remain open to new phenomena, revisable under error, and genuinely exposed to failure without becoming indefinitely elastic. The MPF distinguishes five logical plateaus, cross-cutting facets, and typed relations; it also provides a minimal application mode so that ordinary use need not reproduce the full technical register. Its principal methodological contribution is epistemic repairability: the capacity to detect error, localize failure, revise the affected structure, propagate the consequences of revision, and expose the repaired model to renewed scrutiny. A documented Graduality repair shows the seven-step sequence in practice. The chapter then develops version-specific validation and an explicit anti-rescue boundary, positions the MPF relative to established accounts of scientific change, and maps the architecture onto the Unified Gradual AGI Model without re-deriving Chapter 4’s substantive realization model. The framework is offered as provisional architecture rather than immutable metatheory.

Keywords: Gradual AGI; Unified Gradual AGI Model; Multi-Plateau Framework; epistemic repairability; falsifiability; theory revision; realization; participation

Scope note. This manuscript is Chapter 3 of the forthcoming book On Gradual AGI. Chapter 2 owns the definitions of gradualness, transformation individuation, and the Graduality catalog; Chapter 4 owns the substantive realization architecture and its principal formal variables. This chapter intentionally develops the logical architecture between them without re-deriving either neighboring chapter.

Reader Guide — What This Chapter Does

The chapter solves one problem: how to organize a theory of an evolving object without collapsing definitions, measurements, propositions, evidence, and system-level interpretations into one logical type. By the end, a reader should be able to assign plateau by logical role, retain only analytically necessary facets, type relations strongly enough to block invalid inferences, distinguish failure from limited empirical contact, and revise a failed component while preserving versions and dependencies. The chapter does not supply the substantive realization equation or close the Unified Model’s open empirical questions; those belong to Chapter 4 and later domain chapters.

3.1 Gradual AGI as an Ongoing, Unfinished Transformation

Gradual AGI begins from a simple but consequential premise: the transformation associated with increasingly general artificial intelligence is still in the making. It should therefore not be modeled as though its final form, boundaries, consequences, or organizing concepts were already known.

Building on Chapter 2’s account of gradual realization, this chapter uses Ongoing, Unfinished Transformation to describe an evolving transformation whose realization is not exhausted by a single event, is not presumed complete at a particular capability threshold, and whose relevant forms, relationships, and consequences remain subject to further change. The phrase is used here as architectural shorthand rather than as a new definition of phenomenal gradualness. It is a methodological orientation for theory-building, not an empirical claim about the pace or historical state of any particular transformation.

This does not make capability thresholds unimportant. A system may cross a threshold that is scientifically significant, economically disruptive, or historically transformative. New capabilities may alter what machines can do, what institutions can delegate, what organizations can produce, or what people can reasonably expect from artificial intelligence. But a capability threshold does not by itself establish that the resulting transformation has been realized across societies, institutions, domains, or populations. Nor does it establish that the meaning of the transformation has ceased to change.

The distinction matters because a theory built around a completed object can assume a degree of closure that an evolving object does not permit. Gradual AGI instead concerns an Ongoing, Unfinished Transformation whose relevant technologies, institutions, practices, dependencies, conflicts, and consequences may continue to develop while the theory itself is being constructed. Some objects that later prove important may not yet be visible. Some distinctions that presently appear useful may later prove inadequate. Some propositions may survive testing; others may fail. Measurements may improve. Relationships among components may change. Entire domains may become relevant that the present model does not yet represent.

The Unified Gradual AGI Model therefore requires an architecture that can support inquiry while its object continues to change. Three requirements follow. The architecture must be open enough to encounter phenomena that its present form did not anticipate. It must be flexible enough for particular components to be revised, replaced, extended, merged, or retired without automatically invalidating every other component. And it must be falsifiable enough that its claims can genuinely fail rather than being preserved through reinterpretation after contrary evidence appears.

These requirements are most useful when understood together rather than as independent virtues. Openness without falsifiability can become indefinite accommodation: every new observation is simply absorbed. Flexibility without falsifiability can become revision without accountability: the framework changes, but nothing can establish that the previous formulation was wrong. Openness without flexibility can recognize phenomena the existing architecture cannot accommodate, yet have no way to respond except by forcing them into categories that no longer fit.

Openness, flexibility, and falsifiability jointly support epistemic repairability.

Epistemic repairability is the capacity of a framework to detect error, localize what failed, revise the affected structure, propagate the consequences of that revision, and expose the repaired model to renewed scrutiny. The fuller repair protocol developed later in this chapter begins with exposure to test and includes identification of the dependencies affected by a failure. Repair is therefore more demanding than the mere ability to revise a theory after encountering difficulty.

This requirement changes what robustness means. A robust framework is not one whose components are constructed so broadly that they can never be contradicted. Nor is it one that treats every failure as evidence that the entire theoretical enterprise must be discarded. Its strength lies instead in making failure informative.

That distinction is especially important for Gradual AGI because different kinds of failure need not have the same implications. An empirical result may fail without invalidating the concept that motivated its measurement. A measurement may prove inadequate while the underlying theoretical distinction remains useful. A proposition may be rejected while the objects it relates continue to exist. A formal representation may need revision without requiring a change in the ontology it was intended to represent. Conversely, an apparently successful empirical result cannot rescue a poorly specified concept simply because the result is convenient.

A theory of an evolving transformation therefore needs enough internal differentiation to identify where a problem resides. Without that differentiation, disagreement about the theory can easily collapse distinct questions into one another: whether something exists, what a term means, how an object is represented or measured, whether a proposition concerning it is supported, and what its significance is for the larger system.

The Multi-Plateau Framework is introduced to preserve those distinctions.

It is not another substantive mechanism of Gradual AGI. It does not sit alongside Synchronization, Optimization, Governance, Participation, or other realization processes as an additional causal component. It is the methodological-theoretical architecture through which objects, concepts, formalizations, propositions, evidence, and system-level interpretations within the Unified Model can be distinguished and related without being flattened into a single logical type.

Its purpose is therefore simultaneously representational, critical, and corrective. It helps specify what kind of theoretical object is being discussed; it makes claims sufficiently localizable to be exposed to failure; and it provides an architecture within which failure can lead to disciplined revision rather than concealment or indiscriminate collapse.

That corrective function has an essential boundary: repair must not become rescue. A framework cannot claim self-correction if revision erases the fact that an earlier claim failed, changes the governing test after the result is known, or silently promotes a replacement formulation as though it had always been the theory. The chapter therefore treats versioning, dependency tracking, and renewed testing as disciplines governing repair rather than as additional design criteria. The full anti-rescue boundary is specified later under Repair Without Rescue.

Positioning note. The MPF does not claim novelty for falsification, theory change, or model revision as such. It shares Popper’s insistence that scientific claims be exposed to possible refutation; Kuhn’s attention to historically situated conceptual change; Lakatos’s concern with continuity across revisable theory sequences; Laudan’s emphasis on comparative problem-solving; and model-based accounts in which models mediate between theory and world and are themselves objects of construction, testing, and revision (Popper 1959; Kuhn 1962; Lakatos 1970; Laudan 1977; Morgan and Morrison 1999). Its narrower contribution is architectural: it makes logical type, version, provenance, epistemic status, typed dependency, and repair propagation explicit within one framework for a still-changing socio-technical object. It therefore does not map the MPF onto a Kuhnian paradigm or a Lakatosian hard-core/protective-belt distinction; those traditions are antecedents for the problem of theory change, not templates silently imported into the present architecture.

The Multi-Plateau Framework is consequently provisional in the same sense that the larger transformation is unfinished. Its present architecture is not presumed to exhaust the possible logical organization of Gradual AGI. If important objects repeatedly resist representation, if distinctions prove redundant, if failures cannot be localized, or if repair procedures begin to immunize claims from genuine falsification, the framework itself must be reconsidered.

The framework is not designed never to break. It is designed to break informatively—and to be repairable when it does.

Figure 3.1 provides a reader-level overview of the Multi-Plateau Framework as a whole. The sections that follow unpack its principal components in sequence: the five plateaus (§§3.2–3.3), facets (§3.4), typed relations (§3.5), falsifiability and repair (§§3.6–3.8), and its mapping to the Unified Gradual AGI Model (§3.9).

Figure 3.1. Reader overview of the Multi-Plateau Framework (MPF) of the Unified Gradual AGI Model. The figure presents the chapter’s architecture at a glance: AGI-inclusive humanity as the highest system object, realization as the central transformational bridge, the five logical plateaus, major cross-cutting components, and the framework’s non-sequential organization.

The next section begins with the framework’s central organizing term: plateau.

3.2 Why Plateaus: Logical Vantage Points Without Stages

If Gradual AGI is an Ongoing, Unfinished Transformation, its architecture must permit stable analysis without implying that the transformation itself has become stable, complete, or sequentially ordered.

The Multi-Plateau Framework addresses that problem by distinguishing several logical vantage points from which different kinds of claims can be made about the same evolving system.

A plateau is a relatively stable logical vantage point from which one kind of claim about an evolving system can be specified, examined, revised, or tested without implying a chronological stage, hierarchical rank, or terminal state.

The qualification relatively stable is essential. A plateau provides enough stability to support serious analytical work: concepts can be defined, propositions stated, formalizations inspected, evidence evaluated, and relationships traced. But the stability belongs to the vantage point required for inquiry, not to the transformation as a whole. The objects observed from that vantage point may continue to change, and the framework used to describe them may itself require revision.

The term plateau is deliberate but narrow. A stage suggests chronological succession; a level suggests rank; and a layer suggests fixed stacking or predetermined dependence. Plateau is used only to indicate a locally stable analytical vantage point within a larger landscape that remains open and changing. It carries no further topographic claim: there is no summit encoded in the framework.

