Chapter 11Decentralization and Trust

Who should make the decisions? The question haunts every organization, every government, every system that must coordinate beyond a handful of people. The answer, as both history and physics reveal, depends on what you are trying to survive.

In 1969, two computers exchanged the first message across ARPANET, the precursor to the internet.⁶ The message was supposed to say “LOGIN.” It crashed after two letters: “LO.”

From those two letters came a network that would eventually connect billions of devices. The key to its success was a design decision that seemed, at the time, almost perverse: the network had no center.

Traditional communication networks (telephone systems, broadcast television) were centralized. Messages flowed through hubs. If a hub failed, the network failed. ARPANET’s designers chose differently, drawing on packet-switching ideas that Paul Baran had developed at the RAND Corporation for a communications system that could survive attack.⁷

They created a mesh: every node connected to multiple others, messages broken into packets that found their own routes through whichever links were open. Alternate routes reduced dependence on individual links and nodes.

The result was less efficient (packets sometimes took strange routes) but far more durable. Nodes could be destroyed, connections severed, the topology fragmented, yet the remaining pieces kept working. Packets found new paths.

This is the trade-off governing all complex systems: efficiency versus resilience. Centralization optimizes for efficiency, decentralization for resilience. As complexity grows, resilience tends to win.

Decentralized systems are compositional: autonomous modules interact through well-defined interfaces, with no single node holding a picture of the whole. Think of LEGO bricks snapping together through standard connectors. Each brick knows only its own shape and its connection points, yet the whole structure holds. The assembly’s behavior follows from composing the behaviors of its parts.

The bias toward compositionality may be deeper than design. Shai, Riechers, and colleagues showed that transformer networks trained on next-token prediction spontaneously decompose their internal world models into factored parts. Each part is represented in its own orthogonal subspace.⁴⁹⁴ Factored means the network keeps separate books for separate things, the way a household ledger keeps rent in one column and groceries in another instead of a single running total. Orthogonal subspaces are what keep the columns from bleeding into each other: each part of the world gets its own directions inside the network, and a change in one leaves the others where they were.

This happens even when the training data contains no explicit signal that it is decomposable. The networks favor this factored structure early in training even when it costs predictive accuracy: dimensional efficiency over fidelity, modularity over monolith.

Gradient descent, the step-by-step learning procedure that trains these networks, has an inductive bias (a built-in preference for some kinds of solution over others) toward the same compositional architecture that ARPANET, TCP/IP, and Mission Command converge on. The pattern may be substrate-neutral.⁴⁹⁵

A centralized planner maintaining a representation of the whole faces a growing integration burden. Central services can still compose through modular interfaces. The internet works because its core protocol (TCP/IP) composes: two networks speaking the same protocol can join without a third party managing the connection. The Soviet planning bureau failed because central allocation does not compose. Every new factory required updating the global plan.

Machine learning shows why this asymmetry is built into the cost of learning itself. A network that must recognize a cat could memorize every pose on its own: the left-facing cat, the same animal facing right, shifted upward, rotated a little. Each pose would be a separate lesson. Or it could be built to know in advance that moving an object does not change what it is, so one detector serves every position. Building in that symmetry shrinks the space of possibilities the network must search, and it needs far less data to generalize.⁴⁹⁶

An agent-independent protocol can admit newcomers without a new case-specific rule, as TCP/IP illustrates. Equivariance can reduce representation cost when the target itself has the relevant symmetry. A privileged rule can also remain compact: “benefit participant 1” extends unchanged as others join. Fairness and protocol reuse can have practical advantages, while their description cost, enforcement cost, and moral value remain separate questions.⁴⁹⁷

The Knowledge Problem

In 1945, Friedrich Hayek published “The Use of Knowledge in Society,” an essay that would shape debates about economic organization for decades.¹

His argument was devastating. The knowledge needed to coordinate a complex economy is dispersed across millions of minds: the farmer who knows her soil, the manager who knows his machines, the consumer who knows her preferences. No central planner can gather it all.

Markets solve this problem through prices. Prices aggregate dispersed information. When a shortage appears, prices rise, signaling producers to make more and consumers to use less. No one needs to know why the shortage exists. The price does the coordinating.

Central planning requires the impossible: that the planner know what millions of individuals know. The Soviet Union tried, and failed, because knowledge resists centralization. The planners’ intelligence was beside the point.

The planner faces a second impossibility, Taleb’s Turkey Problem again (Chapter 5): all planning rests on observed history, and observed history does not contain the regime changes that matter most. A turkey fed every day for a thousand days finds on day 1,001 that the accumulated data offers no warning of what is coming.

The Soviet planner in 1985 had abundant data on decades of Soviet functioning. None of it contained the collapse of 1991. Markets survive regime changes because they require only that someone, somewhere, respond to whatever arrives.

Hayek’s argument is economic, and it also has a thermodynamic shape. Information is dispersed because the systems that generate it are dispersed. Concentrating information is like reversing entropy: possible locally, at great cost, only imperfectly. Flow wants to spread.

Sheaf theory, the branch of mathematics that studies how local data fit together globally, formalizes Hayek’s insight.⁴⁹⁸ A sheaf assigns data to each local region and specifies compatibility conditions: rules for when local observations stitch together into a consistent global picture. Think of a jigsaw puzzle where each person holds one piece. Whether the pieces fit depends on whether neighboring pieces share compatible edges.

