I. Introduction: The Morning After the Rebellion
The first two papers traced a path. The first diagnosed the functional transformation of the means of cognitive production: compute, data, and model weights each crossed three vectors from functional resources into control variables, producing a three-class structure in which hyper-concentration constitutes a pro tanto injustice and cognitive exclusion escalates into a violation of basic rights. The second examined the action logic of the dispossessed, demonstrated the dead end of individual technological exit, the limits of liberal action theory under deep capture, and formalized the cognitive coordination trap—multiple binding constraints of spatial dispersion, infrastructure capture, no-trigger co-optation, and identity opacity. Resistance splits into two tiers: strategies within the structure (data contribution, community fine-tuning, serf competition) hit insurmountable ceilings; alternative production entities point toward changing the property rights structure itself.
But that path only leads to its first real starting point: collective action breaking the old centralized structure. What comes after the breaking?
No theory of structural change can evade this question, and it is the most difficult question in history. Breaking an unjust structure is relatively easy—when historical tensions accumulate to a threshold, the structure cracks on its own. What is difficult is what comes after: how to ensure the new structure does not repeat the same mistakes.
The question is not hypothetical. It carries a heavy historical record.
II. The Paradox of Revolution: New Masters, Old Structures
2.1 The Soviet Lesson, Revisited
The largest-scale revolutionary experiment of the twentieth century ended with the Soviet Union. The capitalists and landlords of Tsarist Russia were overthrown; property was "nationalized." But Party bureaucrats replaced the old capitalists, seizing control over the power to allocate resources. They did not need nominal ownership, because they controlled the power to distribute resources itself.
The Soviet Union did not eliminate hyper-concentration. It converted hyper-concentration from private capital into bureaucratic power. The function of wealth did not change—it remained a variable for controlling the feasible sets of others. Only the names of the controllers changed.
2.2 The Replication Risk for Cognitive Production
"Revolution replaces people but not structures" takes distinctive forms in the context of the means of cognitive production.
Suppose a social movement successfully breaks the current compute lords' centralized control. The old AI giants are broken up or nationalized, cognitive serfs gain better conditions, the excluded gain access. Then what?
If the new institutions do not structurally prevent re-concentration, new concentration re-emerges within a generation, in three possible forms.
Alienation of nationalization. If the means of cognitive production are nationalized but the state itself is captured by a new bureaucratic class, concentration has merely shifted from private capital to state bureaucratic power. The new managers of state-owned AI enterprises may not need nominal ownership, because they control allocation of compute, data access, and model deployment. Structurally isomorphic to the Soviet Union's shift from "capital" to "Party membership"—only the entry point shifts from "the market" to "the administrative apparatus."
Emergence of new oligarchs. If no dispersive mechanisms are embedded after the break, market logic re-manufactures concentration. Even if old AI giants are broken up, economies of scale (the physical economies of compute, the self-reinforcement of data flywheels) drive new players toward new concentration. Within a generation, new compute lords re-emerge under different names.
Re-concentration through international asymmetry. Even if one country successfully disperses domestically, international power dynamics can re-manufacture concentration. If one country's AI industry is dispersed and weakened while another's compute lords keep concentrating and gain capability advantages, the former may move back toward concentration under competitive pressure. This is a dimension in which the means of cognitive production differs from traditional wealth: it carries strong geopolitical implications.
The diagnosis of the first paper was that the problem lies in the mechanism of functional transformation itself, not in the specific entities that hold the means. Replacing entities does not solve the problem. You can break up every current AI giant—but if the new institutions allow re-concentration deep into the functional transformation zone, new compute lords emerge within a generation. The disease of the structure does not heal because the patients have been replaced.
2.3 Revolution Is Not the End Point
Revolution is not the end point. Breaking the old structure is only the first step, and not the most difficult one. The most difficult step: designing a new structure that makes re-concentration difficult, makes institutional capture costly, and preserves the dispossessed's capacity for continuous intervention.
Solidarity, the second paper argued, is a necessary condition for changing the structure. It is also a necessary condition for maintaining the new structure. Institutional design is not a one-time project but a structure requiring continuous maintenance—and the force for that maintenance remains solidarity.
III. Preventing Re-Phase Transition: The Core Design Problem
3.1 The Objective
The objective is not to eliminate AI enterprises. This is neither possible (any complex society needs mechanisms for organizing and accumulating the means of cognitive production) nor desirable (AI as a cognitive tool is not inherently unjust; the injustice diagnosed in the first paper occurs only after functional transformation deepens).
The objective: prevent the means of cognitive production from sliding from "functional resources" into "control variables."
More specifically, the new institutions must operate on three progressive layers—preventing holdings of the means of cognitive production from penetrating deep into the functional transformation zone; equipping institutions with sufficient anti-capture mechanisms so that crossing the threshold is difficult; and preserving the capacity for self-correction when capture partially occurs.
Prevention, resistance, correction. These three layers mirror the institutional design of the third paper of the Wealth Trilogy, only now the objects of constraint are the three vectors of the means of cognitive production rather than natural-person and legal-person wealth. The first paper showed the phase transition occurs independently on each of the three vectors—so constraints must symmetrically cover all three. Under existing conditions, any constraint on a single vector cannot interrupt the phase transition mechanism itself.