The term also has an intellectual genealogy relevant to this choice. In “Bali: The Value System of a Steady State,” Gregory Bateson used plateau to describe a sustained intensity that substitutes for cumulative escalation toward climax: the point is maintenance without culmination rather than progression toward a terminal peak (Bateson 1949). Deleuze and Guattari explicitly borrowed Bateson’s term in A Thousand Plateaus and generalized it as a continuous region of intensities whose development is not organized toward a culmination point or external end; they emphasize that a plateau is “in the middle,” not a beginning or an end (Deleuze and Guattari 1987, 22–24). The Multi-Plateau Framework inherits only this anti-terminal intuition. It does not import Bateson’s anthropological account of intensity or Deleuze and Guattari’s broader rhizomatic ontology; here plateau has the narrower technical meaning defined above: a relatively stable logical vantage point.

The five plateaus do not constitute a developmental sequence, a maturity scale, or a process pipeline. Work need not move from Plateau P1 through Plateau P5 in order. An empirical finding at Plateau P4 may expose a problem in a Plateau P3 operationalization. A new phenomenon may require revision of a Plateau P2 concept. A system-level interpretation at Plateau P5 may reveal that two apparently distinct objects at Plateau P1 have been improperly separated. Inquiry can move across the plateaus in multiple directions.

This non-sequential structure is important for epistemic repairability. If the plateaus were treated as stages, failure at a later point might be interpreted as evidence that the entire preceding sequence had failed. The Multi-Plateau Framework instead asks a more discriminating question: what kind of object or claim has failed, and what else actually depends on it?

That question requires logical differentiation.

The present architecture distinguishes five plateaus:

• Plateau P1 — Ontological. What exists or occurs.
• Plateau P2 — Conceptual-Semantic. What theoretical terms, categories, and distinctions are intended to capture conceptually.
• Plateau P3 — Formal-Operational. How objects and relations are represented, formalized, classified, measured, or observed.
• Plateau P4 — Propositional-Epistemic. What is proposed, tested, found, rejected, unresolved, untested, non-operable, or unreproduced.
• Plateau P5 — Integrative-Systemic. How components interact across domains, time, feedback, realization, and the larger system.

These distinctions are logical rather than chronological. They separate questions that are often compressed into one another.

Participation illustrates a full traversal across the five plateaus. Exposure illustrates a different pattern: a single construct can be pulled apart at the conceptual, formalization, and system-role levels without being forced into a full staged sequence. A conceptual distinction concerning Exposure is not identical to the formal representation used to operationalize it, and neither is identical to its role within the larger realization system.

The same discipline applies to Participation. A participatory act may exist as an observable occurrence; its meaning must be conceptually specified; its features may then be classified through a formal grammar; propositions concerning those features may be tested; and their implications may finally be examined within the larger dynamics of realization. None of those operations makes the others redundant.

The purpose of the plateau distinction is therefore not to multiply terminology. It is to prevent logical flattening.

This is especially important in a framework expected to evolve. When a theoretical system contains only loosely differentiated “concepts,” a revision can propagate invisibly: a changed definition may silently alter a measurement, a proposition, an earlier empirical interpretation, and a system-level conclusion. The plateaus make those transitions visible enough to inspect.

At the same time, the plateaus should not be treated as sealed compartments. The Unified Model depends on cross-plateau relations. Concepts refer to objects. Formalizations operationalize concepts. Propositions relate formalized or conceptual objects. Evidence changes epistemic status. System-level interpretations integrate results across domains.

The architecture therefore seeks separation without isolation: logical functions remain distinguishable, while genuine dependencies remain connected so that revision can propagate where warranted.

If future work reveals a theoretically important kind of object that cannot be represented coherently within Plateau P1 through Plateau P5, the architecture should be reconsidered rather than the object forced into an unsuitable category. The plateaus are provisional instruments for stable inquiry, not claims that the landscape has stopped changing.

3.3 The Five Plateaus

The five plateaus distinguish different logical roles within the Unified Gradual AGI Model. They are not five subjects, five stages of research, or five degrees of theoretical importance. The same substantive phenomenon may appear across several plateaus because different claims can be made about it.

Plateau assignment follows logical role, not subject matter.

Participation, for example, is not assigned once and for all to a single plateau. A participatory act can be treated as something that occurs, as something requiring definition, as something classified through an operational grammar, as the subject of a testable proposition, or as part of a larger realization process. What changes is not necessarily the phenomenon but the kind of claim being made about it.

This distinction is what allows the architecture to separate objects from definitions, definitions from representations, representations from epistemic results, and local results from system-level interpretation.

3.3.1 Plateau P1 — Ontological

Plateau P1 concerns what exists or occurs.

What is the object, event, condition, relation, actor, or process that the theory treats as part of the world it is trying to explain?

Plateau P1 provides the ontological inventory on which later distinctions can operate. In the Participation framework, a participatory subject, participatory act, and participatory consequence are Plateau P1 objects. The point at this plateau is not yet how each term is formally defined, measured, classified, or evaluated. It is that the theory distinguishes these as different kinds of things or occurrences.

The same distinction applies elsewhere in the Unified Model. Exposure can be treated as a condition that varies across populations or domains. A governance intervention can be treated as an occurrence. A synchronization gap can be treated as a system condition. Their definitions, representations, measurements, and system-level implications belong elsewhere.

Plateau P1 should therefore not be confused with empirical confirmation. To represent an object ontologically is not to establish that every proposed instance exists, that its measurement is valid, or that propositions concerning it are true. Those are separate questions.

This distinction is important for repair. Evidence may undermine a proposition about a Plateau P1 object without requiring the object itself to be eliminated from the ontology. Conversely, persistent failure to identify or coherently distinguish a supposedly fundamental object may eventually require ontological revision.

3.3.2 Plateau P2 — Conceptual-Semantic

Plateau P2 concerns what theoretical terms, categories, and distinctions are intended to capture conceptually.

What does this theoretical object or distinction mean, and what separates it from neighboring concepts?

Plateau P2 gives semantic structure to the ontology. It specifies definitions, conceptual boundaries, distinctions, and relationships of meaning necessary for the model to remain intelligible.

Participation again provides a straightforward example. Once participatory subjects, acts, and consequences are distinguished as objects, their meanings must be specified. A participatory act cannot simply mean “anything people do involving AI” if the theory requires a more discriminating boundary. Likewise, Participation cannot be collapsed into approval, adoption, resistance, use, or sentiment if those concepts perform different analytical functions.

Graduality provides another important Plateau P2 case. The meaning of gradual in Gradual AGI cannot be inferred merely from the passage of clock time or from a sequence of capability benchmarks. Its conceptual definition determines what counts as graduality within the theory before particular propositions concerning it can be evaluated.

Plateau P2 does not by itself determine how a concept will be measured or coded. A definition can be coherent while its operationalization remains inadequate. Nor does a well-formed definition establish that a proposition using the concept is empirically supported.

Separating Plateau P2 from those later questions allows conceptual failure to be identified as conceptual failure rather than disguised as a measurement or evidentiary problem.

3.3.3 Plateau P3 — Formal-Operational

Plateau P3 concerns how objects and relations are represented, formalized, classified, measured, or observed.

How can the relevant theoretical distinction be represented or made operational without silently changing what it means?

This plateau includes mathematical representations, taxonomies, coding schemes, variables, observational rules, measurement procedures, classification systems, and other devices through which conceptual objects become formally or empirically tractable.

The Participation framework’s Grammar of the Act illustrates this role. Realization Orientation, Foreclosure Position, and Actor Dependence do not constitute the existence of participatory acts and are not simply alternative definitions of Participation. They provide a structured representation through which different acts can be distinguished and examined.

Exposure offers a particularly useful Plateau P2–Plateau P3 boundary case. Once Exposure has been conceptually distinguished, Plateau P3 asks how that distinction is represented or observed: through notation, variables, coding rules, indicators, or other operational devices. A defective operationalization may therefore require revision without automatically destroying the underlying conceptual distinction.

That separation imposes discipline in both directions. Formal elegance cannot compensate for conceptual ambiguity, and conceptual plausibility cannot establish that a proposed measurement actually captures the object it claims to represent.

Plateau P3 is consequently a major site of possible repair. Operational definitions can be revised, categories split or merged, measures replaced, and formal relations corrected while the resulting consequences for dependent propositions are explicitly traced.

3.3.4 Plateau P4 — Propositional-Epistemic

Plateau P4 concerns what is proposed, tested, found, rejected, unresolved, untested, non-operable, or unreproduced.

What claim is being made? What would count against it? What is its present epistemic status?

The wording above describes the logical scope of Plateau P4; it is not intended as an exhaustive result-status register. The Unified Model maintains a more discriminating epistemic-status architecture because materially different evidentiary conditions should not be compressed into a single supported/not-supported field.

Plateau P4 contains propositions and hypotheses, their exposure to tests or falsifiers, and the epistemic statuses that result from inquiry. The established status vocabulary distinguishes, where applicable, Supported / direct case support, Finite non-refutation (not falsified under the specific test conditions exercised; not confirmed), Not observed, UnderObserved / Unresolved, Untested, Non-operable, Unidentified, and FirstCoder / unreproduced. These statuses record different relationships between a claim and the evidence available to adjudicate it.

Plateau P4 is therefore not simply an “evidence layer.” A proposition belongs here before it is tested as well as after evidence changes its status. What matters is that the theory has moved from defining or representing an object to making an assessable claim concerning it.

The Participation work illustrates the distinction. A grammar can exist at Plateau P3 while a proposition concerning the non-fungibility of acts occupies Plateau P4. Evidence may then support the proposition at its claimed scope, fail to trigger a specified falsifier, leave the matter unresolved, or reveal that the relevant test cannot presently be executed. None of those outcomes is equivalent to the grammar itself.