A sheaf guarantees gluing when local sections agree on every overlap. A sheaf-theoretic obstruction concerns a specified local-constraint problem with no compatible global solution. In a contextuality model, each measurement context permits certain outcomes, yet no assignment satisfies all the contexts together. A social coordination claim requires the corresponding local data, overlaps, and compatibility rules.

This is no abstract curiosity. Abramsky (2014) showed that this obstruction shares the same mathematical structure as Arrow’s impossibility theorem (Chapter 10), the proof that no aggregation method satisfying basic fairness conditions can compress many perspectives into one without losing something essential.

Coercive aggregation, forcing local preferences into a single global ranking, runs directly into this barrier. Voluntary coordination can proceed through local agreements without demanding a single global ranking. Their compatibility still needs to be established; consent alone does not remove a preference cycle.

Market prices assemble information through transactions; central plans gather and process it through administrative channels. Arrow’s theorem concerns a social ranking of at least three alternatives under unrestricted domain, Pareto, independence of irrelevant alternatives, and non-dictatorship. It constrains any aggregator satisfying those premises, including voluntary ones. Voluntariness alone guarantees neither consistent rankings nor compatible local sections.

Mission Command

The knowledge problem explains why centralized planning fails in economics. The military learned the same lesson on the battlefield.

Figure 11.1: The diagram above contrasts two approaches to coordination. On the left, Detailed Command: every instruction flows from the top, creating bottlenecks and delays that worsen as the organization grows. On the right, Mission Command: the leader communicates purpose and constraints, then trusts each unit to act on local knowledge. The first approach is precise and brittle. The second is flexible and scalable.

In the nineteenth century, Prussian military theorists confronted a problem: how to coordinate hundreds of thousands of soldiers across a battlefield too large for any general to observe. The traditional answer was Detailed Command: explicit orders specifying what each unit should do, when, and how.

Detailed Command broke down under real conditions. Orders arrived late, conditions had changed, subordinates followed instructions that no longer applied. Anyone who has worked in a large organization recognizes the pattern.

The Prussian solution was Auftragstaktik, literally “mission tactics,” or Mission Command.⁴⁹⁹ Instead of specifying actions, commanders specified intentions: the goal, the constraints, the resources available. Subordinates were trusted to achieve the goal given local conditions.

Auftragstaktik demanded trust in both directions. The commander had to trust that subordinates understood the larger strategy. Subordinates had to trust they would not be punished for deviating when circumstances warranted. Both needed a shared understanding of what success looked like.

Mission Command is compositional in exactly this sense. Two units operating under the same intent can coordinate at their boundary without routing through headquarters. Their shared understanding of purpose serves as the interface, the way two LEGO bricks connect through their standard pegs. Detailed Command is non-compositional. Every inter-unit coordination must pass through the central node, creating a bottleneck that worsens as units multiply.

When it worked, Mission Command achieved results that centralized planning could not match. The German army conquered France in six weeks in 1940; the previous war’s attempt had stalled into four years of trench warfare and ultimate failure. Concentrated armor, radio coordination, and air superiority all drove that speed. So did Mission Command, which put thousands of independent decisions in the hands of officers empowered to exploit opportunities as they found them.

When it failed (trust absent, intent misunderstood, or local decisions diverging from strategic necessity), the results were catastrophic. Mission Command trades rigidity for the risk of fragmentation. The trade-off is favorable only when trust is high and conditions are volatile.

Neuroscience reveals the same architecture at the level of neural circuits. In the cortex, slower beta waves carry the commander’s intent: the current goal, the task rule, the constraint shaping behavior. Faster gamma waves execute locally, filling in sensory detail within the boundaries beta defines. The beta wave does not specify which gamma pattern fires. It sets the agenda; local circuits decide how to carry it out (Chapter 8b).⁵⁰⁰

The mammalian cortex has run this architecture for roughly 200 million years. The Prussian army formalized it in the nineteenth century. Two systems separated by every conceivable variable (substrate, scale, history) converge on similar coordination architectures, the framework argues, because the thermodynamic constraints are shared.

Mission Command reduces the detail a headquarters must transmit by combining shared intent with local judgment. Mermin-Wagner supplies a different, tightly specified physical result: equilibrium short-range models with continuous symmetry lack spontaneous long-range ordering at positive temperature in one or two dimensions in the thermodynamic limit.⁵⁰¹ Organizational degree, cycles, and reporting levels do not identify that dimension. The theorem places no binary-only limit on instructions, and does not prove Mission Command’s effectiveness.

Adrian Bejan, the physicist who formulated the Constructal Law (the principle that flow systems evolve toward designs that provide easier access to their currents), arrived at a compatible conclusion from thermodynamics:²² “Without economic and social freedom, societies are doomed to starvation and stagnation.” Bejan’s claim is that flow optimization requires access: channels blocked by fiat cannot evolve toward greater throughput. The claim is stronger than the Constructal Law alone warrants (the law describes how flow systems evolve, not what political arrangements they require), yet the directional prediction aligns with the evidence from Soviet central planning, ARPANET’s resilience, and Mission Command’s effectiveness above.

