What if technological and social stratification isn’t a statistical regularity — but the residue of strategic choices about who can know what?

Abstract

This paper sits at the intersection of dynamical systems, control theory, differential game theory, and the epistemology of adversarially-maintained regularities. Its central proposal is that the mathematics of directed transport on a hierarchical chain — with the conductance of each transition treated as a strategic variable — provides a unified framework for understanding how technological and social stratification emerge and persist.

Nature provides the form of the cascade equation; nurture, in the strong sense of strategic choice by agents with differential power, provides the throttle profile that determines how capability flows from the frontier to the periphery. We develop the field-layer model, formalize the moat as a localized concentration of the capability gradient, name three consequences of treating the throttle as strategic, situate the framework within a four-layer architecture (kinematic / control / epistemic / mobility), and apply it to three case studies: the nuclear arc, elite higher education, and frontier artificial intelligence. The paper is framed as a worked instance within the author’s larger research program on intrinsic–extrinsic coupling in dynamical systems.

1. Introduction

Stratification is the characteristic shape of any long-lived technology or social order. What begins localized at an origin becomes, over time, a sequence of attenuated forms ordered by distance from that origin. The front retains the original advantage; the rear retains a narrative of the original advantage and, often, the belief that it is converging on it. This is true of nuclear physics, of elite universities, of access to frontier artificial intelligence. The shape recurs too often to be coincidence.

This paper proposes that the recurrence is structural, that the structure has the mathematical form of directed transport through a hierarchical chain, and that the chain in question is not nature-given but strategically shaped. The proposal sits at the intersection of four disciplines. Dynamical systems provides the evolution equations; control theory, as a subfield, introduces agents who can steer trajectories but cannot reason about other agents; differential game theory extends this to reflexive settings in which each agent’s optimum depends on the other’s, which is a genuine second-order effect (the fixed point of mutual best-responses is nonlinear even when the underlying dynamics are linear); and the epistemology of adversarially-maintained regularities — a frame that includes cryptography, coevolutionary biology, and the more recent machine-learning notion of adversarial robustness — supplies the conceptual tools for doing science on a system whose regularities are being actively shaped against the observer.

The central thesis. Nature provides the form of the cascade equation that governs the propagation of capability through a hierarchy of ranks. This is the intrinsic side. The throttle profile that gates the flow between adjacent ranks — together with the source distribution and depreciation — is not nature-given. It is the residue of strategic choices by agents with differential power. This is the extrinsic side, in the strong sense: nurture, not as perturbation to a nature-given baseline, but as the substance determining the chain’s qualitative properties. The observable patterns of stratification are the joint product.

This framing inherits its mathematical backbone from a larger program on intrinsic–extrinsic structure in dynamical systems (Prather 2026). That program establishes a general result: when a system’s observable behavior arises as the joint action of an intrinsic operator and an extrinsic drive, causal attribution is structurally underdetermined — the product carries less information than its factors. The stratification paper is a case study of what happens when the intrinsic operator is no longer independent of the extrinsic drive but is itself shaped by strategic choice, and when the resulting underdetermination is actively exploited by one of the agents. That exploitation — what we will call weaponized underdetermination — is the secondary thread running through the paper, supporting rather than displacing the throttle mechanics that constitute its primary contribution.

The paper is structured as follows. Section 2 develops the throttled-cascade kinematic model and formalizes the moat. Section 3 introduces strategic control of the throttle profile and names three consequences with no analog in standard transport theory. Section 4 situates the contribution within a four-layer architecture (kinematic / control / epistemic / mobility) and identifies which layers this paper develops, which it sketches, and which it defers. Section 5 applies the framework to three case studies. Section 6 discusses the framework’s relation to the larger research program and its limits.

2. The Throttled Cascade Model

2.1 State and equation

Let \(n \in \{0, 1, \ldots, N\}\) index rank in a knowledge/power hierarchy, with \(n = 0\) the frontier and \(n = N\) the periphery. Let \(C_n(t) \in \mathbb{R}_{\geq 0}\) denote the capability level at rank \(n\) at time \(t\). The kinematic equation is

with boundary conditions \(\kappa_{-1,0} = \kappa_{N,N+1} = 0\). Four parameter families do distinct work.

• \(R_n(t)\) — the innovation source at rank \(n\). Bimodal in the natural case: large and continuous at the frontier (\(n = 0\)); small and stochastic at the tail (\(n = N\), “practical novelty without implementation capacity”); near-zero in the middle.
• \(\kappa_{n,n+1} \in [0, 1]\) — the throttle between adjacent ranks. The strategic primitive of the entire framework. Symmetric in flow direction: the same throttle that gates downstream propagation also gates upstream backflow.
• \(\mu_n\) — depreciation. Knowledge ages; position erodes without maintenance.
• Topology. The chain itself is not fixed. Ranks can be inserted, removed, or split; throttles can be opened or closed.