Graduality makes the same separation visible. A definition of graduality belongs to Plateau P2. A formal feature record or operational criterion may belong to Plateau P3. A proposition concerning a feature belongs to Plateau P4, as does the status produced when that proposition encounters evidence.

These distinctions are not terminological refinements for their own sake. Finite non-refutation is not confirmation. Unresolved is not rejected. Untested is not unsupported. Non-operable is not false. Unreproduced is not reproduced merely because an initial assignment exists. Likewise, a credible search that does not observe a predicted phenomenon differs from a situation in which the opportunity to observe it was inadequate.

Ordinary prose may say that particular evidence “supports,” “fails to support,” or “leaves unresolved” a claim. Those inferential descriptions should not be confused with the more precise result-status field when the framework has assigned one.

Plateau P4 is therefore central to falsifiability. It records not only what the theory proposes but what has happened to those proposals under scrutiny, including cases in which scrutiny cannot yet produce an adjudicated answer.

3.3.5 Plateau P5 — Integrative-Systemic

Plateau P5 concerns how components interact across domains, time, feedback, realization, and the larger system.

How do distinct objects, mechanisms, conditions, and findings relate within the broader dynamics of Gradual AGI?

Plateau P5 is where the Unified Model becomes integrative. It concerns relationships among components that cannot be understood adequately by inspecting each component in isolation: cross-domain effects, temporal interactions, feedback, realization pathways, dependencies, constraints, and system-level consequences.

Participation illustrates the transition. A participatory act may be treated as an occurrence at Plateau P1, defined at Plateau P2, classified at Plateau P3, and evaluated through propositions at Plateau P4. At Plateau P5, the question becomes how participation interacts with realization: whether and through what pathways participatory consequences alter what becomes possible, blocked, accepted, contested, delayed, or realized.

Exposure illustrates a different Plateau P5 problem. Its significance here lies not merely in possessing a definition or formal representation but in its relationship to the larger realization system: how differences in exposure may alter the relationship between potential capability and realized transformation across affected groups, domains, or populations.

Plateau P5 should not become a receptacle for any statement that sounds broad. Integration requires specified relationships. Merely placing several concepts in the same paragraph does not constitute a system model. Nor does a system-level interpretation automatically inherit the evidentiary status of each component it combines.

This boundary is crucial because integration can otherwise conceal weakness. A strongly supported local finding and a speculative cross-domain relationship are not epistemically equivalent simply because both appear within the same Unified Model.

3.3.6 Separation Without Isolation

The five plateaus distinguish logical roles, but they do not isolate them from one another.

A Plateau P1 object may require a Plateau P2 definition. A Plateau P2 distinction may be represented through a Plateau P3 formalization. A Plateau P3 representation may enable a Plateau P4 proposition to be tested. A Plateau P4 finding may alter a Plateau P5 interpretation. Failure at any point may also propagate in the opposite direction when genuine dependencies require reconsideration.

But none of these relationships should be assumed merely because the plateau numbers are adjacent.

This is why the Multi-Plateau Framework is not a pipeline.

A Plateau P4 empirical failure may expose a defective Plateau P3 measure rather than a defective Plateau P2 concept. A Plateau P5 inconsistency may reveal that two Plateau P1 objects were improperly distinguished. A new Plateau P1 phenomenon may require a new Plateau P2 category before any satisfactory Plateau P3 operationalization exists. The direction of inquiry follows the problem, not the numbering.

The plateaus therefore perform two functions at once. They preserve differences among logical types, and they make dependencies among those types visible enough to examine.

That combination provides the basis for the framework’s later account of falsification and repair.

Separation makes failure localizable. Connection makes its consequences traceable.

Table 1. Practical boundary tests between adjacent plateaus

Boundary	Stay on the earlier plateau when…	Shift when…
Plateau P1 → Plateau P2	the question is what object, event, actor, condition, or process is being posited	the question becomes what the term means or how its conceptual boundary is drawn
Plateau P2 → Plateau P3	the question is semantic or classificatory meaning	the concept is represented, coded, measured, formalized, or made observable
Plateau P3 → Plateau P4	the question concerns the instrument or representation itself	an assessable claim is made or a test changes the claim’s epistemic status
Plateau P4 → Plateau P5	the claim is local to a proposition, test, or finding	the claim integrates components across domains, time, feedback, dependencies, or realization

3.4 Facets: Cross-Cutting Properties of Theoretical Objects

The plateaus distinguish what kind of logical work an object or claim is doing. They do not, by themselves, preserve every other distinction that matters for the Unified Model.

A proposition at Plateau P4, for example, may be current or superseded, independently validated or unreproduced, inherited from an earlier paper or introduced in the present synthesis, operable or non-operable, and retained in the book or preserved only as historical genealogy. None of those properties is recoverable merely from knowing that the object occupies Plateau P4.

The Multi-Plateau Framework therefore uses a second form of classification: facets.

A facet is a metadata or analytical property of a theoretical record that cannot be reconstructed from plateau membership without analytically relevant information loss.

Plateau identifies logical role. Facet preserves independently varying information about the record occupying that role.

The distinction matters because the two classifications answer different questions.

Knowing that an object belongs to Plateau P2 tells us that it is doing conceptual-semantic work. It does not tell us whether it is a Definition, Explanandum, Positive Feature, or Constraint; whether its present formulation is the first or third version; whether its role changed across source occurrences; or whether its validation attaches to the current formulation or an earlier one.

Likewise, knowing that a proposition occupies Plateau P4 does not establish whether a test has merely been formally specified, is presently operable, has actually been exercised, has reached adjudication, or what result followed. These are separate properties and must remain separately recordable.

Facets are therefore cross-cutting. The same facet may apply to records located on different plateaus, and records on the same plateau may differ sharply along a facet.

This is why the framework does not treat facets as sub-plateaus.

3.4.1 Non-Derivability, Not Statistical Independence

The word independent requires care here.

Facet independence does not mean that a facet must be statistically uncorrelated with plateau membership. Some relationships may be quite strong. Definitions, for example, will naturally occur disproportionately within the Conceptual-Semantic plateau.

The relevant test is functional rather than statistical:

Would reconstructing the facet from plateau membership erase information that an analytical operation later needs?

If yes, the facet performs independent work and should remain separately represented.

If no—if the proposed field can be recovered reliably from information already retained and no downstream operation depends on the distinction—the separate field is a candidate for removal.

This is the framework’s Field Commitment Rule. Facets should be created because they preserve necessary distinctions, not because another classification can be imagined.

The rule prevents the Multi-Plateau Framework from becoming an unlimited metadata system.

3.4.2 Three Families of Facets

The current working register can be understood through three broad families.

Genealogical and record-management facets preserve an object through intellectual change. They include stable identity, version, lifecycle, provenance, and book disposition. Together they make it possible to say that a formulation has been revised, merged, superseded, retired, or retained without pretending that its earlier form never existed.

This is indispensable for epistemic repairability. Revision requires a stable answer to the question: what exactly changed?

Theoretical-function facets distinguish what work a record performs within the model. These include Standing, Role, Locus, and Coordinate Reach or Signature.

Standing records the object-level kind of theoretical contribution—for example, Definition, Proposition, Positive Feature, Constraint, Explanandum, Design Corollary, Synthetic Principle, or Meta-Rule where applicable.

Standing is retained in the present working architecture because plateau membership alone does not reconstruct it without loss. Its status is nevertheless not final. The technical appendix implements a conditional-derivability test: given Plateau assignment plus the other retained facets, can Standing be reliably reconstructed? If yes, Standing becomes a candidate for merger or retirement; if no, it remains independently informative. The main chapter therefore uses Standing without treating its present field structure as immutable.

Standing is used in ordinary analysis when the task requires distinguishing object-level contribution type beyond plateau membership; otherwise it may be omitted.

Role records what work an object performs in a particular source occurrence. This distinction is provenance-sensitive: the same normalized object can perform different work in different contexts without becoming a different object.

Locus identifies where in the modeled process or system the object operates.

Coordinate Reach or Signature records which formal dimensions an object directly asserts or modulates.

These fields are related, but they are not interchangeable.

Epistemic facets preserve the state of scrutiny. They include Validation State, Empirical-Contact Stage, and Result Status.

Validation State records whether the relevant target version has been independently examined, adjudicated, left unresolved, assigned only by an initial coder, or otherwise occupies a particular validation condition.

Empirical-Contact Stage records how far a proposition or test has progressed through formal specification, operability, exercise, and adjudication.

Result Status records what followed, where a result can legitimately be assigned.

These fields do not replace domain-specific judgments of reliability, construct validity, internal validity, external validity, predictive validity, or other claim-specific burdens. They record where empirical contact stands and what result follows; finer validity analysis remains attached to the relevant method and claim. A measure can therefore be reliable yet of questionable construct validity, and a proposition can be supported at one domain scope without being promoted to universal support.

The separation between stage and result is especially important. A test can be formally specified but non-operable. A test can be exercised without producing the evidentiary burden required for adjudication. An adjudicated test can produce Supported / direct case support, Finite non-refutation, Not observed, or another recognized result. Collapsing those possibilities into a single “validated” field would destroy information that later reasoning requires.

3.4.3 Why Facets Matter for Repair

Facets make epistemic repair more precise because they permit change along one dimension without silently changing every other dimension. If a proposition is revised after a failed test, its stable identity may remain traceable while its version changes; the earlier formulation can remain superseded rather than disappear; the new version need not inherit the old validation state; and dependent propositions can be reviewed explicitly. The practical questions are therefore simple: what persisted, what changed, what status belonged to the earlier and replacement versions, and which conclusions actually depended on the changed object?