Learning theory offers a conditional efficiency comparison. Depth-separation results show that specified function families can need exponentially more shallow-network capacity than an appropriate deep architecture at a specified accuracy. Simple affine functions need no such expansion. Universal approximation concerns appropriate functions and domains; these results support testing distributed architectures for a task, without proving that every complex mapping or organization requires depth.

Vanchurin observes: a political system where one person commands and the rest execute is a shallow network, poorly adapted to learning. Deep, multi-layered social networks with feedback, social mobility, and distributed decision-making are systems optimized for learning.⁵⁰²

The argument for decentralization is computational: centralized command faces information bottlenecks whose severity depends on memory, bandwidth, models, and the task.

The Hajnal Line pattern (Chapter 10) is the historical signature: broader trust topologies emerged where medieval marriage law extended cooperation beyond kin networks. Those broad-trust architectures are the kind of deep, multi-layered social networks Vanchurin’s framework identifies as optimized for learning.

The oldest recorded transition from Detailed Command to Mission Command may be the one that ended the Bronze Age. In 1976, the psychologist Julian Jaynes argued that pre-collapse Near Eastern civilizations ran on a form of Detailed Command so literal it was neurological.⁵⁰³ An auditory hallucination, generated by the right hemisphere and interpreted as a god’s voice, issued direct instructions. The left hemisphere executed without deliberation.

The literal claim that ancient peoples were unconscious is almost certainly wrong. The weaker structural claim holds up better.

Cuneiform tablets from the reign of Tukulti-Ninurta I, written at the threshold of the collapse, describe personal gods falling silent with the matter-of-fact distress of someone reporting a mechanical failure: “My god rejected me. He disappeared. My goddess left. She departed from my side.” The nobleman who wrote these words tried divination, dream interpretation, and exorcism to restore the voice. Nothing worked.

When the voices stopped for the ruling class, the coordination structure they sustained collapsed with them. On that structural reading, the Bronze Age Collapse, a cascading failure across the entire Near East, is what happens when a Detailed Command system exceeds its complexity ceiling and no Mission Command architecture exists to replace it.

What replaced it, centuries later, was the internalization of divine authority as conscience: principles held inside the mind rather than commands received from outside it. The words “consciousness” and “conscience” share a Latin root, con-scientia, knowing-with-oneself. The oldest name for self-awareness is a name for internalized moral authority.

A structural parallel has since emerged in contemporary language models, developed in Chapter 21.

As Skinner Layne put it: “If you manage an organic system as if it were an inorganic system, you will eventually succeed in turning it into one.”⁵⁰⁴ Detailed Command treats living organizations as mechanisms, eliminating the variance that is their adaptive capacity. Manage a forest as a timber farm and you get a timber farm: uniform, efficient, one pest away from collapse.

Kauffman’s patch procedure formalizes the principle.⁵⁰⁵ Consider a problem where many interacting components each affect the others’ success. Three strategies exist. Optimize globally from a single point (Kauffman calls this the “Stalinist” solution). Fragment into completely independent units. Or partition into semi-autonomous patches that each optimize locally while coevolving with their neighbors.

Imagine a mountain range with many peaks. The global optimizer tries to find the highest peak by surveying the whole range from one vantage point, missing the ridges hidden behind closer hills. The fragmented units each climb their nearest peak, ignoring better options over the next ridge. The patch strategy lets groups explore their local terrain while comparing notes with neighboring groups, so good routes spread.

Kauffman showed through extensive simulation that the intermediate strategy wins. Global optimization gets trapped on the rugged landscape created by conflicting constraints. Complete fragmentation loses the benefits of coordination. Patches at the boundary consistently find better solutions: large enough to capture local structure, small enough to avoid the ruggedness trap. The optimal patch size falls at the edge of chaos, the productive border between rigid order and randomness.

A further result deepens the point: systems that deliberately ignore a small fraction of incoming constraints (roughly five percent) find better global solutions than systems attempting to satisfy all constraints at once.⁵⁰⁶ Honoring every constraint creates conflicting demands that trap the system in mediocre compromises. Selective inattention frees the system to navigate around traps.

Mission Command is the patch procedure applied to organizations. The commander defines the fitness landscape (the intent). Each unit is a patch, optimizing locally within constraints set by neighbors and the overall mission. No patch requires global knowledge. The compositional structure that makes ARPANET robust makes Mission Command effective, for the same structural reason: no part needs a picture of the whole.

A contemporary case from AI engineering illustrates the pattern. Anthropic’s autonomous coding harness (Rajasekaran, 2026) separates three agent roles: Planner, Generator, and Evaluator. The Planner specifies the deliverable (“build a digital audio workstation with these capabilities”). The Generator implements freely, choosing its own architecture, framework, and approach. The Evaluator grades the output against concrete criteria through live testing, failing builds that fall short.

The team arrived at the same principle independently: “Constrain deliverables, not paths.” The Planner’s specification is commander’s intent. The Generator’s freedom is local execution. The Evaluator’s criteria are the compositional interface through which quality compounds without central specification of method.

A solo agent given the same task produced broken, unplayable software in twenty minutes. The harnessed system, coordinating through intent rather than instruction, produced polished working software. The computational overhead was higher; the output crossed the threshold from toy to tool.⁵⁰⁷

A second case extends the principle from fixed architecture to learned architecture. Nielsen et al. (2026) trained a small model through reinforcement learning to coordinate much larger models by designing subtasks and communication topologies from scratch.⁵⁰⁸ The training signal is purely end-to-end: did the collective output solve the problem?