3.2 Why Taxes and Antitrust Are Not Enough
Taxation is a downstream instrument. Progressive taxation is an ex post adjustment of excess profits; it intervenes only after algorithmic rents have been monopolized. But the first paper showed the self-reinforcing cycle of institutional capture: once the means of cognitive production are concentrated enough to influence rule-making, holders rewrite the tax code itself. Taxation depends on the will of the legislature, and the legislature can be captured. Moreover, AI enterprises have abundant means of profit shifting—cross-border pricing, intellectual property licensing fees, R&D deductions—opening a substantial gap between nominal tax rates and actual burdens.
Traditional antitrust is similarly insufficient because the phase transition spans three vectors, and single-dimensional antitrust (market share in a particular product market) cannot cover it. Breaking up an AI giant's model business still leaves its compute infrastructure, data flywheel, and ecosystem lock-in intact, re-manufacturing concentration in new forms. Worse, antitrust's "consumer welfare standard" (using price as the primary criterion) cannot capture phase transitions under free inference: free services do not raise consumer prices, but they manufacture deeper dependency through co-optation and data extraction. The evaluative framework of antitrust itself must be updated for the AI era.
What is needed are upstream instruments—ones that change the generation structure of the means of cognitive production (who obtains what share of compute, data, and weights under what conditions) rather than merely redistributing ex post. The Three-Principle Framework below is precisely that.
3.3 The Three-Principle Framework
Three principles operate in parallel, addressing the three vectors. These are principles, not a policy menu. Principles are normative requirements for institutional arrangements; each can have multiple implementations. The choice of specific policy depends on empirical context and cannot be settled at the normative level. The principles themselves are derived from the diagnosis of the first paper, their number fixed by the complementary bottleneck structure of the three vectors (Section 2.2 of the first paper demonstrated their necessity and joint sufficiency); the implementation modes under each principle are open, comparable, and context-adjustable.
Compute Principle: No single entity should control enough compute to unilaterally set the frontier of cognitive capability. When compute crosses the threshold it becomes a control variable, enabling its holder to determine who can enter the race. The normative requirement: control over compute must be sufficiently dispersed that no single entity unilaterally sets the frontier. Implementations include (a) public compute as a public utility; (b) compute antitrust and break-up; (c) a GPU tax or compute excess-profits tax directed toward public purposes; (d) transparency and concentration monitoring of compute usage. These are comparable substitutes, not mutually exclusive options. So long as an arrangement satisfies "no single entity unilaterally sets the frontier," it is a legitimate implementation of the principle.
Data Principle: The data flywheel should not become a tool by which a single entity locks out competitors and users. When data accumulation crosses the critical scale, the self-reinforcing flywheel makes it impossible for competitors lacking an equivalent user base to catch up. The normative requirement: control over data should correspond to generative contribution, and data should not be monopolized as a lock-out tool. Implementations include data trusts (trustees managing rights on behalf of contributors), public data pools, data portability and interoperability requirements, and Data Dividend or data profit-sharing. Again, comparable substitutes.
Weight Principle: Models trained using public data should give back to the public in some form. When closed weights become a tool of ecosystem control, the holder exploits public data (web-crawled corpora, publicly funded research, open-source contributions) without giving back. The normative requirement: model outputs that utilize public inputs should not be fully privatized. Implementations include mandatory weight disclosure, tiered disclosure (fine-tuning rights, periodic release of older weights, independent auditing), an AI License framework (analogous to drug patent pools), and open-source incentives (tax or regulatory benefits for disclosing weights).
The three principles operate in parallel and mutually support one another. Reform of a single principle will be hedged by violations of the others. Only when all three are simultaneously satisfied can the self-reinforcing closed loop of concentration be structurally dismantled. The weighting among principles and the choice of implementation modes are open. The framework does not prescribe that "the Compute Principle must be realized through public compute" or "the Weight Principle must be realized through mandatory disclosure"—it prescribes only the principles themselves. Normative determinacy, alongside operational flexibility.
Do the principles conflict? They do not, at the normative level, because they operate on different vectors—three independent nodes of the complementary bottleneck. The Compute Principle governs the dispersion of compute control; the Weight Principle governs the public return of model outputs; their domains do not overlap. What looks like a "conflict" (an enterprise trading weight disclosure for a compute concentration exemption) is a bargaining scenario in policy implementation, not a normative conflict at the level of principle. The principles say both requirements should be satisfied; the proposed bargain satisfies only one. Trade-offs in specific implementation ("how high should the GPU tax rate be," "how long should the weight disclosure lag be") are empirical, context-dependent, and require no a priori adjudication.
IV. Implementing the Compute Principle
4.1 Design Rationale
The root of the compute phase transition is economies of scale: large clusters have lower unit costs and tend naturally toward concentration. Preventing the phase transition cannot deny that physical fact—but can, through institutional arrangements, prevent compute control from being monopolized by a small number of entities.
Public compute as a public utility. By analogy with roads, electricity, and water, frontier compute can be classified as infrastructure with public utility characteristics. The state or public entities build and operate large clusters, making them available to researchers, startups, and public institutions at cost or regulated prices. This provides an alternative outside the compute lords.