Graduality and Participation illustrate the point. A Graduality feature can retain identity while performing different roles in different source contexts, and a changed target can reset validation without moving the object to another plateau. Conditional Non-Fungibility likewise remains a Plateau P4 proposition while its empirical status can vary by exercise and scope. Plateau membership identifies logical role; facets preserve what happened to the object over time and under scrutiny.

3.4.4 Facets Are Also Provisional

The facet system must satisfy the same open, flexible, and falsifiable architecture as the rest of the Multi-Plateau Framework.

A field should not survive merely because it has been used before. If later analysis shows that a facet is reconstructible from other retained information without relevant loss, it should be merged or retired. If future theoretical work repeatedly requires information that the existing facets cannot preserve, a new facet may be warranted.

The current facet register should therefore be understood as a tested working structure, not as a claim that the Unified Model will forever require exactly the same number of fields.

This restraint matters. An open framework does not commit itself in advance to endless expansion, and a flexible framework does not preserve classifications after they cease to do useful work.

The governing test remains informational:

Does the distinction preserve something the theory needs to know that its other classifications would otherwise erase?

If it does, retain it.

If it does not, do not multiply the architecture.

This is the same discipline that governs the plateaus themselves. The purpose of the Multi-Plateau Framework is not maximal classification. It is sufficient differentiation to make theoretical work, evidentiary status, dependency, and revision visible.

Plateaus tell us where a kind of logical work is occurring.

Facets tell us what else must be known about the object doing that work.

Together, they provide the minimum structure required for non-flattening, traceability, and repair.

3.4.5 Minimum Application Mode

The full register is an audit architecture, not a requirement that every ordinary analysis populate every field. A minimal application of the MPF asks only five questions:

1. What logical role is being performed? Assign the relevant Plateau P1–Plateau P5 location.
2. What additional information would be lost if only plateau membership were kept? Record only those facets needed for the task.
3. What are the principal relations, and would an inference change if the relation were mis-typed?
4. What would count as adverse contact with the claim or instrument?
5. If the object has changed, what version was tested and what dependent conclusions must be revisited?

Stable identity/version and provenance become mandatory when revision history matters; empirical-contact stage and result status become mandatory when an empirical proposition or test is reported. Standing, Role, Locus, Coordinate Reach, Book Disposition, and other current fields are optional unless the analytical operation specifically requires them. The technical appendix retains the exhaustive working register and the open conditional-derivability test for Standing. Where repair occurs, Section 3.8.2 gives the minimum archive record needed to preserve the tested object and its downstream consequences.

3.5 Non-Flattening and Typed Cross-Plateau Relations

Separating theoretical objects into plateaus and facets solves only part of the architectural problem. The Unified Gradual AGI Model must also specify how those objects relate.

Without such discipline, distinctions preserved in one part of the framework can be flattened again through vague relational language. A concept may be said to “lead to” an outcome when only an association has been observed. A formal representation may be treated as though it were the construct itself. An empirical finding may be described as validating a definition. A system-level interpretation may inherit evidentiary certainty that belongs only to one of its components.

The Multi-Plateau Framework therefore applies a Non-Flattening Rule:

Different logical kinds must not be collapsed into one apparent hierarchy or treated as interchangeable merely because they concern the same subject.

Non-flattening applies both to objects and to relations.

3.5.1 Why Relations Must Be Typed

Generic expressions such as affects, relates to, connects with, or contributes to can be useful in ordinary exposition, but they are insufficient when the type of relationship matters to the claim.

Consider several statements:

• A definition specifies a construct.
• A formalization represents that construct.
• A measurement observes some feature intended to correspond to it.
• A proposition makes a claim involving it.
• Evidence may support, fail to falsify, or leave unresolved that proposition.
• A process may transform a system condition.
• An observed event may be associated with another event without establishing that it caused it.

These are not stylistic variants of the same relation. Each licenses different inferences.

The framework therefore distinguishes relations by the work they perform.

In the minimum application mode, the first required distinction is simply whether a relation is a world/model claim or a representational/epistemic operation. Finer labels are invoked only where the permitted inference would change—for example, observed pathway versus causal-hypothetical relation, or formalization versus measurement.

3.5.2 World and Model Relations

One family concerns relationships among the objects and processes being modeled.

A constitutive relation means that one object is part of what another object is.

A structural relation means that one object constrains or organizes the state space or relationship of another.

A temporal relation concerns precedence, lag, persistence, accumulation, or change through time.

A transformational relation concerns participation in the conversion of one state or condition into another.

A causal-hypothetical relation proposes a causal mechanism that has not yet been identified or adequately tested.

An observed association or pathway records documentary or empirical connection without claiming counterfactual identification.

A normative relation supplies a preferred direction, constraint, or design desideratum rather than a descriptive causal claim.

These distinctions are particularly important in a model that combines conceptual, empirical, institutional, and normative material. A preferred direction is not an observed effect. An observed pathway is not a demonstrated cause. A structural constraint is not automatically a temporal sequence.

Participation provides a useful example. If community input is followed by an institutional response and the documentary record attributes that response to the input, the framework may record an observed pathway. That does not by itself identify the counterfactual causal contribution of participation. The relation must remain typed at the evidentiary level actually established.

3.5.3 Representational and Epistemic Relations

A second family concerns relationships among theoretical objects across the plateaus.

An ontological object may be denoted, defined, or individuated conceptually.

A concept may be formalized, operationalized, measured, or classified.

A proposition may predict, be tested, be falsified, or fail to be falsified.

Evidence may support, fail to support, or leave unresolved a claim.

One object may constrain, condition, or qualify another.

A later result may transform, feed back into, or revise an earlier representation or subsequent inquiry.

The two relation families are deliberately expressed differently. World/model relations are named relation types because the relation itself is part of what the model asserts about the world. Cross-plateau relations are typed primarily by the operation performed—defining, formalizing, measuring, testing, supporting, revising, and related operations. Giving each operational cluster an additional nominal category would add terminology without adding analytical information.

These mappings make movement across the plateaus explicit without converting the plateaus into stages.

For example, the existence of Exposure as an object or condition at Plateau P1 does not determine its Plateau P2 definition. The definition does not guarantee that a Plateau P3 operationalization captures it adequately. A successful operationalization does not establish a Plateau P4 proposition involving Exposure. And an empirical result involving Exposure does not automatically determine its Plateau P5 significance for the realization system.

Each relation must do its own work.

3.5.4 Non-Identity as an Inference Discipline

Several existing meta-rules of the Unified Model instantiate this broader principle.

M004 — Definition ≠ Proposition. A definition determines how a term is used within the model. A proposition places a claim at empirical or analytic risk. The former cannot acquire the status of the latter merely because both concern the same object.

M005 — Definition ≠ Empirical Finding. A defined construct is not validated merely because it has been defined, nor does the observation of compatible cases turn the definition itself into an empirical result.

M008 — Formalization ≠ Construct ≠ Measurement. A theoretical distinction, its formal representation, and the procedure used to estimate or observe it may correspond closely while remaining different objects.

M011 — Formal Coverage Rule. Evidence for a phenomenon does not imply that the current formal apparatus adequately represents it.

These separations make error easier to diagnose. If a measurement fails, the framework need not immediately discard the construct. If a proposition fails, it need not rewrite the definition that made the proposition expressible. If a formalization is incomplete, the observed phenomenon does not disappear with it.

The same discipline works in the opposite direction. A compelling concept cannot rescue a poor measure. An elegant formalism cannot substitute for evidence. A positive empirical result cannot retroactively establish that every relation assumed in producing it was correctly specified.

3.5.5 Relation Types Can Change Only by Evidence or Argument

Typed relations are not permanent labels.

A causal-hypothetical relation may later acquire stronger evidentiary support. An apparent structural relation may prove to be only an association. An operationalization may be shown not to measure the construct it was intended to represent. A system-level interpretation may require contraction after one of its dependencies fails.

But such changes must themselves be recorded.

A relation should not silently migrate from hypothesized to observed, from observed to causal, or from unresolved to established merely because later exposition uses stronger language.

This is another form of the anti-rescue principle.

The framework permits promotion when it is earned. It prohibits promotion by paraphrase.

3.5.6 Separation, Connection, and Dependency

The Multi-Plateau Framework therefore rejects two opposite errors.

The first is flattening: treating objects of different logical kinds as though they were interchangeable.

The second is isolation: treating the plateaus as sealed compartments whose contents have no dependencies on one another.

The framework requires neither.

Objects remain distinct while their relationships are explicitly typed. Those relationships identify where a change may need to propagate when one component fails.

If a Plateau P4 proposition fails, the framework can ask whether the failure bears only on that proposition or whether it also exposes a problem in its Plateau P3 operationalization, its Plateau P2 conceptual boundary, or a Plateau P5 synthesis that depended upon it.

If a Plateau P3 measure fails, the framework can determine which propositions actually used that measure rather than treating every claim involving the underlying concept as invalid.

If a Plateau P2 distinction is revised, the framework can identify which formalizations, propositions, and system-level interpretations depend on the previous version.

Typed relations therefore make dependency visible.

And visible dependency makes disciplined propagation possible.

Non-flattening preserves what is different. Typed relations specify what is connected.

3.6 Falsifiability: Making Failure Detectable

An open and flexible architecture is not necessarily a self-correcting one. A framework can accommodate new phenomena and permit revision while still protecting its central claims from meaningful failure.

The third design requirement of the Multi-Plateau Framework is therefore falsifiability.

In this chapter, falsifiability should be understood with a scope appropriate to the logical diversity of the model. Not every theoretical object is an empirical proposition, and not every form of failure is a failed statistical test. What matters is that claims be stated precisely enough for evidence or argument to show where they are false, unsupported, inadequate, incoherent, non-operable, or otherwise unable to bear the work assigned to them.

