The Limits of Leviathan

Scale, evaluability, and the Dominion alternative

Ask a large bureaucracy who decided something and you will discover that nobody did. The clerk applied the rule. The rule came from a directive. The directive implemented a statute. The statute compromised between committees whose members have since retired, responding to pressures nobody now remembers. Every person in the chain acted correctly, and no person in the chain can tell you why the outcome is right, whether it is still serving its purpose, or who has the standing to change it. The decision happened. It was not made.

This is usually treated as a defect to be engineered away — more transparency, better incentives, clearer lines of accountability. I want to argue that it is a recurrent condition when coordination outruns a structure’s capacity to remain evaluable. The pressure follows from physical and informational constraints rather than from anyone’s failure. The standard reflex, whenever individuals fail to align their actions, is to add structure: centralize authority, impose shared rules, tighten oversight. The reflex pervades political theory, organizational design, and — most dangerously — contemporary AI alignment discourse. It often assumes that agency aggregates, that a sufficiently sophisticated coordinating structure can be an agent the way its members are agents, only bigger. That stronger attribution has to be earned, and scale makes it harder to sustain.

The formal argument lives in Coalitional Robustness in the Quantum Branching Universe and its architectural sequel Dominions: Plurality Without Closure, with the framework’s governance results gathered in Governance Without Gods. What belongs here is the conceptual shape of the argument and its political consequence — first the diagnosis, then one candidate architecture designed around it.

Institutional Agency as Evaluability

At the institutional layer, the relevant question is whether agency can be attributed to the organization rather than merely to its members. That stronger attribution requires evaluability: outcomes can be meaningfully assessed from within the organization, responsibility can be traced, commitments can be revised, and future actions can be endorsed under reflection. These capacities are not free-floating. Evaluation presupposes a perspective from which consequences can be interpreted. Responsibility presupposes identifiable authorship. Revision presupposes continuity between past commitments and present judgment.

Notice what this test does not require: benevolence, rationality, even competence. A vicious institution may still possess attributable agency — its actions can be assessed, its authorship identified, its course revised. Nor does a failure of collective authorship erase the agency of the people inside it. The test decides when organization-level language such as “the state chose” names more than the causal output of procedures. Institutional agency is a fragile achievement, not a default consequence of scale, and it has maintenance costs.

What Small Coalitions Do Well

Honesty about the dose-response curve matters here, because the case against large-scale coordination is often mistaken for a case against coordination as such. It is the opposite.

Before coordination becomes destructive, it strengthens agency. A small coalition — a family, a crew, a firm of a dozen people, a working group that fits around a table — can sustain redundancy, mutual error correction, and shared semantic load in ways no individual can. When I misread a situation, a partner who shares my context catches it. When my attention fails, someone else’s covers the gap. Interpretation is stabilized, local failures are caught, cognitive effort is distributed — and through all of it, authorship survives. Everyone at the table knows who committed to what and why. Within this limited regime, coordination strengthens the relationships that preserve agency. The mechanisms that will later drive collapse begin as scaffolding.

This is why the argument cannot be waved away as temperamental anarchism. The claim is not that coordination is bad. The claim is that it has a dose-response curve, that the curve can turn over, and that management skill cannot abolish the underlying constraints.

The Migration of Interpretation

As a coalition grows, coordination comes to rely on abstraction. Information must be compressed to travel; interpretations must be standardized so that strangers can rely on them; decisions must be routed through procedures rather than situated judgment, because situated judgment does not scale. Each of these changes is locally rational. Each reduces ambiguity and increases efficiency. And each weakens the link between action and understanding.

The cumulative effect is a migration. Interpretation moves out of individual agents and into the coordinating structure itself — into the forms, the classifications, the escalation paths, the metrics. Decision-making becomes procedural. Responsibility diffuses across the chain until it belongs to no one. Evaluation detaches from authorship: the system’s outputs can still be measured, but there is no longer any perspective inside the system from which they are owned.

The coalition has begun to function as a mechanism rather than as a collective of agents. Nobody chose this. It is what compression does.

Leviathan

The paper gives this outcome its old name and a precise new sense. A Leviathan is a large-scale coordinating structure whose internal evaluability has collapsed. It continues to act, optimize, and enforce constraints, but it lacks a coherent internal perspective from which its actions can be reflectively endorsed, revised, or owned. The definition turns on structural loss of attributable institutional authorship — not on intent, ideology, or moral character.

That last clause carries the political payload, so let me state it plainly. Libertarians have spent generations indicting the state as predatory, and much of this volume has pressed that indictment. But the deepest charge against the state is not that it is evil. Evil would be an improvement. Evil implies an author — something with a perspective, something that could be reasoned with, deterred, reformed, or held responsible. The deeper charge is that once evaluability collapses the state is not a collective agent in that reflective sense. It is a mechanism executing inherited constraints without reflective access to their justification, though the individuals inside it remain agents and particular offices may retain attributable responsibility. When we say “the government decided” or “the state wants,” we risk attributing a unified evaluating perspective to a structure that may no longer have one. The bureaucracy of the opening paragraph is not necessarily hiding a single decision-maker.