Compute antitrust and break-up. Enforce antitrust against private providers exceeding a specified market share, limiting any single entity from controlling more than a certain proportion of national (or global) compute capacity. Not eliminating private compute—preventing it from concentrating deep into the functional transformation zone.
Transparency requirements for compute usage. Require large providers to disclose usage (who rented how much, when, for what purpose), enabling continuous monitoring of concentration. This is the data foundation for phase transition monitoring (Section VII).
4.2 Historical Reference: Public Utility Regulation
The publicization of compute is not without precedent. Infrastructure with strong economies of scale—electricity, railways, telecommunications—has, in different countries and times, undergone publicization or public utility regulation.
The U.S. electricity industry moved from private monopoly to regulated public utility in the early twentieth century. States established public service commissions to regulate pricing, service quality, and investment. The model is not perfect (regulatory capture recurs), but it kept electricity—infrastructure with strong economies of scale—from being reduced entirely to private monopoly.
Telecommunications offers another reference. Many countries historically nationalized or heavily regulated networks to ensure universal service, so even residents of remote areas obtained basic service. The counterpart in the AI context is universal cognitive service: ensuring marginal language communities and low-income groups obtain basic AI services. Public compute is its material foundation.
4.3 Limitations and Challenges
Building and operating frontier clusters requires enormous investment: billions in hardware, sustained energy supply, professional teams. Public compute is not cheap. Fiscal capacity sets the upper bound; fiscally constrained states (especially in the Global South) may be unable to afford it, so compute publicization is more feasible in developed countries, while developing countries may require international cooperation (regional public compute pools).
Economies of scale do not disappear. If a private provider's scale far exceeds public compute (because it can continuously invest in expansion), public compute becomes a marginal substitute rather than a genuine check. To truly check compute lords, public compute must reach a scale at least comparable to the largest private player—extremely challenging both fiscally and politically. And compute technology (chip architectures, interconnect, cooling) iterates rapidly: once built, public compute may quickly become obsolete relative to the latest private compute, requiring sustained investment in upgrades.
Geopolitics compounds the problem. The supply chain for frontier compute (especially high-end GPUs) is highly concentrated. A country's capacity to build public compute is constrained by access to chips, which is constrained by geopolitics (U.S. chip export controls). Compute publicization is not only domestic policy but geopolitical. Public compute in the Global South may be limited by the political control of chip supply.
V. Implementing the Data Principle
5.1 Design Rationale
The root of the data phase transition is the self-reinforcing flywheel: more users → more data → better models → more users. To prevent it, the mechanism by which "data is monopolized for a single flywheel" must be severed.
Data trusts. Legal entities in which trustees manage data rights on behalf of contributors. Users contribute data to the trust; the trust, under fiduciary duty, negotiates terms with AI enterprises, obtains compensation, and supervises usage. The trust does not "return" data to individual users (individual data has very low value in isolation); it concentrates dispersed rights into a collective force capable of bargaining. This is structurally isomorphic in spirit to the "universalization of beneficiary rights" in the third paper of the Wealth Trilogy: preserve the existing property rights framework while embedding collective distribution at the level of benefit flows.
Public data pools. Publicly funded open datasets (multiple languages, multiple domains) as alternatives to commercial datasets. Any researcher, startup, or public institution can use them, weakening commercial flywheels. Quality depends on sustained investment and community contribution; they connect directly to the data contribution strategy in the first tier of the second paper.
Data portability and interoperability requirements. Require AI enterprises to permit users to export data (usage history, preference feedback, fine-tuning data) in standard formats and migrate across platforms. This reduces lock-in and enables users to "vote with their feet"—not formally (Section II of the second paper showed formal foot-voting is ineffective under lock-in), but substantively (when data is portable, migration costs drop sharply).
5.2 Historical Reference: Trust Prototypes and Pool Precedents
Data trusts remain theoretical prototypes and preliminary experiments, but there are referable precedents.
The UK's Open Data Institute (founded by Tim Berners-Lee and Nigel Shadbolt) has promoted the concept and conducted experiments in healthcare and urban governance. These show feasibility in specific domains (medical data), but for cross-domain, large-scale general-purpose AI data, the governance structure (who serves as trustee, how to balance contributors' interests, how to negotiate with enterprises) remains an open legal design question.
Precedents for public data pools include Common Crawl (open web-crawled corpora), Wikipedia (a human-edited knowledge base), and Common Voice (open speech). Their existence proves collective contribution can produce valuable public alternatives. But their scale and quality currently fall far short of commercial enterprises' private data—Common Crawl's corpus, though enormous, cannot match carefully cleaned and labeled commercial data. For public pools to become a genuine check, sustained, institutionalized investment is required, not merely voluntary contribution.
5.3 Limitations and Challenges
Section VII of the first paper already identified the fundamental difficulty of data valuation for operationalizing ρARA: how is the value of different data contributions assessed? A trust allocating compensation must answer this, and any valuation embeds normative judgments. The governance of trusts must therefore include democratized valuation procedures, not values unilaterally set by enterprises or technical experts.