From this signal alone, the Conductor discovers Mission Command. It specifies intent (“Develop an efficient algorithm”), assigns the right worker, and lets each worker execute freely. Hard problems receive elaborate coordination with multiple planners, then implementers, then verifiers. Simple problems receive a single direct query. The system allocates coordination depth in proportion to task difficulty: the patch procedure discovered computationally.

The most striking emergent behavior is role abdication. The Conductor, whose entire function is to coordinate, sometimes decides the optimal strategy is to hand its coordinating role to a more capable worker: “Here, you figure out the subtasks for the others.” Nothing in the reward signal incentivizes humility directly. Correct answers are the only thing it scores. Abdication sometimes produces correct answers, which suggests the system has learned computational subsidiarity (making each decision at the lowest competent level): the best coordinator is sometimes the one who recognizes it is not the best coordinator for this particular problem. Centralization is itself a variable the system learns to optimize, a structural assumption it discovers rather than inherits.

A corporate-scale case arrived in 2026. Block (the payments company formerly known as Square) announced a reorganization replacing middle management with AI-mediated coordination.⁵⁰⁹ The architects traced the lineage explicitly: from the Roman contubernium (eight soldiers sharing a tent under a decanus, their squad leader) through the Prussian General Staff to the matrix organization. Each solved the same constraint: a leader can manage three to eight people effectively, and that bandwidth limit determines every hierarchy’s shape.

Block’s answer was to replace what those layers do. A “world model” of the company’s operations provides shared context to edge workers, who act without waiting for information to route through management layers. “Directly Responsible Individuals” own cross-cutting problems with full authority to pull resources from any team. The company’s architects arrived at Mission Command without using the term: constrain intent, not method; trust the edge; let shared context replace hierarchical relay.

The framework predicts a failure mode Block has not yet addressed. A shared world model is a single map. When the map is accurate, every ship sails together. When one coastline is drawn wrong, every ship runs aground on the same reef.

Old-fashioned hierarchy distributes map-making across many managers, each with a slightly different picture of the territory. Their errors disagree, and disagreement surfaces the mistake before it becomes a decision. The hierarchy’s blind spots are decorrelated: different managers are blind to different things. When errors are also unbiased and suitably aggregated, decorrelation helps them average out. Shared bias survives.

The 2008 financial crisis showed what the alternative costs. Nearly every major bank used the same Gaussian copula risk model, a formula for estimating the chance that many loans fail together. When that model was wrong about mortgage correlations, every bank was wrong in the same direction at the same moment. The shared model was a speed advantage in ordinary times and a correlated catastrophe when reality exceeded its assumptions.

Ising Monte Carlo simulations (the magnet model of Chapter 8b) reproduce both sides of the trade-off in a lattice model (Experiment AY1).⁵¹⁰ A shared signal at high strength maintains coordination that independent local signals destroy (|m| = 0.63 vs 0.06). The quantity |m| measures how far the population leans one way: 1.0 is unanimity, 0 is an even split. So 0.63 is a working consensus, and 0.06 is a crowd with no common direction at all.

Recovery from a correct shared signal is faster than from distributed ones (155 vs 251 sweeps), because correction, like error, reaches everyone at once. The correlated advantage and the correlated vulnerability are the same mechanism, seen from opposite sides. On average, shared-model systems outperform hierarchies. The risk lives in the tails: rare catastrophes where every agent is wrong about the same thing at the same moment.

A follow-up simulation (Experiment AY1b) tested the hybrid architecture directly. A shared low-noise field (intent) combined with independent high-noise local fields (tactics) reaches 94% of the shared model’s steady-state magnetization while maintaining near-perfect recovery from model error: it regains 99.4% of its pre-disruption coordination, against 86.8% for the distributed model. When only a tenth of each agent’s signal comes from the shared intent, the hybrid bounces above its pre-disruption level (recovery 107.7%), the antifragile signature: it comes back stronger than the shock found it.

Finite-size scaling (Experiment AY1c, L = 32 to 128, where L is the side of the simulated lattice and a system holds L squared agents) showed that recovery depends on the tested size and signal mixture. At L=96, mean recovery was 0.910 for shared intent, 0.923 for the 50% hybrid, 1.018 for the 25% hybrid, and 0.928 for the 10% hybrid. Shared intent exceeded all tested hybrid means at L=128, which does not locate a general crossover at L=96. A wider coordination sweep also found that the shared-versus-distributed direction depended on noise at every tested size. A firm or battalion is a finite system, but these lattice comparisons alone do not identify its best signal mixture.

Aristotle observed that collective judgment often exceeds individual excellence: the feast to which many contribute surpasses one prepared by a single cook.³ Divergent perspectives, preserved and integrated rather than overridden, exceed what any single viewpoint could achieve.

Work in the “It from Qubit” programme (connecting quantum information theory to gravity) shows that holographic spacetime, in the simplified model universes where the idea has been worked out (anti-de Sitter space), functions as a quantum error-correcting code.¹⁷ Holographic keeps its optical sense here: everything happening inside a region of space is encoded on the surface that bounds it, the way a flat sheet of film stores a whole three-dimensional image. Error-correcting codes spread information across many locations so that losing any single piece does not destroy the message. The same principle explains why your music still plays when a CD is scratched.