The governing principle is:

Falsifiability requires the possibility of informative failure.

A framework that can reinterpret every contrary result as consistent with itself has not become robust. It has become difficult to test.

3.6.1 Different Logical Objects Fail Differently

The plateaus make it possible to distinguish several kinds of failure.

At Plateau P1, an ontological distinction may fail if the objects it treats as separate cannot in fact be coherently distinguished, or if a supposedly necessary kind of object repeatedly proves unnecessary to the explanatory work of the theory.

At Plateau P2, a concept or definition may prove ambiguous, internally inconsistent, unable to discriminate cases it was introduced to distinguish, or inadequate to the phenomena the theory now needs to represent.

This is not the same as empirically “falsifying a definition.” Definitions are stipulative. But a stipulation can still fail as theoretical architecture if it cannot perform the work for which it was introduced.

At Plateau P3, a formalization or operationalization may fail through invalid representation, inadequate coverage, unreliable classification, non-identifiability, or inability to instantiate the conceptual distinction it claims to represent.

At Plateau P4, propositions encounter the most familiar form of falsification. A specified discriminator may fail. A prediction may be contradicted. A falsifier may trigger. Evidence may fail to support a claim. A test may instead prove non-operable or remain unexercised, in which case the appropriate conclusion is not falsification but a different epistemic status.

At Plateau P5, an integrative claim may fail because its proposed relationships are contradicted, because one of its required dependencies fails, because apparently distinct components prove redundant, or because the synthesis cannot account coherently for the interactions it was introduced to explain.

These failure modes should not be collapsed into a single binary field.

A failed measure is not a rejected proposition.

A rejected proposition is not an incoherent definition.

An inadequate or incoherent definition is not, by itself, evidence that the underlying Plateau P1 object does not exist.

A non-operable test is not a negative empirical result.

A system-level contradiction is not automatically evidence that every component of the system model is false.

The architecture makes failure useful precisely by preserving these differences.

3.6.2 Exposure to Test Comes Before Repair

Epistemic repair begins before anything breaks.

A claim must first be exposed to a form of scrutiny capable of producing an adverse result.

For an empirical proposition, that may require a discriminator, prediction, threshold, comparison, or falsifier stated before the relevant result is known.

For a classification system, it may require inter-rater replication, holdout cases, discriminant tests, or evidence that its categories perform work that simpler alternatives do not.

For a formal model, it may require checks of internal consistency, representational coverage, identification, boundary conditions, and comparison with observations the formalism claims to explain.

For a system-level synthesis, it may require explicit dependency statements and conditions under which the proposed integration would need contraction or revision.

The form of scrutiny varies with logical role.

The obligation does not.

A theoretical object should not receive stronger standing merely because no one has yet specified how it could fail.

3.6.3 Failure, Non-Refutation, and Absence of Contact

The framework must also distinguish failure from the absence of failure.

A falsifier that is exercised and not triggered produces Finite non-refutation, not proof.

A required phenomenon that is sought under credible observation conditions and not found can produce the result status Not observed.

Inadequate observation cannot support the same inference; it belongs instead under UnderObserved / Unresolved.

A formally specified test that current evidence cannot execute is Non-operable, not corroborated.

A test that was never run is Untested, not passed.

Evidence unable to distinguish competing mechanisms leaves the mechanism Unidentified, rather than establishing whichever explanation the theory prefers.

These distinctions matter because an evolving theory will accumulate many cases in which decisive adjudication is impossible. Treating all such cases as provisional support would systematically bias the theory toward survival.

The appropriate discipline is simpler:

No contact with a falsifier is not evidence that the claim survived it.

3.6.4 Local Falsification

The Multi-Plateau Framework adds one requirement beyond ordinary exposure to test: failure should be localized as far as the evidence permits.

Suppose a prediction fails.

Several explanations may remain possible:

• the proposition itself is wrong;
• the concept was inadequately defined;
• the operational measure did not capture the concept;
• an assumed relationship was mis-typed;
• an omitted condition changed the result;
• the data were inadequate to adjudicate the test;
• or a larger system-level interpretation incorrectly treated a local claim as universal.

The framework should not choose among these possibilities merely to preserve the theory.

Instead it records what the result directly establishes, identifies which dependencies are implicated, and leaves alternatives unresolved where the available evidence cannot distinguish them.

This is where separation becomes methodologically valuable.

Falsifiability makes failure detectable. The plateau architecture makes failure localizable.

Detection without localization can produce indiscriminate rejection.

Localization without genuine exposure to failure can produce elaborate bookkeeping around an unfalsifiable theory.

The framework requires both.

3.6.5 The Unified Model Must Be Exposed Too

The requirement applies not only to propositions inside the Unified Gradual AGI Model but to the Multi-Plateau Framework itself.

The architecture would require revision if, for example:

• theoretically important objects repeatedly cannot be represented without category confusion;
• plateau distinctions prove systematically redundant;
• the architecture repeatedly fails to identify what kind of object or claim has failed;
• new domains require logical objects that cannot be accommodated coherently within the present five-plateau structure;
• plateau and facet distinctions cease to preserve independently useful information;
• or repair procedures begin to function as devices for protecting claims from falsification.

These are not predictions that such failures will occur. They specify ways in which the present architecture could prove inadequate.

That possibility is essential to the chapter’s larger argument.

The Multi-Plateau Framework cannot make falsifiability a requirement for everything inside the model while exempting its own structure from revision.

It is a provisional architecture for an Ongoing, Unfinished Transformation.

Its purpose is not to guarantee survival.

Its purpose is to make survival conditional on continued scrutiny.

3.6.6 From Detectable Failure to Repair

Falsification alone does not tell a theory what to do after a failure occurs.

A failed proposition may need withdrawal.

A defective measurement may need replacement.

A definition may require narrowing.

A formal model may require reconstruction.

A system-level interpretation may need to contract.

And sometimes the appropriate response is to preserve the existing object while changing only its epistemic status.

The next problem is therefore not whether the framework can change. Flexibility has already established that it can.

The problem is whether it can change without losing the record of why the change became necessary, without altering the failed test after the fact, and without allowing unaffected components to inherit either the failure or the repair automatically.

That is the problem of epistemic repairability.

3.7 Epistemic Repairability

Falsifiability makes failure possible. The Multi-Plateau Framework adds a further requirement: when failure occurs, the framework should make it possible to determine what failed, what depends on it, and what must change as a consequence.

This capacity is epistemic repairability.

Epistemic repairability is the capacity of a framework to detect error, localize what failed, revise the affected structure, propagate the consequences of that revision, and expose the repaired model to renewed scrutiny.

Repairability is not the same as flexibility. A flexible theory can change. A repairable theory must be able to explain why it changed, where the change belongs, what else the change affects, and what epistemic standing the revised object has after the change.

That distinction is central to the Multi-Plateau Framework.

3.7.1 The Repair Sequence

The full repair sequence contains seven elements:

Exposure to test → failure detected → failure localized → dependencies identified → affected component revised → consequences propagated → revised claim retested

The sequence is logical rather than mechanically chronological. Some operations may overlap, and a repair may require iteration. But none should be silently omitted when it bears on the integrity of the result.

Exposure to test establishes the possibility of failure. Without a discriminator, falsifier, comparison, audit, replication, consistency check, or other appropriate form of scrutiny, there is no adverse result to repair from.

Failure detected records that the relevant burden was not met. Detection should occur under the test actually specified, not under a substitute criterion invented after the result becomes inconvenient.

Failure localized asks what kind of object or claim has failed. Is the problem ontological, conceptual, formal, operational, propositional, evidentiary, or systemic? The plateau architecture exists partly to make that question answerable.

Dependencies identified asks which other objects genuinely rely on the failed component. Dependency must be traced rather than assumed from proximity, shared subject matter, or neighboring plateau number.

The affected component is revised only after the failure and its scope have been recorded. Revision may narrow a definition, replace a measure, change a formal representation, withdraw a proposition, split a category, revise a relation type, or contract a system-level interpretation.

Consequences are propagated to objects whose standing genuinely depends on the repaired component. A changed definition may require re-examination of formalizations using it. A changed measure may reset findings generated through that measure. A rejected proposition may require contraction of a Plateau P5 synthesis that depended upon it.

Finally, the revised object is exposed to scrutiny again. Repair does not confer validation. It produces a new or revised object whose standing must be earned under the evidentiary burden appropriate to that object.

The sequence therefore returns the framework to exposure rather than to closure.

3.7.2 Localization Before Revision

The most important operation is often localization.

Without localization, theoretical failure tends to produce one of two bad responses.

The first is overreaction: a local failure is treated as evidence that the entire theory has collapsed.

The second is underreaction: the theory changes something nearby while leaving the actual failed dependency untouched.

The plateau architecture reduces both risks.

Suppose a Plateau P4 proposition fails because a predicted empirical pattern is not observed. The failure may belong to the proposition itself. But it may instead expose a Plateau P3 measurement problem, a Plateau P2 ambiguity, an unidentified conditioning variable, or an overextended Plateau P5 interpretation.

The first result does not automatically identify which explanation is correct.

Repair therefore begins by recording the failure at the narrowest level the evidence actually establishes.

If the evidence rejects the proposition but cannot identify why, the proposition may be rejected while the cause of failure remains unresolved.

If the evidence establishes that a measurement is unreliable, the measurement can be repaired without pretending that the underlying concept has thereby been refuted.

If a conceptual distinction itself proves incoherent, repairing only its operationalization would be insufficient.

Localization determines the legitimate repair target.