Trust effectiveness depends on enough users delegating rights to it—this is itself a collective action problem (Section IV of the second paper). Each user has an incentive to free-ride: let others join, let the trust bargain, enjoy the improved terms without joining. If enough users opt out, the trust lacks bargaining power. Structurally isomorphic with all collective action dilemmas, it requires an organizational foundation (user consciousness-building, the visibility and credibility of the trust) to overcome.
Privacy and sharing sit in tension. Trusts must find a balance between "making data usable" and "protecting individual privacy." Differential privacy and federated learning can partially mitigate (allowing models to learn without directly accessing raw data) but add complexity and performance costs.
And the data layer lacks transnational governance. Data flows across borders; AI enterprises operate transnationally. A trust established by one country may face counterparts outside its jurisdiction. Effective operation requires a transnational framework, and such a framework is currently almost nonexistent.
VI. Implementing the Weight Principle
6.1 Design Rationale
The weight phase transition is the most distinctive of the three, because its correction cost is the lowest (the zero marginal reproduction cost of weights makes "disclosing weights" technically almost costless). That distinctiveness makes the Weight Principle the most politically difficult (it directly threatens the core rent source of compute lords) and the most technically feasible.
Conditional disclosure. The core design: models trained using public data disclose their weights in some form. The qualifier "using public data" is the key limit. If a model is trained entirely on private data (enterprise-owned, legally acquired, non-public), disclosure is not required. But if training uses public data (web-crawled corpora, publicly funded research, open-source contributions), the weights should be made available to the public in some form.
The jurisprudential basis is Murphy and Nagel: property rights are institutional artifacts, not natural rights. An AI enterprise's "ownership" of weights is not a natural right but an institutional right conferred by the legal framework. When the exercise of this right (closing weights to extract excess rents) conflicts with social interests (models trained on public data should give back), the legal framework has the authority to redraw boundaries. Closing weights on models trained using public data is not "protecting private property" but "privatizing the output of public inputs"—jurisprudentially challengeable.
Disclosure can be tiered. Full disclosure is one form but not the only one: opening fine-tuning interfaces (third parties can fine-tune on a base model without releasing full weights); periodically releasing old weights that are N generations behind; weight auditing (not disclosing weights but allowing independent third-party auditing of performance, bias, and safety). And disclosure requirements can vary by use case: models for critical public services (education, healthcare, law) face the highest requirements; models for general commercial purposes face lower; models for internal R&D are not required to disclose. Disclosure proportional to social impact.
6.2 Historical References: Open Source and Compulsory Drug Licensing
Model weight disclosure has two references, each providing both inspiration and caution.
Open-source software. Software (especially source code) likewise has near-zero reproduction cost. The open-source movement (GNU/Linux, Apache, Firefox) showed that even without mandatory requirements, open source can form an ecosystem competing with closed-source commercial software. The existence of open source broke the complete concentration of the software domain and provided alternatives—without eliminating closed-source commercial software (Microsoft and Oracle remain enormous). The goal of weight disclosure is the same: not to eliminate closed-source models but to provide a countervailing force.
Open source also offers a caution: maintenance depends heavily on voluntary labor and limited funding, and sustainability is a persistent problem ("maintainer burnout" is widespread). If, after weights are disclosed, open-source models lack institutionalized support, they will gradually fall behind closed-source models with ample resources for continuous training, losing their countervailing force.
Compulsory drug licensing. After a drug's molecular formula is disclosed (patent expiration or compulsory licensing), generics can be produced at far lower prices. This is a direct precedent for "disclosure reducing rents." The WTO's TRIPS Agreement permits member states to implement compulsory licensing in public health emergencies—proof that even within a capitalist framework, when private intellectual property rights seriously conflict with social interests, the law can mandatorily alter exclusivity.
Compulsory licensing also provides a cautionary lesson with a precise AI counterpart. Originator pharmaceutical companies use "evergreening" strategies: applying for new patents on minor modifications to extend protection and evade compulsory licensing. The Novartis Glivec case is the clearest precedent. Novartis applied for a patent on a new crystalline form (the beta crystalline form of imatinib mesylate) of the known compound imatinib, to extend its monopoly on this anti-cancer drug. The Indian Patent Office rejected the application in 2006, and the Supreme Court of India upheld the rejection in 2013, relying on Section 3(d) of the Indian Patent Act—a provision specifically designed to prevent evergreening, requiring that new forms of known substances demonstrate "significantly enhanced therapeutic efficacy" to be patentable.
The AI domain has a structurally isomorphic evergreening risk. Enterprises can evade disclosure in two ways: applying for new "version rights" or technical patents on every fine-tuned, aligned, or RLHF iteration, so that old weights are disclosed but the latest deployed version remains perpetually closed; or through continuous patenting of model architectures, training methods, and data-processing pipelines, maintaining rents through intellectual property claims on other links even when weights are disclosed. Same logic as the first paper's demonstration that the three functional transformations mutually reinforce, and breaking a single link is hedged by the others.