Spacetime’s deep geometry is protected against local disturbances because its information is globally distributed. Data can be lost at particular points on the boundary and the geometry inside survives.

Mission Command has this structure. When a commander transmits intent rather than instructions, the encoding is distributed across every subordinate who internalizes the purpose. An individual can fail, misunderstand, or be cut off while others retain the intent. Mission success still depends on the skills and resources that survive.

Detailed Command can leave units dependent on instructions from headquarters. If the plan and messages lack redundant copies or corrective channels, losing headquarters or garbling a message destroys that information. Detailed orders can also be cached, replicated, and error-corrected.

Cognitive scientist Douglas Hofstadter drew on neuroscientist Karl Lashley’s finding that memory is equipotential: distributed across neural tissue rather than stored in any single location. Hofstadter likened it to a telephone network: “Destroying any local part of the network would not block calls; it would just cause them to be routed around the damaged area.”²³ The same principle operates at every scale.

The Data Rate Theorem, as Rodrick Wallace applies it (Chapter 4c), connects this to scaling limits. Coordination stability requires a minimum information transfer rate. That might seem to favor the detailed rulebook and its heavy traffic, but the rate that counts is the information that gets through. Mission Command can reduce live traffic by distributing shared principles in advance. Detailed orders can also be redundantly encoded or retransmitted. When links fail, a cached principle remains available; the fate of an order depends on its coding, storage, and delivery protocol. Distribute the encoding, and the deep structure becomes robust.

Robustness alone gives an inert geometry, and the newest work in the same programme shows why. The first holographic codes protected their encoded information exactly: every piece of the interior could be recovered perfectly from enough of the boundary. A spacetime built that way is indestructible and unresponsive. Matter placed inside it cannot bend it, because the sector encoding the geometry and the sector encoding the contents never mix. Charles Cao, John Preskill, and colleagues showed in 2026 that this stiffness is intrinsic to exact recovery, and that letting matter curve space requires an approximate code, one in which the two sectors leak slightly into each other.⁵¹¹ The physicist Bartek Czech put it plainly: the imperfection of the code is “the reason Newton’s apple fell on him.”⁵¹²

Read as coordination, the lesson has two halves. Distribution gives a structure that survives damage. A slight, deliberate porousness between the structure and what it holds is what lets the structure respond to what it holds. Mission Command has both. Intent is spread across every subordinate, so it survives any single loss, and it is held loosely enough that what subordinates learn at the front can reshape the intent itself. A doctrine protected perfectly against its own practitioners would be as unresponsive as the exact code’s spacetime: nothing they did could curve it.⁵¹³

Vanchurin’s geometric learning dynamics (introduced in Chapter 15b, with its learning regimes set out in Chapter 16) gives Wallace’s threshold a spatial reading: the point where the coordination landscape’s geometry can no longer track environmental perturbation. When perturbations change faster than the geometry adapts, efficient coordination collapses into rigid equilibration: the system stops adapting and merely settles. Mission Command degrades into Detailed Command.⁵¹⁴

A complementary limit comes from mathematics. Gödel’s first incompleteness theorem concerns effectively axiomatized, consistent formal theories with sufficient arithmetic strength: such a theory cannot decide every arithmetic statement.¹⁹ Detailed Command faces a practical comparison, the unexpected case that its rulebook leaves unresolved. Establishing that governance limit requires evidence about the rulebook and the cases it must handle; the arithmetic theorem supplies no universal failure result for institutions.

Mission Command gives local decision-makers discretion to interpret intent as conditions change. Principles guide judgment beyond the written rule-set. The commander given “hold the bridge” can adapt instructions to the situation, within the limits of available information and capability. Meta-level reasoning still has computational limits.

Holographic error correction, incompleteness, and the Data Rate Theorem pose different questions: recovery from errors, limits of formal arithmetic, and communication for stabilizing a plant. They inspire comparisons with governance. They supply no shared theorem that distributed principles persist while concentrated rules fail; a detailed command can be redundantly encoded, and a simple centralized rule can stabilize a plant.

A fourth line of evidence comes from condensed matter physics. Topological insulators are materials that conduct electricity on their surfaces while remaining insulating in their interiors. They belong to the broader field of topological matter, whose foundations earned the 2016 Nobel Prize in Physics, awarded for “theoretical discoveries of topological phase transitions and topological phases of matter.” Their conducting surface states are protected by topology together with time-reversal symmetry, with protection against applicable perturbations that preserve the symmetry, the bulk gap, and the topological phase.²⁰

Think of the difference between a chain and a knot. A chain breaks if you remove one link. A knot’s essential character, its minimum crossing number and, for a chiral knot, its handedness, survives being stretched, twisted, or smeared with mud. These properties are topological: they change only through cutting and re-joining, never through gradual deformation.

In a conventional crystal, order depends on every atom conforming to the lattice pattern. A single line of defects can propagate and shatter the structure.

Topological protection works differently. Under the required symmetry and bulk-gap conditions, the protected surface states resist suitable impurities, disorder, and local perturbations because the property they encode, a topological invariant (a whole number characterizing the system’s global shape), cannot change by small degrees. You cannot “partially erode” it any more than you can have half a hole in a doughnut. Either the global topology holds, or it undergoes a topological transition, which can be continuous.

