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

What “Gradual” Means in Gradual AGI: A Feature Catalog

Publication Version · v1.0

Abstract

Across eight installments, the theory surrounding gradual realization became richer while the minimum meaning of gradual became thinner. This paper reconstructs the vocabulary through which that development occurred and separates two uses of the term. In the phenomenal sense, once the relevant transformation has been fixed, gradual means only that realization is not exhausted by one event; slow, smooth, continuous, or small-step change is not required. The surrounding features explain, structure, constrain, or qualify that extended realization rather than forming a checklist. Across four selected source passages, Potential–Realization Separation (G002) appears in three and is the only shared positive structure in the two whose transformation scopes do not already presuppose extended realization; that smaller result remains an n=2, single-analyst reconstruction rather than an exhaustive or independently coded analysis of the series. In the principled sense, graduality names information-responsive, revisable action under uncertainty. The reconstruction also identifies a prior individuation problem: before asking whether change is gradual, one must specify which change counts as the transformation. The series makes that choice within particular arguments but supplies no domain-general rule. The nine records are therefore a current, revisable catalog rather than a closed definition.

1. Why This Catalog Exists

“Gradual” already carries a practical association outside AI. In Fabian political gradualism, social change was pursued through cumulative reform rather than revolutionary rupture. Carried into an AI title, that familiar sense can suggest a corresponding posture: if AGI is gradual, perhaps urgency is misplaced and the sensible response is to wait, observe, and avoid premature intervention.

That inference is precisely what this series rejects. Gradual first describes how a transformation is realized, not how slowly anyone should respond to it. And when graduality later becomes a principle for action, its point is revisability—acting in ways that allow later evidence to change later decisions—not delay for its own sake. The case for further iteration can weaken when waiting is itself costly.

The confusion persists because the series uses a familiar word for an unfamiliar combination of claims. Across eight installments, it distinguishes what a system can do from how far that capability has been realized in institutions and society; describes different parts of the transformation as moving on different clocks; allows different populations and domains to occupy different states at the same time; and eventually treats the relevant state as multidimensional rather than something that must lie on a single line. In that account, a transformation can be gradual even when important events inside it are sudden.

The vocabulary also accumulated before its relationships were collected in one place. A reader who remembers Ceiling and Floor may not know that the same distinction later appears as potential and realization; someone who encounters “gradual does not mean smooth” may wonder what, then, gradual is supposed to mean; and the later move from description to a principle for action can make two distinct claims look like one.

This catalog reconstructs that vocabulary as nine feature records. Its purpose is not to replace the eight installments with a new definition, nor to make every recurring idea a necessary condition of Gradual AGI. It is to make the conceptual structure visible: what each feature says, what it does not say, where it entered the series, how it relates to the others, and how firmly that relationship is established.

The result is a guide to the word gradual as the series actually came to use it—rather than as the word might first be heard.

2. Two Meanings

One source of confusion in the phrase “Gradual AGI” is that the series eventually uses the same word family for two different kinds of claim. One is descriptive: it concerns how a transformation is realized. The other is practical: it concerns how to act while important facts are still being learned. They are connected, but neither implies the other.

2.1 Gradual as a description

Consider a medical AI system that acquires, in one model release, a capability that had not previously existed. If the question is when the system acquired that capability, the answer may be a single event. Nothing in Gradual AGI requires that event itself to have arrived slowly.

Now change the question. Suppose the transformation of interest is the realization of that capability throughout medicine: clinical validation, regulatory acceptance, integration with hospital records, professional training, reimbursement, liability, patient access, and the reorganization of everyday practice around the new tool. The same model release is still important, but it is no longer the whole transformation. It is one event inside a larger history.

That distinction is the starting point for the descriptive sense of gradual. Once the transformation under discussion has been fixed, its realization is gradual when it is not exhausted by one event. A sudden breakthrough can therefore sit inside a gradual transformation without contradiction.

This is deliberately a thin claim. It does not by itself tell us how fast the larger transformation moves, whether its path is smooth, whether each domain changes in the same way, or whether progress can be summarized on one scale. Those questions are taken up by the feature records that follow. The descriptive core says less: one event does not exhaust the transformation being considered.

The phrase “being considered” matters. The same history can look abrupt at one level and extended at another because the question has changed. The catalog will return to that problem after the feature records, once the reader has seen how potential and realization, heterogeneous adoption, and multidimensional structure make it concrete.

2.2 Graduality as a principle for action

The series later gives graduality a second meaning. In the Optimization installment, graduality is not a prediction about how the world will change. It is a principle for acting when knowledge is incomplete, conditions are changing, and later evidence may matter.

The basic idea is iterative revisability. Act rather than wait for perfect information, observe what happens, learn from the result, and preserve enough room for what is learned to change what happens next. Repetition alone is not enough. A process can repeat the same step indefinitely without learning anything. What matters is that later information can alter later action, and that earlier decisions have not already made meaningful revision impossible.

This use appears earlier in descriptive form. Synchronization portrays capability and society as adjusting to one another through successive rounds of reciprocal learning. Optimization turns the same iterative structure into a recommendation about decision-making under uncertainty. The structure is similar; the claim is different.

A transformation can unfold gradually without anyone wisely governing it.

Conversely, a decision process can be carefully revisable while the external change it confronts is abrupt. A fast process can preserve options; a slow one can be locked in from the beginning. That is why the descriptive and principled meanings must remain separate.

The catalog therefore uses two labels. Phenomenal gradualness concerns the form taken by the transformation. Principled graduality concerns the way action is organized under uncertainty. Most of the nine features bear on the first. Iterative Revisability is the exception: it appears in the series on both sides of the distinction.

3. Eight Installments, One Developing Vocabulary

The catalog did not appear fully formed. Its vocabulary accumulated as the series moved from epistemic extension and resource abundance to synchronization, optimization, and governance. The table reverses the catalog’s usual view: instead of asking where a feature developed, it asks what each installment added or put to work. It is derived from the feature records and their application notes, not maintained as a separate account.