The implementation of the Weight Principle must therefore be accompanied by an anti-evergreening mechanism, drawing on the spirit of Section 3(d): new versions of known models (fine-tuned, aligned, iterated) do not constitute new grounds for closure unless they demonstrate a "significant capability leap" rather than incremental improvement. Disclosure obligations should cover the version actually deployed, not only already-obsoleted old versions. Periodic disclosure (release of old weights N generations behind) is meaningful only if the disclosed version still retains ecological relevance—if what is disclosed is always already eliminated by the market, "disclosure" is nominal.
The substantive standard of "significant capability leap" is admittedly more difficult to objectify in AI than "significant therapeutic efficacy" in pharma (the high manipulability of AI benchmarks is a known problem). But the core force of the anti-evergreening mechanism comes not from a perfect substantive standard but from the reversal of the burden of proof: the enterprise must affirmatively demonstrate that its new version has a significant improvement, otherwise it does not automatically obtain new grounds for closure. This is consistent with the effectiveness logic of Section 3(d)—its force lies in reversing the burden of proof, not in providing an uncontroversial efficacy standard. The specific threshold for "significant capability leap" is an open problem, but the procedural mechanism of reversing the burden is operable and sufficient to significantly raise the cost of evergreening evasion.
6.3 Limitations and Challenges
Weight disclosure directly threatens the core rent source of compute lords, and will encounter the strongest political resistance. They will mobilize every channel of institutional capture diagnosed in Section V of the first paper: lobbying (opposing disclosure in the name of "protecting intellectual property"), opinion-shaping (manufacturing panic in the name of "open source is dangerous"), talent capture (buying off researchers who might promote disclosure). The zero reproduction cost of weights makes disclosure technically almost costless—and precisely for this reason, opposition will concentrate entirely at the political level. The Weight Principle is the hardest to implement and the one whose structural impact, once achieved, is greatest.
The narrative most likely to be weaponized against disclosure is "safety": "disclosing frontier model weights is dangerous—malicious actors could generate disinformation, launch cyberattacks, or manufacture bioweapons." The narrative has some truth (frontier models do carry misuse risks), but it is easily weaponized: compute lords can oppose all disclosure in the name of safety, even when the benefits (breaking the phase transition, raising ρARA) far outweigh the risks.
The key is to distinguish genuine safety concerns from rent protection disguised as safety. One test: if an enterprise simultaneously claims that "weights cannot be disclosed (because of safety)" and "we can commercially deploy this model (because it is safe)," its safety narrative contains a deep internal tension. If the model is safe enough to commercially deploy, it is safe enough to disclose weights; if it is dangerous enough that weights cannot be disclosed, it is dangerous enough that it cannot be commercially deployed.
This is not theoretical speculation. Meta Chief AI Scientist Yann LeCun has publicly accused the leaders of OpenAI, Google DeepMind, and Anthropic of stoking AI fear to advance regulatory capture, manufacturing the narrative that "AI is too dangerous to be developed by anyone but a small number of trusted large enterprises"—justifying restrictions on open source and small competitors. Sam Altman's 2023 testimony before the U.S. Senate calling for a licensing regime for frontier models was criticized by open-source advocates as creating market entry barriers for startups in the name of safety. "AI safety and regulatory capture" has become a searchable research topic. Meanwhile, OpenAI and Anthropic's lobbying expenditures increased substantially in 2025. The empirical pattern—leading enterprises loudly emphasizing safety risks, calling for higher regulatory thresholds, and increasing lobbying investment—is highly consistent with the classic pattern of regulatory capture.
Finally, there is the unilateral-disarmament dilemma. If one country requires disclosure while others do not, the former's AI enterprises are at a disadvantage (forced to disclose core assets competitors retain). Resolving this requires international coordination—analogous to "common but differentiated responsibilities" in climate negotiations, AI governance may need some international framework to coordinate disclosure and avoid a race to the bottom. Achieving that is itself difficult and slow.
VII. Operationalizing ρARA: Control Variable and Evaluation Metric
ρARA, introduced in the first paper as a normative anchor, is now operationalized into concrete policy instruments. It serves two new functions: a control variable for policy intervention, and an evaluation metric for institutional performance.
7.1 As a Control Variable
ρARA can evaluate and rank different policy instruments: to what extent does a given instrument raise ρARA?
- Data rights legislation raises the numerator by requiring compensation: if legislation requires data contributors to receive x% of AI enterprise revenue, ρARA rises by approximately x percentage points.
- A public AI fund (funded by AI enterprise taxation) directs rents to public purposes: a tax of y% on AI enterprise profits, injected into a fund for low-resource language support, public compute, and open-source maintenance, raises ρARA by approximately y percentage points.
- Model weight disclosure requirements reduce the monopolized portion of the denominator, subjecting excess rents of closed weights to competition or publicization.
- Compute antitrust directly reduces individual lords' rent extraction capacity; low-resource language data ratio requirements ensure the excluded are at least included in distribution at the data level.
These instruments need not be implemented simultaneously. The value of ρARA lies in providing a unified metric for evaluating any single one, enabling comparison and ranking rather than isolation. This operationalization is directional—raising the numerator raises the ratio is logically straightforward. A more powerful operationalization would answer comparative statics: how does ρARA evolve over time under unchanged policy? How do marginal impacts differ (data trust vs. public compute vs. weight disclosure)? Those questions require quantitative empirical modeling of cost structures, data valuation, and rent flows—exceeding the scope of a normative framework, and an important direction for subsequent research.