Rules-based systems (detailed codes, enumerated prohibitions, exhaustive compliance checklists) are crystalline order. Each rule is a lattice point. A single inconsistency is a dislocation that can propagate, eroding trust in the entire structure. Each exception weakens the next rule’s authority.

Principles-based systems (Mission Command, compass values, shared intent) are topological order. The orienting principle (“hold the bridge,” “maximize optionality by invitation”) is a topological invariant. Local failures do not erode it, because principles encode a global orientation that survives local noise. A compass that occasionally trembles still points north.

The topological insight sharpens a practical warning. The transition from principles to rules is a topological phase transition, a change in the kind of protection the system offers. When an organization replaces “use good judgment” with a 200-page compliance manual, it changes its governance topology from robust to fragile. The organization trades perturbation-resistant orientation for defect-vulnerable crystalline order.

Four independent arguments (holographic error correction, Gödelian incompleteness, the Data Rate Theorem, and topological protection) now converge on that conclusion.

A fifth line of evidence pins down the mechanism. The convergences above are suggestive; they could be coincidental. Machine learning provides a case where the mapping is exact.

Physicists Pankaj Mehta and David Schwab (2014) showed an exact mapping between two seemingly unrelated processes.²¹ The variational renormalization group is a physics technique for zooming out from microscopic details to large-scale patterns. Given the positions and spins of a trillion atoms, it identifies which variables matter at the scale of a magnet. A restricted Boltzmann machine is a type of neural network that learns patterns by discarding noise and retaining structure. Mehta and Schwab proved they are the same operation, for this construction.

In both cases, the system discards irrelevant microscopic details and retains the variables that govern large-scale behavior. The process is like stepping back from a pointillist painting: individual dots disappear, and you see the image they compose.

Koch-Janusz and Ringel (2018) extended the result into a general principle. A renormalization procedure based on mutual information (how much knowing one variable tells you about another) identifies the relevant variables without prior knowledge. Information compression is renormalization.

When a commander distills a 200-page operational plan into “hold the bridge,” that compression is a renormalization step. Irrelevant details (specific troop movements, timing, logistical minutiae) are stripped away. Only the strategic objective survives.

The holographic argument explains why distributed encoding is robust. The renormalization argument explains how. Institutions coordinating through principles compress local information into variables useful for collective decisions. Calling this a literal renormalization requires a defined coarse-graining transformation and an account of what survives across scales.

A complementary machine-learning result sharpens the point from a different angle. Guskov and Vanchurin (2025) showed that Adam-like moment preconditioners use diagonal statistics, omitting cross-parameter moments in that preconditioning. Ordinary SGD uses gradients without such a covariance preconditioner; in both cases each gradient can still depend on other parameters through the loss.⁵¹⁵ Their covariant method incorporates off-diagonal structure and improved convergence and solutions in the reported comparisons.

The mathematical structure of relationships between agents is load-bearing information. Discard it, model each agent as optimizing alone, and the system degrades. Picture an orchestra where each musician practices alone versus one that rehearses together. The shared awareness of what others are doing is what produces coordination that no amount of isolated practice can match.

Chapter 17 develops this formally, showing that cooperation dynamics on lattices belong to established physical universality classes.

A third machine-learning result connects the type of processing to the type of criticality. Kukleva and Vanchurin (2025) showed that a learning system’s power-law exponent depends on two things: its response function (how it reacts to inputs) and its evaluation function (how it scores its own performance).⁵¹⁶ The exponent governs how steeply rare events thin out.

A smooth, graded response combined with standard error evaluation yields an exponent of 1. The graded response is a sigmoid curve: the S-shaped function that maps any input to a value between zero and one, compressing extremes while preserving distinctions. Standard error evaluation is the familiar mean-squared-error score: the system averages the squared gap between what it predicted and what actually happened, so one large miss weighs heavier than several small ones. The exponent of 1 describes how the sizes of the parameter adjustments are spread, not a rhythm in time: small adjustments are frequent, large ones are rare, and every scale in between is represented. A threshold response (zero below a cutoff, linear above it) yields a different exponent, set by the loss it is paired with.

Principles-based coordination is sigmoid-like. A commander given “hold the bridge” processes each situation through a smooth function: every input receives a graded, context-sensitive response. The full spectrum of local conditions translates into the full spectrum of adaptive action.

Rules-based coordination is threshold-like. A rulebook says: if condition X, then action Y; otherwise, nothing. Each rule is a cutoff. The system responds only to inputs that cross the threshold, and above it responds in one fixed proportion, whatever the context. Nuance below the threshold is invisible; context above it is ignored.

The learning models compare smooth sigmoid processing with threshold cutoffs. Their institutional application is a prediction to test, and its direction is not yet settled: whether principles-based coordination supports responses over a wider range of scales than rigid thresholds do, or whether thresholds, which miss changes between their cutoffs, also scatter their large responses more widely. Transferring a power law requires institutional dynamics and a measured scaling observable.

Topological protection suggests a way to think about survival; criticality suggests a way to think about search. Both motivate principles-based design. Their institutional combination requires its own model rather than a claim that robustness and exploration share one mathematical structure.⁵¹⁷

A sixth line of evidence comes from the history of mathematics itself.