Installment	Features	What it contributed
#1  Epistemic Extension	G001 · G004 · G007 · G009	Rejects a single arrival event; allows uneven, domain-specific development; introduces Depth and Width; and describes Cumulative Insight as compounding over time.
#2  Abundant Resources	G002 · G003 · G004 · G006	Separates cognitive capability from societal realization; contrasts technological and infrastructural clocks; treats abundance as distributed; and allows discontinuous cognitive emergence within gradual societal realization.
#3  Synchronization	G002 · G003 · G005	Turns potential and realization into Ceiling and Floor, makes their relative movement explicit through the Slope, and describes adoption as successive cycles of reciprocal learning.
#4  Ceiling, Floor, and Slope	G002 · G003 · G004 · G008	Formalizes capability-realization dynamics, heterogeneous thresholds, and assimilation rates; separates genuine convergence from measurement-window, population-mixture, and benchmark-ceiling effects.
#5  Optimization framework	G005 · G006 · G007	Recasts graduality as iterative refinement under preserved adaptability, allows drift and punctuated change, and generalizes evaluation toward multidimensional optimization.
#6  Optimization empirical tests	G004 · G005 · G007	Applies the earlier features under formal and empirical contact: heterogeneous multi-agent optimization, iteration under non-stationarity, and scalar-to-tensor or multi-objective structure.
#7  Governance framework	G001 · G002 · G003 · G005 · G006 · G007	Makes the phenomenal/principled distinction explicit, defines the governed object through potential and realization, carries relative rate differences into piecewise realization, derives a partial-order view, and treats AGI as expansion across domains.
#8  Governance measurement	G005 · G007 · G008	Uses a vector-valued governance profile, separates observables from underlying properties, shows representation changing under measurement, and applies revision as an ongoing mode of governance.

Seen this way, the vocabulary does not grow by repeatedly redefining one word. The early papers introduce distinctions and recurring patterns; the middle papers formalize or operationalize them; the later papers clarify what is being described, what can be measured, and when graduality becomes a principle for action. Some installments therefore introduce a feature, while others mainly apply one that already exists.

That history matters for how the catalog should be read. A familiar term may disappear while its underlying feature survives under another name, and a later application need not represent a new meaning. The nine records that follow collect those continuities without treating every reuse as a conceptual revision.

4. How to Read a Feature Card

Each card can be read in sequence or used as a stand-alone reference. Its fields answer four practical questions: what is this feature, what nearby reading does it rule out, where did the series use it, and how firmly is the reconstruction supported?

Name, ID, and aliases

The reader-facing name is paired with a stable G-number. “Also known as” routes older wording to the same feature; it is not a list of loose associations.

Kind and bears on

“Kind” says what work the feature does—what is being explained, an explanation or constraint, a question of observation, or a guide to action. “Bears on” keeps phenomenal gradualness separate from principled graduality.

What it doesn’t mean

This is not a caveat field. It carries the catalog’s corrective guardrails: G003 rules out absolute slowness as the meaning of gradual; G006 rules out smoothness as a requirement. These boundaries are part of the feature.

Where it appears

This is a selective genealogy: first clear appearance, material change, and, when useful, mature form. Mere application need not appear in the visible trail.

Related features

These links distinguish relations such as “combines with,” “distinct from,” and “helps explain.” A dagger (†) marks a connection that depends materially on the present reconstruction rather than the source’s own wording.

How firmly established

This separates direct source claims from cross-installment reconstruction, which may carry a different level of confidence.

“Remember this” is not another definition. It is the shortest faithful sentence for readers who do not need the technical backplane.

Feature and Term Index

Use this page either by feature name or by a term remembered from the series. The entries appear in the same reading order as the catalog; the full alias and related-term lookup appears in the back matter.

ID	Feature	In one sentence	Also known as
G001	Extended Realization	The transformation is realized across more than one event rather than being exhausted by a single arrival.	AGI as process rather than event; no single arrival moment
G002	Potential–Realization Separation	What has become possible and what has actually taken hold in the world are two different things.	Ceiling and Floor; potential and realization; cognitive versus societal emergence; growth versus spread
G003	Relative Rates and Lag	Parts of a transformation can run on different clocks, and what matters is their relation rather than either speed alone.	assimilation rate; mismatched clocks; synchronization rate; software speed versus infrastructure speed
G004	Heterogeneous Realization	Different people, institutions, places, and domains can realize the same capability to different extents at the same time.	distributed abundance; heterogeneous thresholds; coexisting regimes
G007	Multidimensionality and Partial Order	Progress need not lie on one scale: some states can be ordered while others differ across dimensions without either being simply ahead.	Depth and Width; holistic multidimensional evaluation; scalar-to-tensor; governance profile
G009	Accumulation and Compounding	Contributions can persist and build on one another rather than each replacing the last.	Cumulative Insight; compounding over time
G006	Local Discontinuity	A gradual transformation may contain abrupt jumps; smoothness is permitted, never required.	drift versus punctuated shift; gradual does not mean smooth
G008	Observational Partiality	What we can measure may capture only part of the underlying state, and may not identify it at all.	measurement-dependent convergence; observable versus latent; recognized variety
G005	Iterative Revisability	Later information can change later action, and early choices leave room for that to matter.	reciprocal learning; mutual learning; iterative refinement; graduality as an optimization principle; preserved adaptability

Quick Reference: Nine Lines to Remember

These are the catalog’s shortest faithful takeaways. They are memory aids, not substitutes for the feature definitions or their evidentiary qualifications.

Feature	Remember this
G001 · Extended Realization	The transformation is larger than any one moment inside it.
G002 · Potential–Realization Separation	A capability can arrive before its consequences do.
G003 · Relative Rates and Lag	What matters is not how fast either side moves, but whether their clocks stay together.
G004 · Heterogeneous Realization	One society can contain several stages of the same transformation at once.
G007 · Multidimensionality and Partial Order	Not every kind of progress fits on one line.
G009 · Accumulation and Compounding	An achievement becomes cumulative when later work can stand on it.
G006 · Local Discontinuity	Jumps do not disqualify it.
G008 · Observational Partiality	A measurement can be accurate about what it counts and still miss what matters.
G005 · Iterative Revisability	Move in a way that lets what you learn change what you do next.