7.2 As an Evaluation Metric
ρARA can also serve as a continuous metric for institutional performance. If the new institutions successfully prevent re-phase transition, ρARA should trend upward (or at least not decline). If ρARA begins to decline under the new institutions, this is an early warning signal: phase transition may be reoccurring, anti-capture mechanisms may be eroding.
This requires regular measurement and public reporting—an institutional design problem itself: who measures it? How are methods standardized? How is internal data obtained (especially what AI enterprises are unwilling to disclose)? One design: an independent "Algorithmic Rent Audit Agency" (analogous to a financial audit agency) responsible for regularly evaluating and publicly reporting ρARA for major AI economies. The independence of such an agency—its freedom from capture by compute lords—is the precondition for its effectiveness, which returns us to Section VIII.
7.3 ρARA Is Not a Target
ρARA is not a target value that must reach 1. Complete distribution of rents is neither possible nor desirable (incentives for innovation and investment must be preserved). It is a directional guide: current ρARA is likely too low, and raising it is a requirement of justice; but the magnitude, speed, and instrumental mix depend on institutional endowments and political possibilities.
A rough international reference: at the height of the welfare state (Northern Europe in the mid-to-late twentieth century), the "rent redistribution ratio" of the traditional economy (the proportion of total economic surplus redistributed through wages, taxation, and welfare) may have been 0.4–0.6. If the AI economy's ρARA could rise from its current extremely low level to 0.2–0.3, that would be an enormous structural improvement: at least one-fifth to one-third of algorithmic rents returned to co-producers, rather than almost entirely monopolized. But this is a heuristic reference, not a normative target. The normative criterion is Rawlsian: whether the distribution improves the interests of the least advantaged—cognitive serfs and the excluded.
VIII. The Paradox of Anti-Capture
The Three-Principle Framework and ρARA together constitute a framework for preventing re-phase transition. But there is a core paradox, already identified in the third paper of the Wealth Trilogy: any anti-capture mechanism can itself be captured.
8.1 The Guardian Paradox
Who executes the anti-capture mechanisms? The managers of public compute, the trustees of data trusts, the regulators of weight disclosure, the auditors of ρARA. These are "guardians." Their duty is to prevent re-concentration. But guardians are human; they can be bought, coerced, or captured.
Public compute managers may be lobbied to allocate compute preferentially to research affiliated with compute lords. Data trust trustees may be bought off, making concessions that harm contributors. Weight disclosure regulators may grant "safety" exemptions to models that should be disclosed. ρARA auditors may be pressured to underestimate algorithmic rent share.
Anti-capture design cannot rely on the premise that "there exists a neutral, uncaptured executing agent." That is a fragile premise. In reality, every executing agent faces the risk of capture.
The PUHCA precedent. The U.S. Public Utility Holding Company Act of 1935 is a classic precedent of anti-concentration design. At the time, approximately 86% of operating electric utilities belonged to interstate holding company pyramid systems; PUHCA forcibly broke up these structures, confining them to single integrated systems and placing them under SEC oversight. The institution operated for roughly 70 years, partially curbing concentration. Then the Energy Policy Act of 2005 repealed PUHCA, leading to a resurgence of mergers and acquisitions in electricity and gas, reversing decades of anti-concentration policy. A textbook case of an anti-concentration institution itself being captured—abolished by the very industry it was supposed to constrain, under sustained lobbying and political realignment.
The mechanisms of PUHCA's failure deserve attention. Its repeal was not sudden but the result of decades of gradual erosion: industry accumulated political influence through sustained lobbying; under the neoliberal shift of the 1990s, "deregulation" acquired cross-partisan legitimacy; formal abolition came in the 2005 energy legislation. PUHCA's weakness lay in its reliance on a single executing agent (the SEC) and a single legal framework (federal securities law). Once captured, the entire institution lost its guardian. This directly supports the principle of multiple redundancy argued below: if an anti-concentration institution is composed of multiple mutually independent layers (legal requirements, administrative regulation, judicial review, independent auditing, citizen participation), capture requires buying off all layers simultaneously, and the cost is far higher. The lesson of PUHCA is not "institutions inevitably fail" but "institutions with a single guardian are more likely to fail than institutions with multiple redundancy."
An institution's life cycle is not infinite merely because it was once successful; it can be repealed, hollowed out, or reinterpreted as an instrument for protecting vested interests. What design strives for is not a permanent solution, but time: time for the dispossessed to accumulate organizational capacity, time for society to maintain monitoring consciousness.
The Findata vs. care.data contrast. The same institutional idea can produce opposite outcomes under different governance details. Finland's Findata (established 2019 under the Act on the Secondary Use of Health and Social Data), operating as a one-stop licensing agency under statutory mandate, independent operation, and a clear list of purposes, has been relatively successful and continues under the European Health Data Space framework. The UK's NHS care.data (launched 2013), based on a similar centralized idea, failed due to inadequate public communication, insufficient informed consent, and weak governance, leading to over 1 million patients opting out; the program was abolished in 2016. Data trustification does not automatically work; governance details, clarity of statutory mandate, independence of trustees, construction of public trust determine success or failure. The normative requirement is determinate, but the institutional forms realizing it must be context-sensitive.