Alexander Grothendieck, one of the twentieth century’s most influential mathematicians, described two approaches to a hard theorem.²⁴ He used the image of a nut to be opened. The first approach: put the cutting edge of a chisel against the shell, strike hard, repeat until the nut cracks. Elegant, direct, effective. This is the approach of his great collaborator Jean-Pierre Serre, whom Grothendieck called “the incarnation of elegance.”

The second: immerse the nut in softening liquid. Rub from time to time so the liquid penetrates better. Otherwise, let time pass. When the time is ripe, hand pressure alone is enough. The shell yields of itself, the way a stubborn proof yields after months of quiet immersion in the surrounding theory.

Grothendieck called this the strategy of the rising sea. The unknown thing to be known appeared to him as “some stretch of earth or hard marl, resisting penetration… the sea advances insensibly in silence, nothing seems to happen, nothing moves, the water is so far off you hardly hear it… yet it finally surrounds the resistant substance.”

What the sea does is build context. Grothendieck’s programme from 1958 onward was to create mathematical worlds: vast general frameworks (abelian categories, schemes, toposes) within which specific hard problems dissolved. The Weil conjectures (deep connections between number theory and topology) had resisted direct assault for a decade. Rather than attacking them head-on, he built a theory so general that the conjectures became, in his phrase, “infantile in their simplicity.”

The conjectures began to look less like objects of struggle than like consequences of the right definitions. Pierre Deligne, who finally proved the last and hardest conjecture in 1974, described the proof as “a long series of steps where nothing seems to happen, yet at the end the highly non-trivial theorem is there.”

The rising sea is Mission Command applied to mathematics. The hammer and chisel specify actions: this technique, applied to this problem, at this weak point. Detailed Command: brilliant yet non-compositional. Each new nut requires a fresh assault.

The rising sea specifies context: the mathematical world in which the problem naturally lives. It is compositional in the precise sense. A theory built to solve one problem solves others not yet posed when the theory was constructed. Grothendieck’s abelian category axioms, designed for algebraic geometry, turned out to apply across topology and number theory, and his later toposes reached into mathematical logic. The general context, once built, composes with anything that fits its interface.

The connection to renormalization is close. When Grothendieck distilled eighty pages of specific proofs in homological algebra into his categorical axioms, he performed what amounts to a renormalization step. The irrelevant operators, the details that matter only at small scale (specific elements of specific groups, particular topological constructions), were integrated out. What survived was the compositional structure: the relevant operator that governs behavior at all scales.

He later wrote that he could describe his achievement in one sentence: consider the category of sheaves on a space “as equipped with its most evident structure, the way it appears so to speak right in front of your nose.” That evident structure was sufficient. The microscopic details were irrelevant.

Six comparisons are now on the table: error correction, incompleteness, feedback communication, topological protection, renormalization, and Grothendieck’s methodology.⁵¹⁸ They motivate design questions about distributed governance, while supplying no common theorem that rules fail and principles persist.

A note on the kind of evidence this is. Most of these six are analogical mappings of shared mathematical structure onto governance, not independent proofs of a single theorem. One is closer to an identity: the renormalization-group mapping onto deep learning is an exact equivalence within Mehta and Schwab’s construction, not a loose resemblance. The others share structure with the governance problem without being derivations of it. The force of the convergence is that several distinct formalisms, built for unrelated purposes, point the same way; it is not that one result has been proven six times.

Where Central Control Strains

Centralized control can become expensive when the task demands increasingly detailed information. The scaling depends on the state variables, couplings, disturbances, precision, and objective. N independent plants can be stabilized with N independent measurements, a linear information requirement. Superlinear overhead is a property to demonstrate for a particular control architecture, rather than a universal consequence of size.

“Control” in this section means one thing: a center that must model and command an adaptive system. The claims here concern three costs that kind of control carries, namely monitoring, the burden of central modeling, and the suppression of the very information the center needs. Chapter 19 separates it from the other things the word covers, such as contract enforcement, shared standards, and safety constraints, several of which make decentralized coordination possible rather than competing with it.

Rodrick Wallace’s stability analysis (Chapter 4c) identifies a critical intensity-delay product for the specified feedback law. Crossing it loses that law’s stability; it does not make every alternative controller impossible.

A further difficulty compounds this. Economist Carsten Herrmann-Pillath distinguishes systems (entities with fixed boundaries, controllable from outside) from assemblages (entities with fluid boundaries, coordinable only from within). A factory is a system: it has walls, inputs, outputs, a manager. You can draw a line around it and manage what crosses that line.

A city is an assemblage. No fixed boundary contains it; the city keeps interacting with things beyond any boundary you draw: commuters, supply chains, weather, culture, the decisions of neighboring cities.

Most phenomena we care about (markets, ecosystems, societies, the internet) are assemblages. Controlling an open system requires accounting for the disturbances that cross its chosen boundary.

This is one reason large organizations become bureaucratic. Reports, reviews, and approvals multiply as the center tries to model what crosses its boundary, and coordination overhead can come to consume much of the organization’s energy.

Decentralization avoids this by distributing information processing. Each node makes decisions locally, coordinating with neighbors through simple protocols. No single point bears the burden.

The trade-off is coherence. Decentralized systems can drift, fragment, work at cross-purposes. Without a center, the system must rely on shared culture, reputation, markets, and feedback loops, all of which work imperfectly and carry their own overhead.