Related terms worth knowing: Slope points to G003 but is a model variable rather than an alias; expanding region points primarily to G001 and also to G007; group-indexed exposure points to G004. The back-matter lookup gives the full term-by-term routing.

5. The Feature Catalog

The nine entries are presented together as one catalog. G001 is different from the records that follow: it names what the phenomenal part of the catalog is trying to explain—extended realization—while the surrounding features help explain, structure, constrain, or qualify it. Their sequence is thematic only, proceeding from Extended Realization through features that structure, qualify, observe, and guide action under it. The ordering is a reading aid, not an additional classification system; each record remains a catalog entry identified by its stable G-number.

5.1 Extended Realization · G001

Compact record — this feature names what the catalog explains.

Canonical label	Extended / non-pointlike realization
Also known as	AGI as process rather than event · no single arrival moment
Kind	What is being explained
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	The transformation is realized across more than one event rather than being exhausted by a single arrival.

What it names

Extended Realization is the simplest idea in the catalog and the one the other phenomenal features largely surround.

A transformation is extended when no single event completes it. Something important may happen suddenly—a new capability, a threshold crossing, an institutional decision—without that event exhausting the larger change being considered.

The first installment already rejects the idea that one breakthrough marks the arrival of AGI. It describes progress as uneven and domain-specific and later treats AGI as a continuing state of affairs rather than a sequence of isolated feats. The Governance installment makes the idea sharper: the object being tracked is an expanding region across domains, realized piece by piece rather than at one moment.

What it doesn’t mean

Extended does not simply mean long-lasting. A long period of waiting followed by one event is still one-event realization.

G001 makes no further claim about the shape of the path between realization events. Its boundary is narrower: the transformation is not exhausted at one point.

Questions of pace, shape, and step size are taken up in G003 and G006.

Where it appears

#1 · Epistemic Extension — §III, “Depth as an Epistemic Continuum,” and §V, “Width.” No single breakthrough constitutes AGI; epistemic progress becomes recurring, persistent, and socially embedded. Read source

#7 · Contestation — §3.1, “Definition 0 — AGI.” The mature formulation treats AGI as an expanding region across domains and derives gradualness from piecewise rather than momentary realization. Read source

How firmly established

The rejection of a singular arrival event is stated directly in the series. The stronger catalog claim that G001 is what the other phenomenal features help explain, rather than simply another feature beside them, is a cross-series reconstruction and has not yet been independently checked.

Remember this: The transformation is larger than any one moment inside it.

5.2 Potential–Realization Separation · G002

Canonical label	Potential–realization separation
Also known as	Ceiling and Floor (#3, #4) · potential and realization (#7) · cognitive versus societal emergence (#2) · growth versus spread (#2)
Kind	Explains
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	What has become possible and what has actually taken hold in the world are two different things.

The question

Why can a new capability appear almost overnight while its consequences take years, or decades, to arrive?

The idea

Something becomes possible at the moment it can be done. It becomes realized only when it has entered the world: people can reach it, know when to rely on it, fit it into existing work, adjust rules around it, and see its effects in what actually happens.

These are two states, not one state seen twice. The series keeps them apart deliberately and treats the passage between them as having its own structure—a mapping that can be fast or slow, even or uneven, and that institutions can act upon.

Either half is easy to overlook: a demonstration can be mistaken for realized change, while limited use can be mistaken for absent capability.

Why it matters

Without this separation, an advance and its consequences collapse into a single event, leaving no place to ask why access is unequal, institutions adapt at different rates, or a capability reaches one domain long before another. It also explains how a capability can arrive abruptly while the transformation around it remains extended: the abruptness is real, and so is the extension.

What it doesn’t mean

The separation itself says nothing about the speed or direction of the mapping. Realization may track capability closely or lag behind it, and the gap may narrow, widen, or persist.

Examples

Medicine. Suppose a system becomes able to produce diagnoses at the level of an experienced specialist. The capability might appear within one model release. Its realization still needs clinical validation, regulatory approval, integration with hospital records, liability rules, training, reimbursement, and enough public confidence that patients accept it. The capability came first. What followed was a separate process with obstacles of its own.

Shipping containers. Standardized containers made a new way of moving freight possible, but realizing that possibility required ports, cranes, ships, rail links, customs procedures, contracts, and labor practices to change around them. The core capability became available well before world trade was reorganized to exploit it.

Contrast. If possibility and realization were the same thing by definition, the first successful demonstration would also be the completed transformation. There would be no gap to study, no adoption to explain, and nothing for policy to shape in between.

Where it appears

#2 · Abundant Resources — §6, “Cognitive Emergence and Societal Emergence.” Introduces the split between the creation of advanced capability and its integration into civilization. Read source

#3 · Synchronization — §3.2–§3.4. Turns the split into a moving relationship: the Ceiling of what is possible, the Floor of what is realized, and the relation between them. Read source

#4 · Ceiling, Floor, and Slope — §2. Formalizes Ceiling and Floor as distinct dynamical quantities and makes their gap an object of analysis. Read source

#7 · Contestation — §3.2, “Definition 1 — The governed object.” Generalizes the distinction into capability potential, realized outcomes, and the mapping between them. Read source

Related features

Enables → Relative Rates and Lag (G003). Two clocks can only be compared once there are two distinct things to compare.

Combines with → Heterogeneous Realization (G004). The gap between possibility and realization can differ across populations and domains.

Helps explain → Extended Realization (G001). The gap is one reason a transformation can outlast the moment a capability appears.

Distinct from → Multidimensionality and Partial Order (G007). G002 separates potential from realization; G007 concerns how a state is structured and compared.

How firmly established

The distinction is stated directly across four installments, and an independent check confirmed the relevant relation in #2. In the present reconstruction, G002 appears in three of four selected source passages and is the only shared positive structure in the two whose transformation scopes do not already presuppose extended realization. That smaller result is based on a single analyst’s cross-series reading and has not yet been independently rerun. These passage-level counts are not an exhaustive analysis of every argument in the series.

Remember this: A capability can arrive before its consequences do.