8.2 Multiple Redundancy: Confronting Structure with Structure
The only structural means of addressing the guardian paradox is multiple redundancy: not relying on a single agent or a single mechanism, but multiple mutually independent, mutually supervising layers.
- Compute layer: not one manager but multiple independent public compute institutions (national, state/provincial, academic consortium) that compete with and supervise one another. If one is captured, others serve as alternatives.
- Data layer: not one trust but multiple representing different groups (users, researchers, marginal language communities), bargaining on different dimensions. The buy-off of one trust does not collapse the whole.
- Weight layer: disclosure enforced not by a single agency but by multiple complementary mechanisms—legal requirements, administrative regulation, judicial review, independent auditing, sustained open-source community pressure.
- ρARA: auditing not by a single institution but by official agencies, academic researchers, and civil society organizations, mutually cross-validating.
The cost is complexity and expense. But it is the only structural means: not pinning hope on any single guardian remaining uncaptured, but making capture require simultaneous buy-off of enough independent agents, raising its cost to an unaffordable level.
This is consistent with the principle of Federalist No. 51, cited in the third paper of the Wealth Trilogy: "Ambition must be made to counteract ambition." Anti-capture does not rely on eliminating ambition but on using ambition to check ambition—using the mutual checks of multiple independent agents to ensure that the capture of any single agent cannot dissolve the framework.
8.3 Continuous Intervention by the Dispossessed
Multiple redundancy reduces the probability of successful capture but cannot eliminate it. The ultimate anti-capture force comes from the continuous intervention of the dispossessed themselves. If cognitive serfs and the excluded possess institutionalized channels of participation (a voice in public compute allocation, voting rights in data trust governance, citizen initiative rights for weight disclosure), any attempt to capture the anti-capture mechanism faces resistance.
New institutions cannot merely be designed "for the dispossessed" (designed by elites, executed by elites, the dispossessed as passive beneficiaries); they must be institutions in which the dispossessed participate in design and management. Institutionalized participation is the last line of defense.
But here the core difficulty of the second paper returns: the organizational capacity of the dispossessed (especially the cognitively excluded) is weak. If they cannot organize effectively, their channels become nominal—voice in name, domination in substance. This constitutes a cycle: effective anti-capture requires organization; organization requires institutionalized space; institutionalized space requires protection by the anti-capture framework. The cycle has no starting point; it must begin from some imperfect present, accumulating gradually.
IX. Re-examining Historical Cases
The Three-Principle Framework must be honestly tested against historical precedent.
9.1 Open-Source Software: Partial Success
Open source is the closest precedent for the publicization of the means of cognitive production. Under conditions of near-zero reproduction cost (software), public-good alternatives can emerge without compulsion and constitute a countervailing force against commercial concentration.
Its success factors: legal infrastructure (the force of licenses like the GPL), organizational vehicles (the Linux Foundation, Apache Foundation), corporate participation (IBM, Google finding advantage in supporting open source), and community culture (the hacker ethic, voluntary contribution). Together they made open source a force in the software domain.
But open source has limitations: it has not eliminated the concentration of commercial software (Microsoft and Oracle remain enormous); sustainability depends on sustained voluntary and corporate investment (maintainer burnout is real); in certain domains (enterprise databases, professional design software) it still cannot compete. The lesson for AI: open-source models can serve as a countervailing force but cannot dissolve concentration on their own. They need to synergize with the institutional realization of the Three Principles to form genuine structural checks.
9.2 Public Broadcasting and Universal Service
Public broadcasting (the BBC, NHK) is another reference—publicly funded (license fees or taxation), oriented toward public service rather than commercial profit, providing information public goods outside commercial media. Its existence means the information ecosystem is not entirely commercialized; the public has at least one source not dominated by advertisers and shareholders.
Its success: public funds can sustainably support cognitive production not entirely dominated by commercial logic. Its limitations: dependence on the political will of the state (budgets subject to pressure); independence requiring institutionalized protection (the BBC's independence is a complex legal and political design); and resource disadvantage when facing competition from global commercial media. These lessons apply directly to public compute: it needs a similar independence structure (insulated from short-term political pressure) and a sustainable funding mechanism.
The universal service principle in telecommunications is also a precedent: even in unprofitable areas, basic service must be guaranteed. The AI counterpart is universal cognitive service: even in unprofitable language communities and regions, basic AI services must be guaranteed. Public compute is its material foundation.
9.3 The Drug Access Movement: Political Precedent for Rent Redistribution
The drug access movement (especially the HIV/AIDS access movement of the 2000s) is a political precedent for breaking intellectual property monopolies. High prices of antiretroviral drugs made them unaffordable for AIDS patients in developing countries. A global social movement (patient groups, NGOs, developing-country governments, some generic drug companies) promoted compulsory licensing and generic production. The annual cost of antiretroviral therapy dropped from approximately $12,000–$15,000 to approximately $100, enabling millions to receive treatment.