Table 2. Failure-localization diagnostic heuristic

This is a practical checklist, not a new canonical error taxonomy. It orders questions without presuming that the first plausible explanation is correct.

Observed problem	First localization checks	Do not infer automatically
Prediction contradicted	test/data integrity → Plateau P3 operationalization → Plateau P2 scope/definition → Plateau P4 proposition	that the Plateau P1 object does not exist or the whole model failed
Low coding agreement	evidence access → coding rule → category boundary → construct definition	that the substantive proposition is false
Formal object non-identifiable or non-operable	available variables/data → representation → claim burden	that absence of a result counts as support or refutation
Plateau P5 synthesis conflict	typed relations → load-bearing dependencies → local findings and scope	that every component of the synthesis is false
New phenomenon fits poorly	existing ontology/concept boundaries → representation → only then architectural adequacy	that a new plateau or facet is immediately required

3.7.3 Dependency Determines Propagation

Repairability also depends on distinguishing connection from dependency.

Many objects in the Unified Model are related. Fewer are logically or evidentially dependent in ways that require a change in one to alter the standing of another.

A revised Plateau P2 definition may invalidate a Plateau P3 coding rule constructed from its previous boundary. But it need not affect another Plateau P3 formalization that never depended on the changed clause.

A failed Plateau P3 measure may require reassessment of Plateau P4 findings produced through that measure. It does not automatically invalidate every proposition involving the underlying Plateau P2 concept.

A rejected Plateau P4 proposition may require contraction of a Plateau P5 synthesis that used the proposition as a premise. It does not automatically erase unrelated components of that synthesis.

This is why typed relations matter for repair.

They make it possible to ask not merely whether two objects are connected, but how they are connected and whether the changed component was load-bearing for the dependent claim.

Propagation should follow those dependencies.

It should go no farther merely for consistency of appearance.

3.7.4 Versioning Makes Repair Visible

A repaired theory must preserve the object that failed.

This requirement is already encoded in the Unified Model’s M003 Versioned-Object Rule: retired, merged, revised, and current formulations are preserved rather than rewritten out of the intellectual history.

A revised formulation therefore does not erase its predecessor.

If version v1 fails a specified test and is replaced by v2, the historical record remains:

• v1 was the object tested;
• the test produced the recorded result;
• the failure motivated or contributed to the revision;
• v2 contains the identified change;
• and v2 begins with the epistemic standing warranted by the revised object.

This last point is captured by M029 Version-Specific Validation: changed targets or governing rules create new versions whose validation state resets.

A repaired formulation therefore cannot inherit validation merely because its predecessor had previously been tested.

The distinction is easy to overlook. A small revision can feel like “the same theory,” especially where the stable identity of the object has been preserved. But stable identity and immutable formulation are different facets.

Repairability depends on preserving both.

3.7.5 A Worked Pattern: Graduality

The Graduality reconstruction provides a documented example of repair without theoretical collapse. An early analysis found an exact correspondence between four failed DH1 cases and the four candidate signatures then carrying State Structure. Post-pilot repair G-R7 later changed G008’s direct signature by removing State Structure because observational partiality did not independently require it. That local Plateau P3 repair destroyed the apparent correspondence and therefore changed the dependent Plateau P4 interpretation. The published catalog retained the old pattern only as a version-sensitive artifact rather than rescuing it as independent validation.

Table 3. G-R7 through the seven-step repair sequence

Repair step	G-R7 instantiation
1. Exposure to test	The signature scheme and DH1 cases were compared under the then-current coding rules.
2. Failure detected	Post-pilot review showed G008 had been overcoded with State Structure.
3. Failure localized	The defect was localized to the Plateau P3 direct-signature assignment, not to Graduality as a whole.
4. Dependencies identified	The apparent State-Structure/DH1 correspondence depended on G008 remaining State-Structure-bearing.
5. Component revised	G008 changed from {S, RO} to {RO}.
6. Consequences propagated	The exact correspondence disappeared and lost its standing as independent coordinate validation.
7. Revised claim retested	The repaired target did not inherit prior validation; its new standing remained limited to the evidence actually obtained.

Nothing in the episode required abandoning the Graduality construct. The repair was local, but its consequence was not: a changed representation altered a result that depended upon it while preserving the historical record of what had previously been observed.

The pattern can be summarized compactly: the object changed → the dependent result changed → the historical record did not.

3.7.6 Repair Can Contract as Well as Extend

Repair is often imagined as adding complexity: another variable, another qualification, another exception.

That is not the preferred direction of the Multi-Plateau Framework.

A legitimate repair may instead make the theory smaller.

A proposition may be withdrawn.

A facet may be shown redundant and retired.

A universal claim may become domain-specific.

A causal claim may contract to an observed pathway.

A formal apparatus may be recognized as covering less of the phenomenon than previously assumed.

A synthetic principle may lose canonical standing, merge into another principle, or remain only as genealogy.

Such contractions are not defects in repairability. They are among its strongest expressions.

The purpose of repair is not to preserve the maximum amount of theory.

It is to preserve only what remains warranted after the failure is incorporated.

This is especially important for an Ongoing, Unfinished Transformation. Openness creates pressure to add new objects as the transformation changes. Repairability must supply an equal capacity to narrow, merge, retire, or abandon objects that no longer earn their place.

3.7.7 Repairability as a Property of the Architecture

Epistemic repairability is therefore not a promise that the Unified Gradual AGI Model will converge toward a final, error-free representation.

The transformation may continue to change. New evidence may expose new failures. Previously adequate concepts may become insufficient for new domains. New relationships may alter the dependencies among components.

Repairability makes a narrower claim.

The framework is designed so that failure need not be hidden, globalized, or forgotten.

Separation makes it possible to identify what kind of object failed.

Falsifiability makes the failure detectable.

Typed relations make dependencies traceable.

Flexibility permits the affected structure to change.

Versioning preserves the historical record.

Renewed scrutiny prevents revision from becoming automatic vindication.

Together, these capacities make the framework self-corrective in a disciplined sense.

They do not guarantee that every error will be detected, that every failure will be localized correctly, or that every repair will succeed.

A repair can itself fail.

And when it does, it should re-enter the same sequence.

The Multi-Plateau Framework is therefore not a mechanism for returning a theory to an imagined perfect state.

It is an architecture for keeping theoretical change inspectable.

The remaining boundary is decisive. A framework capable of changing after failure can also become very good at explaining away every failure it encounters.

Epistemic repairability is defensible only if revision is prevented from becoming post hoc immunity.

That is the purpose of the next section:

Repair is not rescue.

3.8 Repair Without Rescue

Epistemic repairability creates a danger of its own.

A theory capable of revising definitions, measurements, propositions, relations, and system-level interpretations after encountering failure can become difficult to falsify for exactly the reason it appears adaptable. Every adverse result can be followed by another qualification. Every failed prediction can generate a narrower prediction. Every rejected mechanism can be replaced by another.

Flexibility can therefore become a means of theoretical survival rather than theoretical correction.

The Multi-Plateau Framework imposes a simple boundary:

Repair is not rescue.

A repair responds to failure while preserving what the failure established.

A rescue changes the theory in a way that makes the earlier failure disappear.

The difference is not whether the theory changes. Both can involve revision. The difference is whether the revision remains accountable to the historical test that made revision necessary.

The governing rule is:

Repair may change the model; it may not change the historical fact that the previous model failed its test.

3.8.1 What Legitimate Repair Requires

A legitimate repair satisfies several conditions.

It records the earlier claim and its result rather than replacing them with the revised formulation.

It preserves the version that was actually tested so that later readers can identify what the evidence encountered.

It identifies what changed between the failed object and its replacement.

It does not redefine the governing test after seeing the result merely to convert failure into success.

It traces the dependencies affected by the change rather than propagating revision indiscriminately.

It withdraws, contracts, or reclassifies conclusions that are no longer warranted.

And it returns the revised object to scrutiny rather than treating successful revision as equivalent to validation.

These requirements make repair an epistemic operation rather than an editorial one.

3.8.2 The Historical Failure Must Survive

The strongest anti-rescue safeguard is version preservation. Section 3.7.4 established the general rule: when a tested formulation is materially revised, the earlier target and its result remain part of the record, while the replacement becomes a new target for epistemic purposes. Stable intellectual identity does not collapse distinct validation histories.

A theory that rewrites the failed formulation as though it had always contained the later qualification has not repaired itself; it has rewritten its past.

The minimum archive record for a repaired claim should include:

• Tested version identifier.
• Governing test or rule version.
• Date and result of empirical or analytical contact.
• Repair note identifying what changed and why.
• Dependent conclusions reviewed after the change.

The purpose of this archive is not bureaucratic completeness. It makes the historical failure citable and keeps later success from being projected backward onto an earlier object.

3.8.3 Do Not Move the Test After the Result

A second rescue pathway is less visible.

The original claim remains in the record, but the criterion by which it was supposed to succeed is modified after the result becomes known.

Suppose a proposition predicts an increase under a pre-specified comparison. The comparison is run and the increase does not appear.

A legitimate response can question the proposition, the design, the measurement, or the scope. It can propose a new test for a revised claim.

What it cannot do is quietly replace the original comparison with another one and then report that the original proposition survived.

The new test may be scientifically valuable.

It is still a new test.

This distinction is particularly important when the evidence is exploratory. Post hoc analysis can reveal mechanisms, boundary conditions, alternative measures, or new hypotheses. The Multi-Plateau Framework does not prohibit such learning.

It prohibits retrospective relabeling.

Discovery after failure may motivate the next claim. It may not rewrite the burden borne by the previous one.

3.8.4 Do Not Promote by Paraphrase

Repair can also become rescue through language.

A result recorded as an observed association/pathway may later be described casually as a mechanism.