5.3 Relative Rates and Lag · G003

Canonical label	Relative-rate / temporal asymmetry
Also known as	assimilation rate (#4) · mismatched clocks (#2) · synchronization rate (#3, #4) · software speed versus infrastructure speed (#2)
Kind	Explains
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	Parts of a transformation can run on different clocks, and what matters is their relation rather than either speed alone.

The question

If Gradual AGI does not mean slow AI progress, where does time enter the idea?

The idea

Two things can both move quickly and still fall out of step. They can both move slowly and remain closely synchronized. That is the point of Relative Rates and Lag.

The series first puts the intuition plainly: intelligence can scale at software speed while civilization scales at infrastructure speed. Synchronization then names the relationship through the Ceiling, Floor, and Slope: frontier capability may advance faster than institutions and society can assimilate it. CFS turns the same idea into a dynamical model in which realized state changes relative to a moving capability frontier.

The important word is relative. A gradual transformation need not be absolutely slow. Its components may simply move on different clocks.

Why it matters

Once capability and realization move at different rates, a gap appears. That gap can widen, narrow, stabilize, or vary across domains.

This shifts the central question from “How fast is AI improving?” to “How fast are the other parts of the transformation moving relative to it?” That is often the more useful question for institutions, infrastructure, adoption, and governance.

What it doesn’t mean

Lag is not inevitable. Two processes may remain closely synchronized.

Nor does a slow process automatically count as gradual. Absolute slowness says nothing about whether the relevant transformation is exhausted by one event.

Slope is not another name for this feature. It is one model variable used to represent the relative-rate relationship.

Examples

Clinical adoption. A diagnostic system improves every few months while hospitals revise procedures every few years. Both are changing; their clocks differ.

Housing. A city can add residents faster than it builds homes. Neither population growth nor construction needs to be slow for a housing gap to widen. What matters is their relative rate.

Contrast. If capability and realization rise together at roughly the same rate, there may be little persistent lag even if both are changing rapidly.

Where it appears

#2 · Abundant Resources — §7, “Hard Takeoff and the Limits of Tech Determinism.” Contrasts software-speed capability change with infrastructure-speed civilizational change. Read source

#3 · Synchronization — §3.3, “The Slope.” Makes relative movement an explicit relation between frontier capability and realized assimilation. Read source

#4 · Ceiling, Floor, and Slope — §2. Formalizes the relationship as a non-constant synchronization rate coupled to the remaining gap. Read source

Related features

Depends on → Potential–Realization Separation (G002). Relative rates require distinct quantities whose movement can be compared.

Combines with → Heterogeneous Realization (G004). Different populations can have different rates and therefore different lags.

Helps explain → Extended Realization (G001). In Synchronization and CFS, differing rates directly account for realized transformation continuing beyond a capability event.

How firmly established

The relative-rate idea is stated directly and formalized across #2–#4. An independent check confirmed its use in #4. The broader conclusion that G003 is load-bearing in two of four selected source passages remains part of the present single-analyst reconstruction.

Remember this: What matters is not how fast either side moves, but whether their clocks stay together.

5.4 Heterogeneous Realization · G004

Canonical label	Heterogeneous realization
Also known as	distributed abundance (#2) · heterogeneous thresholds (#4) · coexisting regimes (#4)
Kind	Explains
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	Different people, institutions, places, and domains can realize the same capability to different extents at the same time.

The question

Why can one part of society look transformed while another appears almost untouched?

The idea

Realization does not have to happen uniformly.

A capability may be commonplace in one profession and rare in another; routine in one institution and prohibited in another; readily available to one population and difficult for another to reach.

The first installment already allows progress to be uneven and domain-specific. Abundant Resources shifts the idea into distribution: a capability may exist without being ubiquitous, accessible, or affordable across society. CFS then makes heterogeneity explicit by giving different adopting populations different thresholds for when a capability becomes usable or worthwhile. Several realization states can therefore coexist.

Why it matters

Averages can hide the transformation they are meant to describe. If half a population has reorganized around a capability and the other half has scarcely encountered it, an average can make both groups look moderately transformed when neither is.

Heterogeneity therefore matters whenever we ask who has actually realized a capability, where, and under what conditions.

What it doesn’t mean

Heterogeneity does not by itself make a transformation gradual. Different groups can occupy different states at a single moment.

It also does not imply that the groups will eventually converge, or that one group is necessarily ahead in every relevant respect.

Examples

Workplaces. A large company may integrate an AI tool into everyday operations while a small firm in the same industry still lacks the data, staff, or systems required to use it.

Payments. A new payment method can become routine among urban retailers while remaining uncommon among businesses with poor connectivity or older equipment.

Contrast. If every relevant group adopted a capability at the same time and to the same degree, heterogeneous realization would disappear even though the overall transformation might still be extended.

Where it appears

#1 · Epistemic Extension — §III, “Depth as an Epistemic Continuum.” Progress is explicitly allowed to be uneven and domain-specific. Read source

#2 · Abundant Resources — §4.1–§4.3, “Ubiquity, Accessibility, Affordability.” Realization becomes a question of distributed availability rather than mere existence of capability. Read source

#4 · Ceiling, Floor, and Slope — §2.2, “Threshold-gated coupling, generalized to heterogeneous thresholds.” Different populations receive distinct adoption thresholds, producing coexisting realization regimes. Read source

Related features

Combines with → Potential–Realization Separation (G002). Different groups can sit at different distances from the same capability potential.

Combines with → Multidimensionality and Partial Order (G007). Realization may differ across both groups and dimensions.

Can be obscured by → Observational Partiality (G008). Aggregating unlike populations can make several regimes look like one smooth trend.

Distinct from → Extended Realization (G001). Heterogeneity can exist in one cross-section; it does not by itself establish temporal extension.

How firmly established

Heterogeneous realization is stated directly in several installments. Independent review led to a narrower definition of what belongs in the feature. In the present reconstruction, G004 is not a load-bearing premise in any of the four selected source passages, so the catalog does not treat it as a universal explanation of extended realization.

Remember this: One society can contain several stages of the same transformation at once.

5.5 Multidimensionality and Partial Order · G007

Canonical label	Multidimensionality / partial order
Also known as	Depth and Width (#1) · holistic multidimensional evaluation (#5) · scalar-to-tensor (#6) · governance profile (#8)
Kind	Explains
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	Progress need not lie on one scale: some states can be ordered while others differ across dimensions without either being simply ahead.