2.2 Continuum limit

For continuous analysis the natural limit is

on \(x \in [0, L]\), a parabolic PDE with strategic conductivity \(\kappa(x, t)\). Real-valued, no oscillation, no interference. The discrete chain is preferred for argumentation; the continuum is preferred for visualization and for connection to the broader literature on inverse-coefficient problems.

2.3 Strain

The local gradient \(C_{n-1} - C_n\) (equivalently, \(-\partial_x C\)) measures the capability step between adjacent ranks. Where \(\kappa\) is small, capability piles up upstream and the step steepens — adjacent ranks pull apart in capability terms. Where \(\kappa\) is large, the step flattens. The visible profile across all ranks is a sequence of steps of varying size. A strategically engineered moat appears as a kink in the profile, not as a smooth slope.

2.4 The moat, formalized

A moat is not a coefficient. It is a profile feature: a localized region of low \(\kappa\) that concentrates the capability gradient at one position. We define

— the share of the total cumulative gradient concentrated at throttle \(n^*\). A flat \(\kappa\) profile gives \(M \approx 1/N\) everywhere; a strategically engineered moat gives \(M \to 1\) at one position.

In the steady state of a uniform chain with constant frontier rate \(R\) and uniform \(\kappa\), the profile is linear: \(C_n = C_0 - n R/\kappa\). A strategic in-group at rank \(n^*\) deviates from linearity by depressing \(\kappa_{n^*, n^*+1}\) relative to its neighbors, producing a measurable kink. The location of the kink identifies the in-group; its sharpness measures the strength of the moat.

2.5 The bimodal source: a falsifiable prediction

Innovation sources peak at both ends of the hierarchy: continuously at the frontier (where capability concentrates and feeds new work) and intermittently at the tail (where novel ideas occasionally appear without implementation capacity). The middle is largely silent. This bimodal structure, together with the symmetry of the throttle, produces a prediction that has no analog in unidirectional transport models. A throttle restrictive enough to protect frontier rent against downstream leakage is also restrictive enough to block tail-to-frontier backflow of peripheral innovations. Strong-moat regimes therefore systematically miss peripheral innovation. This is the first sharp falsifiable claim of the framework, and Section 5 will check it against the historical record.

3. Strategic Control of the Throttle

3.1 The move

In standard transport physics, \(\kappa\) is a property of the medium, given by the physics and discovered by experiment. In the framework of this paper, \(\kappa\) for the medium through which technological capability propagates is chosen. The coefficients are not nature-given constants but the residue of strategic activity by a subset of agents — the in-group — who have access to control levers the out-group does not. The in-group’s problem is to choose a trajectory \(\kappa(t)\) that optimizes its objective. The resulting throttle profile is a negotiated object, and any attempt to model it as nature-given misses what it is.

This is a small mathematical modification. It is a larger conceptual one. Once \(\kappa\) is strategic, three consequences follow that have no analog in standard transport mechanics.

3.2 Observer-dependent throttle

The in-group’s choice of \(\kappa(t)\) depends on what the out-group believes. When the out-group develops tools to measure the capability gap, the in-group revises \(\kappa\) — or, more often, revises the publicly visible signal \(C^{vis}\) that the out-group’s beliefs track — to frustrate the measurement. In ordinary transport theory the conductivity of a medium does not care whether it is being observed. In strategic transport, it does. The observer’s epistemic state is a parameter of the dynamics — a feature the framework inherits from its game-theoretic structure rather than from transport mechanics proper.

Formally: let \(B(t)\) denote the out-group’s belief about the gap. Then the in-group’s optimal \(\kappa\) is a function \(\kappa^*(t;\,B(t))\), and the belief dynamics

close the loop. Neither the in-group’s policy nor the out-group’s belief is exogenous; each is a fixed point of the other.

3.3 Adversarial throttle (weaponized underdetermination)

The in-group does not merely conceal the throttle profile. It shapes \(\kappa\) and the visible profile \(C^{vis}\) to produce specific misperceptions in the out-group’s model of the dynamics. If the out-group believes the chain is uniform, they will predict eventual convergence between frontier and rear. An in-group that benefits from sustained stratification benefits from allowing this belief to persist while maintaining a sharp kink at \(n^*\). The throttle profile is chosen, in part, to be adversarial with respect to the out-group’s inference procedure.