This is why reform so reliably disappoints. Reform is addressed to an agent — it assumes something inside can hear the argument, weigh it, and revise. A Leviathan can be perturbed, redirected, defunded, or dismantled, but it cannot be persuaded, because persuasion requires a listener and evaluability has left the building.

The Thermodynamics of Comprehension

Why does this pressure recur? Because maintaining evaluability across a large coalition is physically expensive. Shared understanding is not free: it requires continuous investment in information fidelity, interpretive alignment, and contextual preservation. Every layer of hierarchy, every additional participant, every added domain of responsibility raises the cost of keeping the whole structure comprehensible to itself.

As systems scale, these costs can outrun the benefits of coordination. Centralization often buys efficiency by discarding interpretive resolution — by compressing away the context that made evaluation possible. The loss is cumulative and hard to reverse because restoring institutional authorship requires reconstructing context already thrown away. You cannot decompress a file whose details were deleted in the compression. Leviathan is therefore a recurrent risk under large-scale coordination, not proof that every large institution has already collapsed into it. Transparency, modularity, narrow jurisdiction, traceable delegation, and real exit can move the threshold; once those safeguards are absorbed into procedure rather than restoring authorship, exit and rebuilding at viable scale may be the only remedy left.

Alignment Is Downstream of Agency

The implication for artificial systems is immediate enough that alignment researchers should find this chapter uncomfortable. Alignment presupposes an agent — something capable of understanding, endorsing, and revising its actions in light of consequences and authorization. When a system exceeds the scale at which evaluability can be preserved, alignment ceases to have a well-defined referent. There is nothing left to align.

Worse: the standard prescriptions can accelerate the disease. Expanded oversight, tighter control, and increased centralization intensify the coordination pressures that dissolve agency unless they restore traceable judgment. Pursued without that constraint, they yield a system that enforces policy without comprehension and optimizes objectives without accountability: alignment’s stated nightmare, arrived at through alignment’s own methods. Alignment is downstream of agency, not a substitute for it. The same holds for institutions. Accountability mechanisms help only when they preserve identifiable authorship, bounded jurisdiction, and revision; otherwise they are absorbed into the procedural mass they were meant to audit.

What survives this analysis is a narrow region: coalitions bounded or modular enough that shared context is preserved, authority remains revocable, decision pathways remain traceable, and exit remains feasible without systemic collapse. The region resists unbounded scaling by its nature. Its stability depends on limits rather than ambition, and extending it without preserving those properties converts coordination into mechanism. Cooperation succeeds only when it respects the structural constraints imposed by agency itself.

The Constructive Question

So far this is demolition, and demolition alone would leave the volume’s project incomplete. Against Utopia ruled out final world designs; the Sacrifice Pattern showed that optimizing systems collapse when performance improves through captive agency loss; this chapter has rejected the assumption that large benevolent structures acquire attributable agency merely by coordinating at scale. Together these constraints pose a question that deserves a constructive answer:

What kind of social architecture remains coherent once agency, pluralism, and value drift are treated as structural facts?

One proposed answer is an architecture, not a moral theory and not a world design. It is called the Dominion architecture. Its conditional claim is deliberately narrow: among governance architectures that preserve agency, reject outcome coercion, and satisfy the model’s assumptions, federated virtual Dominions are structurally optimal within their domain. The domain is digitally mediated social space — a scope restriction I will return to, because it is not a footnote.

A Dominion is a sovereign virtual jurisdiction created by an agent or group of agents, defining its own rules, norms, and internal affordances. Five properties define the architecture. Entry is voluntary: agents join only by invitation and explicit consent to the rules. Enforcement is bounded: the only sanction available to a Dominion is expulsion — no punishment, no fines, no coercive penalties. Exit is supreme: agents retain the unconditional ability to leave any Dominion at any time. There is no global value aggregation: the system does not rank Dominions, reconcile their values, or enforce shared outcomes. And the shared substrate is thin: it enforces identity persistence, consent verification, capability isolation, and expulsion — and adjudicates nothing else, not meaning, not morality, not success.

Readers of Anarchy Is Not Chaos will recognize the exit criterion doing the same work it did there: a Dominion may be internally hierarchical and remain admissible only while exit is informed, realistic, asset-portable, and unpunished, with voice and remedy for burdens that exit cannot cure. Dominions take part of that principle and make it constitutive — built into the substrate rather than depending only on an operator’s restraint.

Why Exit Requires Asset Portability

Exit supremacy sounds like a slogan until you ask what makes exit real, and the answer is a causal chain worth spelling out link by link.

Exit has meaning only when it remains affordable. For that reason the architecture requires that persistent identity, assets, and reputation belong to agents rather than to Dominion operators. A Dominion may deny you continued participation; it cannot confiscate what gives you continuity across jurisdictions. This requirement is constitutive, not a convenience, because its negation runs straight downhill to collapse: when operators control assets, exit costs rise. Rising exit costs convert voluntary membership into captivity. Captivity reintroduces sacrifice gradients — populations that can be made to absorb costs because they cannot leave. And sacrifice gradients reintroduce exactly the collapse dynamics that Against Utopia traced, the standing sacrifice this architecture exists to exclude by construction.