The question

What does it mean to say that AGI has advanced if different capabilities or outcomes move in different directions?

The idea

Some comparisons are easy because they use one scale. Ten kilometres is farther than five. Many important comparisons are not like that.

One AI system may be stronger at mathematics while another is more reliable at planning. One social outcome may improve access while worsening concentration. Unless we have a defensible rule for combining those differences into one number, saying that one state is simply higher than the other hides information.

A partial order allows some comparisons without forcing all of them. One state may clearly exceed another across the relevant dimensions; two others may remain incomparable because each is stronger in a different respect. The series moves toward this idea in stages: Depth and Width in #1, multidimensional evaluation in #5, and an explicit partial order in Governance.

Why it matters

A single scale invites a single crossing point: below the AGI line, then above it. A multidimensional structure permits progress across different domains without producing one universally meaningful moment of arrival. In Governance, this becomes an expanding region of domains rather than a dated threshold event, with gradualness derived from the piecewise expansion of that region.

What it doesn’t mean

It does not mean that nothing can be compared, or that every dimension matters equally.

G007 therefore rejects the assumption that AGI progress must be representable by a single number or crossed at a single universal threshold.

Nor does the series claim that capability is inherently incapable of scalar representation. Its narrower claim is that no defensible aggregation rule has been supplied for the use at hand; if one were supplied, the analysis would have to change.

Examples

Students. One student excels in mathematics and another in writing. Without first deciding how those skills should be weighted, “Which student is better?” has no single neutral answer.

AI systems. One system may be better at scientific reasoning while another is more dependable at long-horizon planning. A single ranking requires an aggregation rule, not merely two sets of scores.

Contrast. If every relevant dimension could be defensibly reduced to one common measure, a total ranking would become possible and much of the partial-order argument would no longer be needed.

Where it appears

#1 · Epistemic Extension — §IV, “Depth and Width as Orthogonal Axes.” Introduces two distinct dimensions rather than one AGI scale. Read source

#5 · Optimization — §3.6, “Holistic, Multidimensional Evaluation.” Explicitly rejects a single scalar metric for civilizational evaluation. Read source

#7 · Contestation — §3.1, “Definition 0 — AGI.” Makes partial ordering explicit and characterizes AGI as expansion across a region of domains. Read source

Related features

Combines with → Heterogeneous Realization (G004). Different groups may occupy different positions in a multidimensional space.

Observed through → Observational Partiality (G008). A summary measure may fail to preserve distinctions in the underlying state.

Individuates → Extended Realization (G001)†. In Governance, the partial-order structure helps fix the relevant transformation as region expansion rather than one local capability crossing. The source directly derives gradualness from the expanding region; describing this as individuation work is the reconstruction’s interpretation.

† This connection depends materially on the current unreproduced interpretive reconstruction; the source-level relationship itself is stated more narrowly.

How firmly established

The non-scalar and partial-order claims are stated directly in the source series. An independent check confirmed the relevant use in #7. The stronger claim that G007 helps fix what counts as the transformation is a single-analyst cross-series reconstruction and has not yet been independently checked.

Remember this: Not every kind of progress fits on one line.

5.6 Accumulation and Compounding · G009

Canonical label	Accumulation / compounding
Also known as	Cumulative Insight (#1) · compounding over time
Kind	Explains
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	Contributions can persist and build on one another rather than each replacing the last.

The question

What turns a sequence of impressive achievements into continuing development?

The idea

Not everything that happens repeatedly accumulates.

A system may solve one difficult problem after another while each result remains isolated. Accumulation begins when earlier contributions persist, interact with later ones, and alter what can be done next.

Epistemic Extension calls this Cumulative Insight. A contribution matters not simply because it is impressive, but because later reasoning can reuse it, extend it, or build on the structure it created. The present is partly made out of what earlier stages left behind.

Why it matters

If achievements disappear after each episode, repeated success need not amount to durable development.

Compounding changes that. A result becomes a method; the method supports another result; several results reorganize a field. What persists becomes part of the conditions for what follows. This is one way extended realization can acquire depth, but it is not required by the minimum meaning of gradual.

What it doesn’t mean

Accumulation is not simply more. Ten isolated achievements need not compound.

Nor does accumulation tell us how large or regular the individual gains must be; its concern is whether earlier contributions persist into later ones.

And it is not the same as Iterative Revisability. G009 concerns whether earlier contributions persist and build; G005 concerns whether later action changes in response to later information.

Examples

Research. One result settles a useful lemma; later researchers reuse it to prove broader results. The first contribution has become part of the structure of later work.

Learning. A person who retains and combines earlier skills can use them to learn something more difficult. Repeating lessons that are forgotten after each session produces activity without much accumulation.

Contrast. Imagine a system that repeatedly produces excellent answers, but every answer disappears and influences nothing later. The achievements are successive, but they do not compound.

Where it appears

#1 · Epistemic Extension — §III.3, “Cumulative Insight: Compounding Over Time.” Depth increases when contributions persist, interact, and form a basis for later reasoning rather than remaining isolated achievements. Read source

Related features

Distinct from → Iterative Revisability (G005). Accumulation concerns persistence; revisability concerns responsiveness to new information.

Distinct from → Extended Realization (G001). A transformation can extend across several events even if later states replace rather than accumulate earlier ones.

How firmly established

Cumulative Insight is directly stated in #1. The current reconstruction first treated accumulation as a mechanism and later restored it as a feature because the source makes compounding part of the epistemic state itself. G009 does not appear as an isolated load-bearing premise in the selected source arguments, and the series does not make accumulation necessary to gradual realization.

Remember this: An achievement becomes cumulative when later work can stand on it.

5.7 Local Discontinuity · G006

Compact record — this feature constrains what an acceptable account may require.

Canonical label	Local discontinuity compatible with global graduality
Also known as	drift versus punctuated shift (#5) · “gradual does not mean smooth” (#7)
Kind	Constrains
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	A gradual transformation may contain abrupt jumps; smoothness is permitted, never required.