This is the point at which the paper’s secondary thread — adversarial epistemology — enters with the most force. The underdetermination theorem of Prather (2026) establishes that passive observation cannot separate intrinsic response from extrinsic drive when behavior is their joint product. In the non-strategic case this is an epistemic obstacle that controlled experiment can overcome. In the strategic case the obstacle is being actively exploited. The in-group does not merely benefit from the underdetermination of the intrinsic-extrinsic partition; it manufactures that underdetermination, and it does so in ways designed to route the out-group’s causal attribution into the wrong factorization. Narratives of meritocracy, of natural selection within competitive markets, of the inherent difficulty of frontier work — each is a candidate wrong factorization that the in-group has reason to sustain.

The throttle framing makes the inversion concrete. Given an observed visible profile \(C^{vis}\) and assumed knowledge of \(R\) and \(\mu\), the steady-state estimator is

If \(C^{vis}\) is being shaped to be a wrong factorization — for instance, made uniform when the underlying chain has a sharp moat — then the estimator returns a profile that is uniform when the actual profile is concentrated. The visible-profile narrative the in-group propagates does not need to deceive the out-group about specific measurements; it needs only to control which measurements the out-group treats as informative about \(\kappa\).

3.4 Self-stratifying belief

The third consequence is that the observer’s belief is not exogenous to the system — it is part of the dynamics. The out-group’s belief \(B(t)\) about the gap influences how it allocates resources, which influences its actual realized capability, which closes back on the gap. If the out-group believes the chain is uniform, it under-invests in moat-breaking strategies; this under-investment makes the chain effectively more stratified than it would have been under a correct belief. Belief is not merely an epistemic state but a coupled dynamical variable. The framework predicts that successful in-groups will work to maintain out-group beliefs that are not just incorrect but self-defeating — beliefs whose action consequences entrench the very stratification the belief denies.

4. The Architecture in Layers

The full system has four coupled layers, each with its own state variables and dynamics:

1. Kinematic. The chain itself: \(C_n(t)\), the cascade equation, the throttle profile \(\kappa\). This is what Section 2 develops.
2. Control. The optimization problem the in-group solves to choose \(\kappa(t)\). Standard control theory if the out-group is non-strategic; differential game if both groups optimize. Section 3 sketches this layer.
3. Epistemic. The belief dynamics \(B(t)\) and the visible profile \(C^{vis}(t)\) the in-group projects. Includes narrative management, signal control, and the inversion problem the out-group faces. Sketched in §3.2–3.4.
4. Mobility. Agents migrating between ranks under capability shocks or strategic openings. The agent layer in the companion simulator; not formalized in this draft.

This paper develops layer 1 in full, sketches layers 2 and 3, and defers layer 4. The framework’s claim to unify nuclear, educational, and AI stratification rests on layer 1 alone; the strategic-throttle move of layer 2 is what distinguishes this account from non-strategic adoption-curve models; layer 3 is what makes the framework adversarially robust against the in-group’s narrative work; layer 4 is what makes it predictive about individual trajectories rather than only field-level shape.

5. Case Studies

5.1 The nuclear arc

From the Manhattan Project to non-proliferation regimes to the current AI-and-nuclear nexus: nuclear physics has the cleanest historical record of a chain where the throttle profile was negotiated explicitly. The Atomic Energy Act of 1946 codified \(\kappa\) at one transition (frontier → states) as approximately zero, then later admissions (1953 Atoms for Peace, 1968 NPT, modern reprocessing rules) re-opened it selectively. The bimodal-source prediction holds: peripheral innovation in nuclear technology has been systematically routed back through the frontier (DOE laboratories, the IAEA) rather than being permitted to develop independently in the periphery. Strong moats produced the predicted side effect of missed peripheral innovation, most visible in the underdevelopment of thorium and small-modular-reactor technologies in the 1970s–90s — not because the ideas were absent in the periphery but because the throttle blocked their backflow to the frontier where implementation capacity lived.

5.2 Elite higher education

The chain runs from research universities (rank 0) through regional institutions and selective liberal arts colleges (intermediate ranks) to community colleges and online providers (the tail). The throttle profile is shaped by accreditation, transfer credit rules, and the structure of the labor market signals that employers read. The visible profile \(C^{vis}\) — what employers infer about graduates — is actively shaped by both the in-group (signaling research-university quality without revealing the components of that quality) and intermediate institutions (working to compress the visible gap). The Hoxby–Avery result on missing one-offs is the bimodal-source prediction in this domain: high-achieving low-income students at the periphery represent a peripheral innovation source that the throttle blocks rather than promoting to the frontier.