Anyone who has tried to leave a platform that owns their identity, their audience, their purchase history, and their social graph has felt the first links of this chain personally. The Dominion architecture treats that feeling as a design constraint of constitutional rank.

Airtight Vessels

The second constitutive constraint governs relations between Dominions. They exist as capability-isolated execution contexts: no direct access to one another, no substrate resources beyond assigned quotas, no writes to shared state except through explicit, substrate-mediated bridges. The design will be familiar to engineers as operating-system process isolation or object-capability security.

The visceral version: separate airtight vessels rather than a shared room. In a shared room, peace depends on everyone agreeing to refrain from violence, and on enforcement when someone defects. In separate vessels, nobody must agree to anything, because violence across the boundary is not physically possible. Inter-Dominion aggression is infeasible rather than prohibited — a constitutive constraint, not a behavioral rule. This is the same move the whole architecture makes, applied at the boundary: wherever possible, replace norms that must be enforced with structures under which the violation cannot be expressed.

Freedom Density and Drift

What does the architecture optimize? Not happiness, not welfare, not any global objective — those were ruled out three chapters ago. The proposed quantity is freedom density: the number of distinct value-consistent future trajectories available per unit of coercive constraint. It rises in the model when constraints remain local to voluntary contexts, enforcement remains minimal and reversible, and divergence across jurisdictions remains unconstrained. The paper’s sharper result is conditional: within its specified space of non-coercive, agency-preserving architectures, no alternative improves outcomes for some agents without reducing them for others. That result does not establish frictionless exit, complete preferences, harmless substrate governance, or physical-world optimality; those assumptions and limits govern how far the conclusion travels.

The architecture’s answer to value drift completes the picture. Agents change; values evolve; any governance design that pretends otherwise is utopia wearing engineering clothes. Nation-states, federations, and consensus communities handle drift through reform — collective renegotiation in which everyone must adjust when anyone diverges, which is why their politics is permanent and embittered. Dominions handle drift through mobility. When an agent’s values change, the agent moves. No one else must change in response; no system-wide renegotiation occurs. The principle deserves to be stated as the design maxim it is: localize coordination, globalize exit. Coordination stays small enough to remain inside the viable region where evaluability survives; exit operates globally, so that no local arrangement can become a trap. The two halves of this chapter are the two halves of that maxim.

There is a corresponding honesty about what the architecture declines to provide. No shared meaning, no large-scale coordination, no global public goods, no epistemic convergence. Agents who want such goods may build them voluntarily within Dominions; the substrate refuses to enforce them across the federation. This is not indifference. It is the discipline of a design that has understood the diagnosis: every global good enforced from above is a scale commitment, and scale commitments are how Leviathans begin.

Lineage, Scope, and the Rhyme

The lineage should be acknowledged. Nozick’s Anarchy, State, and Utopia¹ closed with a framework for utopias — a meta-structure hosting many communities, none authorized to impose itself on the rest. Dominions are recognizably that proposal re-founded. But the grounding has changed: where Nozick argued from moral rights, this architecture argues from agency preservation, the non-composability of value, and robustness under drift — and it enforces its guarantees through technical constraints rather than through anyone’s respect for rights. The framework for utopias stops being a moral aspiration and becomes an engineering specification.

And the scope restriction must be kept honest, because the argument’s precision depends on it. The optimality claims apply to the governance layer of digitally mediated interaction — the domain where jurisdictions can be instantiated at will, isolation can be made airtight by construction, and exit can be low-cost and asset-portable. Physical scarcity, energy, biological dependency, and material political economy persist outside that layer, and no governance architecture dissolves them. Exit to Protocol argued that the transition away from coercive governance begins where coercion is weakest; this chapter has specified what the destination must preserve. But an architecture for digital space is not yet an answer for bodies, which cannot be forked and must breathe.

That harder problem has a proving ground, and the design constraints travel better than one might expect. A settlement on Mars — the extreme case of physical dependency, where even the air is infrastructure — will need analogous guarantees: exit that is engineered rather than promised, assets that belong to agents rather than operators, enforcement that cannot metastasize into sovereignty. Exit supremacy and asset portability in virtual space; engineered exit and portable assets under physical dependency. The same dependency rule across two different substrates is the structural rhyme on which this volume closes, and cashing it out is the work of Axiocracy.

The state’s defenders have always had one argument that outlives every rebuttal: and yet it coordinates. This chapter concedes the fact and denies the conclusion. Coordination is real, valuable, and bounded — and beyond its evaluable bounds, what coordinates may cease to be traceably anyone’s. Leviathan does not need to be slain. It needs to be recognized as what it is — a mechanism mistaken for a mind — and outgrown by architectures that keep coordination owned and exit real.