What it rules out

Any account of gradual realization that makes smoothness, continuity, or small step size a condition of the term.

The commitment is negative and precise: gradual does not entail smooth. It does not entail the opposite either. A larger transformation may unfold through a handful of large jumps and still be gradual if its realization is not exhausted by any one of them.

This matters because the assumption is nearly automatic. A reader carrying it into the series may see a contradiction in calling a transformation gradual while allowing sudden capability releases. The apparent contradiction disappears once the local jump and the larger transformation are distinguished.

What it doesn’t mean

It does not say that abrupt change is itself gradual, or that every gradual transformation must contain jumps. It says only that local discontinuity does not rule out global graduality.

Example

Example. Suppose an AI system crosses a major medical-reasoning threshold in one release. That local change may be sudden. If the transformation being tracked includes adoption across specialties, hospitals, regulators, insurers, and patients, however, the one jump need not exhaust it.

Where it appears

#2 · Abundant Resources — §7, “Hard Takeoff and the Limits of Tech Determinism.” Introduces the compatibility claim explicitly: cognitive emergence may be discontinuous while societal emergence remains gradual. Read source

#5 · Optimization — §3.3, Proposition 2. Treats gradual drift and punctuated shift as two possible forms of development rather than as a gradual/non-gradual divide. Read source

#7 · Contestation — §3.1, “Definition 0 — AGI.” States the mature guardrail: a discontinuous capability jump is compatible with the definition, and gradualness does not require smoothness. Read source

Related features

Constrains → Extended Realization (G001). This is the point of the record: extended realization requires more than one realization event, not a smooth path between them.

How firmly established

The compatibility claim is stated directly in #2, #5, and #7. Treating it as a Constraint rather than a positive feature is the catalog’s reconstruction of the role that claim plays.

Remember this: Jumps do not disqualify it.

5.8 Observational Partiality · G008

Canonical label	Observational partiality
Also known as	measurement-dependent convergence (#4) · observable versus latent (#8) · recognized variety (#8)
Kind	How it is known
Bears on	Phenomenal gradualness
Technical basis	Graduality technical catalog v0.8
In one sentence	What we can measure may capture only part of the underlying state, and may not identify it at all.

The question

If gradual realization is complex, how much can we infer from the indicators we happen to observe?

The idea

A measurement can be perfectly accurate about what it measures and still be incomplete about the thing we care about.

CFS gives an early example. Apparent convergence can be produced by the chosen measurement window, by combining populations in different regimes, or by a bounded benchmark approaching its own ceiling rather than by genuine synchronization in the underlying process.

The governance measurement paper makes the distinction harder to ignore. A count of recognized governance categories can fall while specificity and stringency rise. The lesson is not that measurement fails. It is that an observable and the underlying theoretical property are not automatically the same thing.

Why it matters

Without this distinction, changes in an instrument can be mistaken for changes in the world.

A flat benchmark may conceal continued improvement. An average may conceal population differences. A governance count may miss changes in precision, commitment, or transparency. Good measurement therefore requires knowing not only the number, but what relation that number bears to the state we are trying to understand.

What it doesn’t mean

Observational partiality is not a claim that reality is unknowable, or that measurements are arbitrary or useless.

It says only that the connection between an observable and an underlying state must be established rather than assumed.

Examples

AI benchmarks. A system approaches the maximum score on a fixed test. The score changes less and less, but that does not by itself show that the underlying capability has stopped changing.

Education. An exam can measure performance on the questions it contains while missing collaboration, judgment, creativity, or durable understanding. The score may be valid without being complete.

Contrast. If an observable mapped uniquely and without information loss to the underlying property, observational partiality would disappear for that measurement.

Where it appears

#4 · Ceiling, Floor, and Slope — §2.5, methodological caveat on apparent convergence. Separates measurement-window effects, population mixture, and benchmark ceilings from genuine underlying convergence. Read source

#8 · Measuring Governance Under Empirical Contact — §4.7 and §4.13. Shows recognized variety moving independently of specificity and stringency and concludes that the minimum observable governance profile is vector-valued. Read source

Related features

Can obscure → Heterogeneous Realization (G004). Aggregation may hide distinct population states.

Qualifies evidence about → Relative Rates and Lag (G003). A measured rate depends on appropriate windows, populations, and proxies.

Distinct from → Multidimensionality and Partial Order (G007). G007 concerns the structure of the underlying state; G008 concerns what survives when that state is observed or represented.

How firmly established

The distinction between an observable and the underlying state is explicit in #4 and #8. An independent check confirmed the relevant use in #8. Later review narrowed G008 to the observation relation itself rather than treating a particular state-space structure as part of the feature.

Remember this: A measurement can be accurate about what it counts and still miss what matters.

5.9 Iterative Revisability · G005

Canonical label	Iterative revisability
Also known as	reciprocal learning (#3) · mutual learning (#3) · iterative refinement (#5) · graduality as an optimization principle (#5) · preserved adaptability (#5)
Kind	Explains · how to act
Bears on	Phenomenal gradualness and principled graduality
Technical basis	Graduality technical catalog v0.8
In one sentence	Later information can change later action, and early choices leave room for that to matter.

The question

How should anyone proceed when both the technology and their understanding of its effects are still changing?

The idea

Not every important decision can be made once and left alone. When knowledge is incomplete and conditions are still moving, a process can run in rounds: act, watch what happens, learn, revise, act again.

Repetition alone is not enough. A process that repeats the same step regardless of what it observes is iterative and nothing more. What matters is that new information can change what comes next—and that earlier choices have not already foreclosed the possibility.

This feature appears in the series in two distinct ways. In Synchronization it describes something: capability and society adjust to one another in successive rounds, each shaping what the other does next. In Optimization it becomes a recommendation: under real uncertainty, proceed so that evidence arriving later can still make a difference.

The same structure, doing two different jobs. That makes G005 the only feature in the catalog that bears on both meanings of graduality—how a transformation unfolds, and how one ought to act while it does. A transformation can unfold gradually without anyone wisely governing it.