5.3 Frontier AI

The third case is the most recent and the one where the throttle is being negotiated in real time. The chain runs from frontier labs (rank 0, currently three to five organizations) through enterprise customers, academic researchers, open-source communities, and ultimately end users. The strategic throttle has multiple components: capability disclosure (papers vs. system cards vs. red-team-only releases), model weights (closed, gated, open), API access (rate limits, deployment restrictions), and the narrative framing of risk (safety-coded restrictions on what gets released). The GPT-3 staged release (2020), the LLaMA leak and its consequences, the EU AI Act, and the current frontier-model regulation debate are all visible negotiations of \(\kappa\). The framework predicts — sharply — that the bimodal source structure will produce missed peripheral innovation if the throttle is set too high, and that the visible profile will be shaped to make this miss look like an absence of capability rather than a suppression of flow.

6. Discussion

6.1 Relation to the larger research program

This paper is a worked instance within the program developed in Prather (2026), On Dynamical Systems. That program establishes that when observable behavior is the joint product of an intrinsic response and an extrinsic drive, causal attribution is structurally underdetermined. The stratification paper extends this in two directions. First, the intrinsic operator is itself a strategic variable — it is not just that we cannot separate response and drive from observation, but that the response side of the partition is being actively shaped. Second, the underdetermination is being exploited — not merely tolerated as an epistemic limit. The two extensions together give the framework its distinctive character: it is a theory of how a known epistemological limit becomes a strategic asset for one of the parties to a dynamical system.

6.2 What we are not claiming

We are not claiming that every stratified field is the result of conscious strategic choice. Much of what looks strategic from the outside is the unintentional consequence of incentive structures, regulatory accidents, and historical contingency. The framework treats strategic in the game-theoretic sense — agents responding to expected payoffs — not in the conspiratorial sense. The Atomic Energy Act of 1946 was the product of a specific political coalition with specific information about specific risks; the resulting throttle was strategic, but it was not the work of a unitary actor pursuing a fixed objective over decades. The framework asks how the resulting profile looks, not whether it was designed.

6.3 What the framework gets right, and what it does not

The framework captures three features that simpler adoption-curve accounts miss: the widening of the frontier-to-public gap over time rather than its closing; the coexistence of apparent democratization with persistent and sometimes increasing stratification; and the symmetric-throttle prediction about peripheral innovation, which has empirical content. It also provides natural homes for self-stratifying belief, regime-switching control via topology change, and adversarial narrative management.

The framework leaves much unmodeled. Capability is treated as a real-valued scalar at each rank, which is an abstraction. The single-in-group analysis of §3 is restrictive; real fields typically have multiple competing in-groups whose interactions shape the throttle profile in ways this paper does not represent. The narrative channel is handled functionally rather than structurally; a full treatment would model \(C^{vis}\) as its own dispersive field coupled to the observable. The mobility layer — agents migrating between ranks under capability shocks — is sketched but not developed. None of these is a fatal omission, but each marks a direction the framework can be extended in.

6.4 Closing

The framework is a claim about shape, not about necessity. Stratification on the pattern described here is what happens when strategic incentives are followed. The historical counterfactual — Tesla wanting a uniform \(\kappa\) profile while Morgan wanted a sharp moat — is a standing reminder that the throttle profile is chosen. Open-source movements, commons-based governance, and the long arc of scientific publishing each represent arrangements with systematically more uniform \(\kappa\) than the closed default. What the framework tells us is where to look if we want to understand why most arrangements are not of this kind, and what would need to change for them to be. Nature sets the form of the cascade; nurture, in the strong sense, sets the throttles — and it is the throttles, far more than the form, that determine how stratified the medium will be.

References

• Akerlof, G. A. (1970). The Market for “Lemons”: Quality Uncertainty and the Market Mechanism. Quarterly Journal of Economics 84(3): 488–500.
• Bourdieu, P. (1977). Outline of a Theory of Practice. Cambridge University Press.
• Chetty, R., Hendren, N., Kline, P., and Saez, E. (2014). Where is the Land of Opportunity? The Geography of Intergenerational Mobility in the United States. Quarterly Journal of Economics 129(4): 1553–1623.
• DiPrete, T. A., and Eirich, G. M. (2006). Cumulative Advantage as a Mechanism for Inequality. Annual Review of Sociology 32: 271–297.
• Hoxby, C. M., and Avery, C. (2013). The Missing “One-Offs”: The Hidden Supply of High-Achieving, Low-Income Students. Brookings Papers on Economic Activity.
• Merton, R. K. (1968). The Matthew Effect in Science. Science 159(3810): 56–63.
• Prather, C. (2026). Describing the Structure of Dynamical Systems: A Unifying Mathematical Framework. peirastes.com/projects/on-dynamical-systems.html.
• Rhodes, R. (1986). The Making of the Atomic Bomb. Simon and Schuster.
• Schelling, T. C. (1960). The Strategy of Conflict. Harvard University Press.
• Scott, J. C. (1998). Seeing Like a State. Yale University Press.
• Spence, M. (1973). Job Market Signaling. Quarterly Journal of Economics 87(3): 355–374.