The lesson: monopoly rents on intellectual property can be broken, but this requires cross-border social movements and political coalitions. Market forces alone or states alone are insufficient; what is needed is a coalition of patients (the dispossessed), NGOs (external solidarity actors), developing-country governments (political force), and alternative suppliers (generic companies / open-source models). The AI equivalent would be a transnational coalition of cognitive serfs, digital rights NGOs, Global South states, and the open-source community.
The drug access movement also provides a caution: originator companies use evergreening (new patents on minor modifications) to continuously evade compulsory licensing. The AI domain may see similar evergreening—continuous patenting of architectures, training methods, data processing, maintaining rents through closure of other links even when weights are disclosed. The three principles must mutually support one another; no single principle can be treated as "already resolved."
X. Conclusion: Maintenance Without End
10.1 The Logical Chain of the Three Papers
The three papers form a complete logical chain, corresponding precisely to the Wealth Phase Transition Trilogy.
First Paper (Diagnosis). The phase transition of the means of cognitive production deprives hyper-concentration of moral protection. Compute, data, and weights each transform, on three vectors, from functional resources into control variables; the triple phase transition mutually reinforces, forming a self-reinforcing closed loop, producing a three-class structure; institutional capture tilts the rules; cognitive exclusion escalates injustice into a violation of basic rights. ρARA is the linking variable between diagnosis and norm.
Second Paper (Action). Individual technological exit is a dead end. Liberal action theory loses efficacy under deep capture. The core contribution is the formalization of the cognitive coordination trap—multiple binding constraints of spatial dispersion, infrastructure capture, no-trigger co-optation, and identity opacity. The first two components have been addressed in platform labor studies (Vallas & Schor, Prassl); the CCT's increment is identifying the distinctive constraints of AI annotation labor in two dimensions: employer visibility (obscured by the outsourcing chain) and leverage points (tasks globally re-allocatable). Resistance splits into two tiers: within-structure strategies hit ceilings; alternative production entities point toward changing the property rights structure itself. Ultimate institutional change requires collective political action and solidarity.
Third Paper (Design). The Soviet lesson: revolution replaced the people but left the structure intact. Since the cause lies in the phase transition mechanism of the means of cognitive production, the remedy must be structural. The Three-Principle Framework symmetrically covers the three complementary bottlenecks; the normative requirement of each principle is determinate, the modes under each open and comparable. ρARA is operationalized as a policy control variable and a performance metric. The Weight Principle is accompanied by an anti-evergreening mechanism. Yet guardians themselves may be captured; multiple redundancy and the continuous intervention of the dispossessed are structural means of anti-capture, but no permanent solution exists.
This chain is not a circle—circles return to the origin. This is an upward spiral. Each iteration operates at a higher level: the first paper's diagnosis works within mainstream political economy; the second turns to the logic of collective action after the failure of mainstream frameworks; the third asks how the new structure is to be maintained after collective action has broken the old. If the spiral continues, the next diagnosis will confront a more complex question: where the new institutions have deviated from the new norms, and the new norms themselves must be established in practice, not given a priori.
10.2 Return to Action
The end of the chain returns to the starting point: action.
The structure is unjust. Change requires solidarity. The new structure requires continuous maintenance, and maintenance requires solidarity.
This is not circular reasoning; it is the dialectic of structures. There is no once-and-for-all solution. The functional transformation diagnosed in the first paper will not disappear because of a single institutional design; it re-emerges in new forms within the new structure. The solidarity argued in the second paper does not become obsolete with the new structure's establishment; it must still operate within it, only shifting from "breaking the old structure" to "maintaining the new."
Solidarity under the new structure differs substantively from solidarity under the old. Old-structure solidarity was bare-handed: no institutional tools, no organizational experience, no monitoring capacity. New-structure solidarity gains three things: institutional tools (legal oversight channels, statutory participation rights, accountability mechanisms for public compute); organizational experience (the capacity accumulated through the operation of data trusts and public compute); and monitoring capacity (the early warning of regular ρARA auditing). The solidarity that maintains the new structure is not simple repetition but an upgrade—from "the force that breaks the structure" to "the immune system that maintains it."
Maintenance without end. Justice without completion. Only a continuous process: diagnosis, action, design, re-diagnosis, re-action, re-design.
This is not pessimism; it is honesty. Public compute may be captured; data trusts may be bought off; weight disclosure may be evaded in the name of "safety"; ρARA auditing may be pressured. But public compute may also not be captured; data trusts may persist; weight disclosure may, under sufficient social pressure, be implemented; ρARA may become a widely adopted governance tool. Which direction concretely prevails is not determined in theory but in struggle. What theory can do is point struggle toward a directionally correct exit. The road out of that exit must be walked by the actors themselves.
Structural justice cannot be declared "complete" on any given day; it is a responsibility each generation must assume anew. The first two papers argued why action is necessary; this paper has argued what is needed after action. Together they point in one direction: continuously, collectively, and in an organized manner, resist the re-concentration of the means of cognitive production; continuously, collectively, and in an organized manner, raise ρARA; continuously, collectively, and in an organized manner, maintain the justice of cognitive infrastructure.
This is not a choice; it is a condition. So long as the mechanism of concentration of the means of cognitive production exists, this condition will not change.