Why it matters

Large transitions are rarely governed with complete information. Policies reveal costs after deployment; practices improve as people learn where a tool is reliable; new capabilities can change the problem itself. Iterative revisability makes that learning part of the process rather than something that arrives after commitments have hardened.

What it doesn’t mean

It does not mean going slowly. Revisions can come quickly; a fast process can be revisable, and a slow one can be locked in from the first step.

It does not mean hesitation. Revision requires something to revise.

It does not require every consequence to be reversible—only that the process keeps meaningful room to change course.

And the case for it is conditional, not universal. Where a problem is well understood and stable, or where waiting is itself costly, the argument for further iteration weakens. The practical question is therefore not whether more information would be useful in the abstract, but whether its expected value exceeds the cost of delaying action while preserving opportunities to revise afterward.

Examples

Fisheries. A fishing quota cannot be set once with certainty; estimates of the stock change as new evidence arrives. A quota can be set, the stock surveyed, the estimate corrected, and the next quota changed. What matters is that later evidence can still alter later action.

Building codes. Seismic standards were not derived from theory alone. Major earthquakes revealed which assumptions failed, and codes changed around those failures. Revision is part of the method.

Contrast. An organization makes one irreversible commitment before observing any consequence and afterwards cannot meaningfully change course. Time still passes, but nothing learned along the way can alter the outcome.

Where it appears

#3 · Synchronization — §4.2, “Reciprocal Learning.” Describes transformative adoption as successive cycles of mutual adjustment and characterizes Gradual AGI as an iterative process of mutual learning. Read source

#5 · Optimization — §3.1 and §5.1. Defines graduality as iterative refinement through observation, learning, and feedback, explicitly not as a measure of speed, and states the conditions under which the principle is worth adopting. Read source

#7 · Contestation — §3.1. States plainly that gradual as a description of the phenomenon and graduality as a principle of action are different claims. Read source

Related features

Combines with → Relative Rates and Lag (G003). Repeated adjustment matters especially when different parts of a system are moving on different clocks.

Distinct from → Accumulation and Compounding (G009). Iteration concerns whether later steps respond to what was learned; accumulation concerns whether earlier contributions persist.

Distinct from → Extended Realization (G001). A transformation can be extended with no learning in it at all; a fixed schedule unfolding over decades is not thereby revisable.

Tolerates → Local Discontinuity (G006). Revisability and the size of any one move are separate questions.

How firmly established

The two source uses are stated directly. An independent check confirmed the iterative process in #3. The stronger reading of #5—that revisability operates both as a process structure and as a decision principle—remains based on a single analyst and has not yet been independently checked.

Remember this: Move in a way that lets what you learn change what you do next.

6. What the Catalog Shows

The nine records do not add up to a nine-part definition. That is the central result of putting them side by side. Some features help explain why realization can extend beyond one event; some describe the structure of that realization; one rules out a common assumption about smoothness; one concerns what survives observation; and one also becomes a principle for action. Their importance lies in the work they do around the idea of gradual realization, not in becoming nine conditions that every gradual transformation must satisfy.

6.1 No single surrounding feature defines “gradual”

The core commitment remains permissive. Once the relevant transformation has been fixed, calling its realization gradual requires only that the transformation not be exhausted by one event. Potential–Realization Separation (G002), Relative Rates and Lag (G003), Heterogeneous Realization (G004), Multidimensionality and Partial Order (G007), and Accumulation and Compounding (G009) can all make that extended realization intelligible in different ways. None is part of the minimum meaning by itself.

The present reconstruction does find an asymmetry among them. Across four selected source passages, Potential–Realization Separation appears in three, and it is the one structure common to the two arguments whose transformation scope does not already build extended realization into the object they examine. In those two clean cases (n=2), G002 therefore does more than happen to be frequent: it performs non-circular explanatory work. But this topology is a reconstruction of four selected passages, not an exhaustive or independently coded analysis of the entire series; the n=2 result remains single-analyst. Nor is G002 necessary: one of the selected passages derives extended realization through multidimensional domain expansion without depending on it.

This is why the catalog becomes richer while the word at its center becomes thinner. The later installments add machinery around gradual realization without steadily adding requirements to the meaning of gradual itself.

6.2 Gradual relative to what?

The cards also reveal a prior question that ordinary uses of gradual often leave implicit: what exactly is the transformation being judged?

Consider one history. An AI system crosses a major medical-reasoning threshold in a single release. If the question is when the system acquired that capability, the relevant transformation is the capability crossing, and it may be sudden. If the question is when that capability became realized across medicine, the transformation includes adoption across specialties, hospitals, regulators, insurers, and patients, and the same history may be gradual. If the question concerns wider civilizational assimilation, the object is broader again. These are not three verdicts about one fixed transformation; they are three differently bounded transformations.

The technical reconstruction calls this choice individuation. The term is less important than the order of reasoning it names: first specify which transformation is under discussion; only then ask whether that transformation is exhausted by one event. Much apparent disagreement about whether AGI is “sudden” or “gradual” can therefore arise before the predicate is even applied.

That freedom is not unlimited. A transformation boundary should be fixed by the research or policy question rather than widened after the answer is visible merely to obtain a gradual verdict. And if the chosen scope already defines the transformation as extended across domains, institutions, or time, that built-in extension cannot then be offered as independent evidence that realization is extended. The technical companion applies both checks to the series’ own arguments—whether the scope has an independent reason for being chosen, and whether extended realization is already built into it—and reports where those conditions hold and where they do not. An embedded scope can still be a legitimate object of analysis; it simply carries less independent evidential weight for the claim that realization is non-pointlike.

The catalog does not supply a universal boundary. The eight installments fix the relevant transformation within particular arguments—epistemic development, societal realization, synchronization, optimization, or governance—and the catalog reconstructs those choices rather than replacing them with a domain-general rule. That limit matters because the minimum predicate is simple only after the object has been fixed. The meaning of gradual can be thin while the classification of a real history remains substantive.

6.3 A current catalog, not a closed catalog

G001–G009 are the nine canonical features recovered by the present reconstruction. The number nine should not be read as a claim that the conceptual space has been exhausted. A later installment, a broader comparison, or a stronger reconstruction could reveal a feature that deserves its own record. Conversely, a future analysis could show that something now treated as a feature is better understood as an alias, mechanism, condition, consequence, or connection of another.