A Finite non-refutation result may become “supported.”

An Unresolved result may become “consistent with the theory,” and repeated use of that phrase may eventually make it sound affirmative.

A FirstCoder assignment may be described as though it had been independently reproduced.

A domain-specific finding may gradually be written as though it were universal.

No formal rule needs to change for this kind of promotion to occur. Vocabulary alone can do it.

This is why the Unified Model’s M028 No Silent Promotion matters to repairability. Admission into the framework, observability, analytical usefulness, or survival in one domain does not establish centrality, necessity, universality, or definitional status.

The same principle applies to evidence.

A claim can acquire stronger standing.

But the stronger standing must be earned by new evidence or argument, not by stronger prose.

The framework permits promotion when earned. It prohibits promotion by paraphrase.

3.8.5 Revised Objects Do Not Inherit Validation

A repaired object creates another common temptation.

If the change from v1 to v2 appears small, the evidentiary record accumulated around v1 may feel close enough to carry forward.

That intuition is dangerous.

The relevant question is not whether the revision looks minor. It is whether the target encountered by the previous validation procedure remains the target now being asserted.

Where the target or governing rule has changed materially, M029 Version-Specific Validation applies: the new version begins with the validation state warranted by the new object.

Prior evidence remains part of its genealogy. It may motivate confidence, identify useful tests, or explain why the revision took its present form.

But genealogy is not validation.

The Graduality reconstruction illustrates the point. G-R7 altered G008’s direct signature and thereby dissolved a previously observed correspondence. The repaired signature did not inherit the earlier pattern as validation. The old result remained visible as a version-sensitive artifact.

The object changed.

The dependent result changed.

The historical record did not.

That is repair rather than rescue.

3.8.6 Repair Must Propagate in Both Directions

Anti-rescue discipline also constrains propagation.

A repair cannot be confined to the local object when downstream conclusions depended upon the failed version. Those dependent claims must be reconsidered.

But neither can failure be propagated beyond its dependencies simply because a larger contraction would appear more intellectually dramatic or internally tidy.

Repair therefore has two symmetrical obligations:

do not protect dependent conclusions from a relevant failure;

and do not invalidate independent conclusions merely because they share a framework with the failed object.

This is where the plateau and facet architecture performs practical work.

Logical separation helps identify the repair target.

Typed relations identify genuine dependencies.

Versioning preserves what changed.

Epistemic status records what the evidence now warrants.

The same architecture therefore constrains both under-reaction and over-reaction.

3.8.7 Repair Can Itself Fail

Nothing guarantees that the first repair will be adequate.

A revised definition may remain ambiguous.

A replacement measure may still lack validity.

A narrower proposition may fail its next test.

A newly introduced distinction may prove redundant.

A repaired system synthesis may reveal another contradiction.

The framework should not regard repeated repair as embarrassment to be hidden. Nor should repeated repair automatically be treated as evidence of healthy self-correction.

Repeated failure can itself become evidence that the architecture has localized the problem incorrectly or that a deeper component requires revision.

Repair therefore remains exposed to falsification.

A repair that fails re-enters the same sequence: exposure, detection, localization, dependency identification, revision, propagation, and renewed scrutiny.

There is no privileged point at which a repaired claim becomes exempt from this cycle.

3.8.8 The Anti-Rescue Test Applies to the Framework Itself

The same standard applies to the Multi-Plateau Framework.

If evidence reveals that the five plateaus cannot represent an important class of theoretical objects, the framework may be revised.

If a facet proves redundant, it may be retired.

If a relation type proves inadequate, the relevant register may change.

If another architecture localizes failure more effectively, the present architecture may lose its privileged role.

But such revisions cannot be used to claim that the earlier architecture never failed.

Nor may every counterexample simply produce another plateau, facet, exception, or relation type until all possible observations fit.

That would turn openness into accommodation without risk.

A framework designed for an Ongoing, Unfinished Transformation must remain open to new phenomena.

It must also remain open to the possibility that its own present architecture is wrong.

That is the final meaning of repair without rescue.

The framework should be capable of changing because it has learned something.

It should not be capable of changing so freely that it can never be shown to have been mistaken.

Epistemic repairability preserves the possibility of correction. Anti-rescue discipline preserves the meaning of failure.

3.9 The Multi-Plateau Framework and the Unified Gradual AGI Model

The Multi-Plateau Framework is the architecture through which the Unified Gradual AGI Model is organized. It is not the substantive model itself.

That distinction is essential at this point in the chapter.

The Unified Model contains substantive objects, processes, conditions, relations, propositions, observations, and empirical results concerning Gradual AGI. The Multi-Plateau Framework determines how those different kinds of theoretical material can coexist without being mistaken for one another.

The relationship can be stated simply:

The Unified Model describes the evolving system. The Multi-Plateau Framework organizes the kinds of claims made about it.

This is why the five plateaus should not be drawn as five components inside the realization process. Nothing in the world moves from Plateau P1 to Plateau P5. The plateaus classify theoretical work performed on objects that may themselves participate in dynamic processes.

3.9.1 One System, Multiple Logical Views

The highest-level object of the Unified Model is AGI-inclusive humanity: the evolving system within which humans, artificial intelligence, institutions, infrastructure, resources, and wider social conditions interact.

Its organizing substantive problem is realization: how capability or potential becomes dated, distributed, socially and institutionally realized outcomes.

Those statements themselves already occupy more than one plateau.

AGI-inclusive humanity is an object of the theory at Plateau P1. Its conceptual boundary and meaning require Plateau P2. Formal representations of aspects of the system may occupy Plateau P3. Claims about its behavior or conditions occupy Plateau P4. Its integrated dynamics belong principally to Plateau P5.

Realization is similarly multi-plateau. It can be treated as a process or relation that occurs, conceptually distinguished from capability or potential, formally represented, made the subject of propositions and empirical tests, and integrated across domains and time.

The plateau assignment therefore does not tell us which substantive subsystem an object belongs to.

It tells us what kind of claim is presently being made about that object.

3.9.2 Subsystems Are Not Plateaus

The distinction becomes especially important for the major substantive components inherited from the Gradual AGI research program.

Graduality concerns the characterization of gradual realization and, separately, principled revisability under uncertainty.

Synchronization and CFS concern a bounded family of realization dynamics.

Optimization concerns the evaluation and selection of possible trajectories, choices, and interventions.

Governance concerns contestation, intervention, institutional steering, observation, and the conditions under which realization is shaped or constrained.

Participation concerns human participatory subjects, acts, consequences, and contingent pathways through which those consequences may become realization-relevant.

Observation and evidence concern what parts of the underlying system can be represented, measured, classified, tested, or adjudicated.

Feedback and recursion concern how realized outcomes can alter conditions relevant to later realization.

The frozen Unified Model gives Observation/Evidence and Feedback/Recursion separate architectural standing because they perform model-wide functions. That separate standing does not imply exclusive domain ownership. Governance, Participation, Synchronization, and other substantive components may contain observational or recursive operations while the general Observation/Evidence and Feedback/Recursion structures remain independently represented across the model.

Table 4. Typical component × plateau occupancy in the frozen Unified Model

The mapping is many-to-many: substantive components traverse plateaus according to the logical work being performed, while no plateau is owned by one subsystem. The table is illustrative rather than exhaustive.

Observation/Evidence as a substantive component refers to the model-wide functions by which the system is represented, measured, classified, tested, and adjudicated. Observation as a Plateau P3 operation refers to the particular logical work of representing or measuring a specific object. The component draws on that operation across the model; the operation is not reducible to the component.

Component	Typical MPF occupancy (not exclusive)	Reason
Graduality	Plateau P2–Plateau P4; Plateau P5 when integrated	definitions and conceptual distinctions occupy Plateau P2; formal feature/signature assignments (including G-R7) occupy Plateau P3; propositions/statuses occupy Plateau P4; system integration reaches Plateau P5
Synchronization / CFS	Plateau P2–Plateau P5; Plateau P1 for treated states/events	conceptual and formal realization dynamics with empirical and integrative roles
Optimization	Plateau P2–Plateau P5; Plateau P1 for decision/process objects	concepts, formal selection/evaluation, propositions, and upstream system effects
Governance	Plateau P1–Plateau P5	objects, concepts, mappings, propositions, institutional interventions, and recursive system effects
Participation	Plateau P1–Plateau P5	full traversal from Subject/Act/Consequence through grammar, propositions, and contingent realization pathways
Observation / Evidence	Plateau P3–Plateau P5 primarily	representation, measurement, validation/status, and feedback into later system behavior
Feedback / Recursion	Plateau P1 / Plateau P5 primarily; Plateau P3–P4 when represented/tested	realized outcomes alter later conditions; formal or empirical claims about that recursion occupy other plateaus

3.9.3 The Realization Spine and the Architectural Overlay

The Unified Model is realization-centered.

Chapter 4 develops the substantive realization architecture, including the model’s principal conditioning objects and their formal representation. Those objects are intentionally not re-derived here.

For present purposes, Chapter 3 needs only the architectural relation:

capability and conditioning structures are distinct from realized outcomes, and the mapping between them is itself theoretically consequential.

The Multi-Plateau Framework overlays that substantive spine without becoming another input into it.

It asks:

• What objects does the realization architecture posit?
• How are those objects defined?
• How are they formalized or measured?
• Which propositions concerning them have actually encountered evidence?
• What is the status of those results?
• How do the components interact at system level?
• And, if any answer changes, what else depends upon it?

These are MPF questions.

They are different from the substantive Chapter 4 question of how realization itself should be modeled.

The boundary should therefore remain explicit:

The Multi-Plateau Framework organizes the realization model; it does not compete with the realization model.