The value of a catalog is therefore not that it closes the vocabulary. It gives the vocabulary a stable enough form to be inspected, used, challenged, and revised without pretending that the present arrangement is the last possible one.

7. Conclusion: Three Things to Remember

First: gradual does not mean slow or smooth.

A gradual transformation can contain rapid change and abrupt local jumps. What matters is not that every step be small or every curve be smooth, but that the transformation being considered is not exhausted by one event. Relative rates and discontinuities belong to the story; neither supplies the definition by itself.

Second: a capability event and a realized transformation are not necessarily the same thing.

Something can become technically possible before it becomes widely accessible, institutionally integrated, socially accepted, or consequential in practice. Different populations and domains can move on different clocks and occupy different states at the same time. This is why a dramatic capability release can be real without being the whole transformation.

Third: graduality also names a way of acting under uncertainty.

In the later installments, the same family of ideas acquires a principled use: act, observe, learn, and preserve enough room for later information to change later decisions. That recommendation is not entailed by the descriptive fact that a transformation is gradual. The descriptive claim and the practical recommendation must therefore remain distinct.

The nine records in this paper are the current reconstruction of that vocabulary, not a claim that the vocabulary is complete. Some relationships are stated directly in the source installments; others are reconstructed across them and carry correspondingly narrower confidence. The detailed coding record, validation history, provenance machinery, admission and retirement rules, and version history belong to the technical companion rather than to this reader-facing catalog.

Back Matter

Alias and Related-Term Index

This index routes terms used across the eight installments to the catalog record that now carries them. An alias is another term for substantially the same catalog feature. A related term is useful for lookup but is a model variable, application, conditioning device, or adjacent formulation rather than another name for the feature.

Term	Routes to	Status / note
AGI as process rather than event	G001 · Extended Realization	Alias
assimilation rate	G003 · Relative Rates and Lag	Alias
Ceiling and Floor	G002 · Potential–Realization Separation	Alias
coexisting regimes	G004 · Heterogeneous Realization	Alias
cognitive versus societal emergence	G002 · Potential–Realization Separation	Alias
compounding over time	G009 · Accumulation and Compounding	Alias
Cumulative Insight	G009 · Accumulation and Compounding	Alias
Depth and Width	G007 · Multidimensionality and Partial Order	Alias
distributed abundance	G004 · Heterogeneous Realization	Alias
drift versus punctuated shift	G006 · Local Discontinuity	Alias
expanding region	G001 · Extended Realization; see also G007	Related term — source formulation links extended realization to multidimensional region expansion
governance profile	G007 · Multidimensionality and Partial Order	Alias
gradual does not mean smooth	G006 · Local Discontinuity	Alias
graduality as an optimization principle	G005 · Iterative Revisability	Alias
group-indexed exposure	G004 · Heterogeneous Realization	Related term — conditioning/indexing device, not an alias
growth versus spread	G002 · Potential–Realization Separation	Alias
heterogeneous thresholds	G004 · Heterogeneous Realization	Alias
holistic multidimensional evaluation	G007 · Multidimensionality and Partial Order	Alias
iterative refinement	G005 · Iterative Revisability	Alias
measurement-dependent convergence	G008 · Observational Partiality	Alias
mismatched clocks	G003 · Relative Rates and Lag	Alias
mutual learning	G005 · Iterative Revisability	Alias
no single arrival moment	G001 · Extended Realization	Alias
observable versus latent	G008 · Observational Partiality	Alias
potential and realization	G002 · Potential–Realization Separation	Alias
preserved adaptability	G005 · Iterative Revisability	Alias
reciprocal learning	G005 · Iterative Revisability	Alias
recognized variety	G008 · Observational Partiality	Alias
scalar-to-tensor	G007 · Multidimensionality and Partial Order	Alias
Slope	G003 · Relative Rates and Lag	Related term — model variable/formalization, not an alias
software speed versus infrastructure speed	G003 · Relative Rates and Lag	Alias
synchronization rate	G003 · Relative Rates and Lag	Alias

Series Source Links

The feature trails in the catalog are selective genealogies. The links below provide the eight installment-level sources from which the reader-facing reconstruction is drawn.

#1 · Gradual AGI as Epistemic Extension. January 19, 2026.

#2 · Gradual AGI as Abundant Resources. June 11, 2026.

#3 · Gradual AGI as Synchronization for Transformative Adoption. June 29, 2026.

#4 · Ceiling, Floor, and Slope: A Falsifiable Dynamical Model of Synchronization for Gradual AGI. July 4, 2026.

#5 · Gradual AGI as Optimization: A Conceptual Framework. July 22, 2026.

#6 · Gradual AGI as Optimization: Formal Models and Empirical Tests. July 24, 2026.

#7 · Gradual AGI as Contestation: A Framework for Governance. August 3, 2026.

#8 · Gradual AGI as Contestation: Measuring Governance Under Empirical Contact. August 11, 2026.

Context source · Richard R. Weiner, “Fabianism,” International Encyclopedia of Social Theory. Wiley-Blackwell, 2017. DOI 10.1002/9781118430873.est0648.

Technical Companion and Reproducibility

The reader-facing catalog is backed by a separate technical reconstruction. Its current basis is Graduality: A Feature Catalog of Gradual AGI, technical catalog v0.8, together with Appendix A, Feature Record Schema v0.3, and Appendix B, Feature Index and Candidate Register v0.3.

Those materials retain the machinery intentionally omitted here: candidate admission and retirement, stable feature IDs, axis and role coding, provenance subtypes, validation chronology, agreement statistics, route analysis, cross-checks, and the distinction between source-level claims and later reconstruction. The v0.8 provenance backplane explicitly records application-only occurrences—including G003 in the Governance framework—so the installment view can be regenerated as Genealogy plus Applied in without adding those uses to the cards’ visible genealogies.

The present paper should therefore be read as the public-facing map of the vocabulary, not as a substitute for its audit trail. Where a card states that a stronger cross-series conclusion remains a single-analyst reconstruction or has not yet been independently rerun, that qualification is preserved by design.