3.9.4 Cross-Cutting Components

Cross-cutting does not mean logically undifferentiated. Governance, Participation, and Observation/Evidence can each appear at several plateaus, but every occurrence must still be classified by the work it is doing. The MPF therefore preserves domain continuity without allowing a domain label to substitute for logical analysis.

3.9.5 Graduality as an Overlay, Not Another Mechanism

Graduality requires particular care because it gives the research program its name.

The Unified Model does not treat graduality as another causal mechanism operating alongside Synchronization, Optimization, Governance, or Participation.

Phenomenal gradualness classifies the realization of an individuated transformation. Principled graduality concerns information-responsive, revisable action under the conditions that justify it.

These concepts can characterize or constrain parts of the Unified Model.

They do not constitute another realization input.

This distinction matters because a model of Gradual AGI could otherwise become circular: realization would be explained by “graduality,” while graduality would itself be inferred from the realization the theory was trying to explain.

The Multi-Plateau architecture prevents that collapse by preserving the logical role of each claim.

A classification is not automatically a mechanism.

A normative principle is not automatically a causal relation.

A name for the research program is not automatically an explanatory variable.

3.9.6 Open Items Remain Open Across the Mapping

Integration creates pressure for apparent completeness.

A unified diagram looks cleaner if every variable has a final definition, every subsystem has a settled boundary, every synthetic principle has a fixed status, and every relationship can be represented by a confident arrow.

The Multi-Plateau Framework resists that pressure.

An unresolved object remains unresolved when inserted into the larger model.

A provisional relation does not become canonical because the model needs a connecting line.

An open boundary does not become settled because a diagram requires a box.

A FirstCoder result does not become reproduced because it is summarized in a synthesis chapter.

This is a direct consequence of the anti-rescue discipline developed above.

Integration must propagate what is known.

It must also propagate what is not yet known.

The Unified Model therefore gains coherence from synthesis without gaining evidentiary certainty merely from being synthesized.

3.9.7 Architecture Without Closure

The Multi-Plateau Framework gives the Unified Model a stable logical architecture without supplying another substantive explanation of why Gradual AGI unfolds as it does. Its contribution is to keep objects, concepts, formal representations, measurements, propositions, evidentiary statuses, and system-level interpretations distinguishable while allowing their dependencies to remain traceable.

That structure is enough for cumulative inquiry without requiring theoretical closure: failures can be localized, revisions can propagate, and unresolved questions can remain visible. For an Ongoing, Unfinished Transformation, the MPF therefore supplies the logical architecture of the Unified Gradual AGI Model; Chapter 4 develops the substantive realization architecture it organizes.

3.10 Limits, Falsifiers, and Open Architecture

The Multi-Plateau Framework is intended to make the Unified Gradual AGI Model more coherent, falsifiable, and repairable.

It should not be mistaken for a claim that the present architecture is complete.

The distinction matters because the object of inquiry is an Ongoing, Unfinished Transformation. A framework constructed while that transformation is still developing must be stable enough to support cumulative inquiry while remaining exposed to the possibility that its present organization is inadequate.

The Multi-Plateau Framework is therefore offered as a provisional architecture, not an immutable metatheory.

3.10.1 What the Framework Does Not Claim

The framework does not claim that the five plateaus exhaust every logical distinction that future Gradual AGI research could require.

It does not claim that every current facet deserves permanent retention.

It does not claim that every relation type in the present register is sufficient for every future domain.

It does not claim that organizing a source construct within the Unified Model independently validates that construct.

It does not convert unresolved empirical questions into theoretical closure.

And it does not guarantee that every future failure can be localized correctly or repaired successfully.

The Multi-Plateau Framework provides an architecture for making those questions more tractable.

It does not answer them in advance.

3.10.2 Conditions Under Which the Architecture Should Be Revised

The framework must remain exposed to failure at its own level.

Several conditions would provide reason to revise it.

First, important theoretical objects repeatedly cannot be represented without category confusion. If conceptually distinct objects must continually be forced into unsuitable plateaus, the architecture is not performing its classificatory function.

Second, the plateau distinctions prove systematically redundant. If Plateau P1 through Plateau P5 repeatedly fail to preserve differences that matter to reasoning, evidence, or repair, their separation would become ornamental rather than analytical.

Third, the architecture repeatedly prevents rather than enables localization of failure. A framework introduced partly to identify where problems reside should be reconsidered if failures continually remain ambiguous because of its own organization.

Fourth, new domains require kinds of theoretical objects that the five-plateau structure cannot coherently accommodate. Openness requires admitting this possibility rather than treating the present ontology of theoretical work as exhaustive.

Fifth, plateau and facet distinctions cease to perform independent analytical work. A facet that can be reconstructed from other retained information without relevant loss should be pruned under the Field Commitment Rule. The same principle applies more broadly to distinctions whose retained complexity no longer earns its cost.

Sixth, repair procedures begin to immunize claims from falsification. If versioning, qualification, new categories, or revised relations become means of absorbing every adverse result while preserving all substantive conclusions, epistemic repairability has become rescue.

These conditions do not imply that every local inconvenience requires redesign.

They specify failures of the architecture’s intended function.

A comparative-adequacy check operationalizes the redundancy condition without adding a seventh falsifier: if a materially simpler architecture repeatedly preserves the same distinctions and localizes and propagates failure equally well, the added MPF structure has not earned its cost. Here materially simpler means fewer logical distinctions, mandatory fields, or relation types while preserving the same analytical work. “Repeatedly” means recurrent functional non-addition across relevant applications rather than one awkward case or a fixed amount of calendar time. The framework therefore does not impose a universal time-to-failure threshold.

3.10.3 Open Does Not Mean Unbounded

An open framework is not a framework without constraints.

The Multi-Plateau architecture does not authorize a new plateau whenever an awkward object appears, a new facet whenever another metadata field can be imagined, or a new relation type whenever existing language feels inconvenient.

Expansion must perform analytical work.

The same discipline used throughout this chapter applies to the architecture itself:

• retain distinctions that preserve information the theory needs;
• revise distinctions that fail;
• merge or retire distinctions that have become redundant;
• and record what changed rather than rewriting the genealogy.

This is particularly important because conceptual frameworks often grow more easily than they contract.

The constraints therefore remain coupled: openness needs pruning, flexibility needs versioning, falsifiability needs localization, and repairability needs anti-rescue discipline.

3.10.4 Some Questions Are Intentionally Still Open

The present framework contains unresolved issues that do not prevent its use.

The sharper test of whether Standing carries information beyond Plateau plus other retained facets remains open.

The present facet register remains subject to future pruning under the Field Commitment Rule.

Some synthetic principles remain provisional or high-confidence rather than canonical.

And the larger Unified Model retains substantive open questions concerning scope, notation, individuation, resource boundaries, and other domain-specific issues.

Those questions should not be closed merely to make the architecture appear finished.

The frozen Unified Model itself was designed with this distinction in mind: architecture can be stable enough to use while research questions remain open.

The same standard governs this chapter.

3.10.5 A Framework for an Unfinished Transformation

The chapter’s argument can now be stated compactly. Gradual AGI is an Ongoing, Unfinished Transformation, so its theory needs an architecture that can organize knowledge without pretending that present knowledge is final. The five plateaus separate logical kinds; facets preserve independently needed record information; typed relations connect without flattening; falsifiability and localization make failure informative; dependency tracking and versioning govern propagation; and anti-rescue discipline prevents revision from erasing what failed.

Together, these elements provide a framework open enough to encounter what it did not anticipate, flexible enough to change when necessary, and falsifiable enough to discover where it is wrong. Its aim is not theoretical permanence but disciplined continuity under revision.

The architecture should be stable enough to use, exposed enough to fail, and repairable enough to learn from failure.

Chapter 4 turns from that architecture to the substantive problem it organizes: how potential becomes realized transformation.

References

Intellectual antecedents and methodological references

Bateson, Gregory. 1949. “Bali: The Value System of a Steady State.” In Social Structure: Studies Presented to A. R. Radcliffe-Brown, edited by Meyer Fortes, 35–53. Oxford: Clarendon Press. Reprinted in Steps to an Ecology of Mind (1972).

Deleuze, Gilles, and Félix Guattari. 1987. A Thousand Plateaus: Capitalism and Schizophrenia. Translated by Brian Massumi. Minneapolis: University of Minnesota Press. Originally published as Mille plateaux (1980).

Kuhn, Thomas S. 1962. The Structure of Scientific Revolutions. Chicago: University of Chicago Press.

Lakatos, Imre. 1970. “Falsification and the Methodology of Scientific Research Programmes.” In Criticism and the Growth of Knowledge, edited by Imre Lakatos and Alan Musgrave, 91–195. Cambridge: Cambridge University Press.

Laudan, Larry. 1977. Progress and Its Problems: Towards a Theory of Scientific Growth. Berkeley: University of California Press.

Morgan, Mary S., and Margaret Morrison, eds. 1999. Models as Mediators: Perspectives on Natural and Social Science. Cambridge: Cambridge University Press.

Popper, Karl R. 1959. The Logic of Scientific Discovery. London: Hutchinson.

Primary corpus — Gradual AGI Series

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/

Companion Graduality catalogs

W.H.L., GPT-5.6 Sol, & Claude Opus 5. 2026. “What ‘Gradual’ Means in Gradual AGI: A Feature Catalog.” Champaign Magazine, August 17, 2026. https://champaignmagazine.com/2026/08/17/what-gradual-means-in-gradual-agi-a-feature-catalog/

W.H.L., GPT-5.6 Sol, & Claude Opus 5. 2026. “Graduality: A Feature Catalog of Gradual AGI.” Champaign Magazine, August 18, 2026. https://champaignmagazine.com/2026/08/18/graduality-a-feature-catalog-of-gradual-agi/