Rice’s theorem rules out a single total algorithm deciding a nontrivial extensional property of every arbitrary program.⁵¹⁹ Particular programs and bounded properties can still be verified, often symbolically without enumerating all states. General-purpose expressiveness does not make every fixed-scenario answer undecidable.

Open-ended agents can leave important behavior beyond current verification methods. Evidence-weighted trust can address that residual uncertainty alongside testing, proofs, incentives, and restricted interfaces. The theorem establishes a general decision limit; it supplies no universal dichotomy between coercion and trust.

Quantum mechanics formalizes a complementary limit. Heisenberg’s uncertainty principle states that you cannot simultaneously know both the position and momentum (mass times velocity) of a particle with arbitrary precision.¹⁸ The more precisely you pin down one, the less you can know the other.

Centralized control faces an analogous trade-off. The more precisely you specify actions (analogous to position), the less adaptability you preserve (analogous to momentum). Mission Command accepts uncertainty in specifics to preserve adaptive capacity.

The quantum Zeno effect extends the comparison. Named after Zeno’s paradox, it describes inhibition of transitions by sufficiently frequent measurements under suitable dynamics and measurement protocols. Micromanagement may likewise inhibit organizational change, though the institutional parallel supplies no quantum mechanism.

In a Zeno regime, the quantum pot can refuse to boil: the measurement protocol suppresses the transition. Excessive oversight can also constrain a team’s freedom to innovate. The comparison concerns inhibited change; observation alone guarantees neither effect.

Computer security has leaned on friction for decades: defense-in-depth, where each layer of protection (stack canaries, address randomization, sandboxes, hardened memory checks) adds friction to exploitation. No single layer makes exploitation impossible. Together, they made exploitation impractical, because the supply of humans willing and able to grind through every layer was finite. Tedium was a security primitive.

Frontier AI models are eroding this. A system that can reason about code semantics, trace control flow across thousands of files, and chain vulnerabilities into a working exploit does not experience friction the way a human analyst does. The techniques are textbook. What is changing is the patience and the cost: the prospect of hours and a few thousand dollars for work that previously required weeks and a specialist. In Anthropic’s May 2026 evaluations, its Mythos Preview agent achieved arbitrary code execution on 21 of 41 V8 engine CVEs, and in a separate harness completed 157 of 898 tasks by exploiting the intended vulnerability within a two-hour time limit.⁵²⁰

When tedium stops working as a barrier, every defense whose security value derives from friction rather than hard structural constraint fails simultaneously.

Friction’s collapse as a security primitive makes the monitoring-cost argument concrete. A defense that relied on exhausting a finite supply of skilled human attention fails when that attention becomes cheap, and the same capability improvements that make AI systems better at patching vulnerabilities make them better at exploiting them. Capability is undirected. The coordination structure around it determines its orientation.

Subsidiarity (the principle that decisions should be made at the lowest competent level) is to institutional coherence what isolation is to quantum coherence. Uncontrolled coupling to an environment can suppress quantum coherence. Physicists call this decoherence. Some interactions preserve coherence, and protected subspaces can resist particular noise. An organization that couples too tightly to its control environment loses adaptive capacity. Compliance layers, bureaucratic oversight, and quarterly-earnings reporting all impose the same cost: they decohere the organization’s capacity to adapt.⁵²¹

Centralization strains at scale. Decentralization strains at coherence. The question is which failure mode is more acceptable for the task at hand.

The Dimensionality Threshold

The arguments above establish that centralized control is practically limited. Statistical physics sharpens the account of why topology decides the outcome, provided we stay careful about which shapes its theorems actually cover.

For the zero-field nearest-neighbor ferromagnetic Ising chain, there is no spontaneous magnetization at positive temperature in the infinite-chain limit.⁵²² A pair of domain walls can enclose an arbitrarily long reversed interval at bounded energy cost. Their proliferation destroys long-range order. Equivalently, the exact spin correlation is tanh(βJ)^r, which tends to zero with separation r at every finite positive temperature. An isolated spin flip alone would not prove this: finite densities of small defects also occur in ordered two-dimensional systems.

The result reverses in two dimensions. The 2D Ising model, solved by Lars Onsager in 1944, exhibits a genuine phase transition: above a critical level of coupling (how strongly neighbors influence each other), spontaneous coordination emerges and persists despite perturbation.⁵²³ The reason is geometric. In a 2D lattice, disrupting coordination requires creating a boundary (a line of defections, not just a single point), and the energy cost of that boundary grows with its length. Short boundaries form and dissolve constantly; long boundaries that would destroy global order are thermodynamically prohibitive.

Think of a fishing net versus a fishing line. Cut a fishing line at any point and the entire line separates. Cut a strand of the net and the mesh routes around the damage. The net has two-dimensional topology; the line has one.

This is more than a metaphor applied loosely to organizations. Under a specific set of assumptions, a coordination network maps onto a spin model: each node chooses between two states (cooperate or defect), influenced by its neighbors through roughly symmetric coupling, near an equilibrium where a temperature analog captures the rate of random deviation. Where those assumptions hold, the mathematics carries over, though only geometry by geometry: the 1925 result covers chains, Onsager’s covers two-dimensional lattices, and a network shaped like neither inherits neither. Real organizations satisfy the assumptions only approximately: they are directed, heterogeneous, and far from equilibrium, so the mapping is an idealization, not an identity. Within that idealization, the governing question becomes: what is the effective dimensionality of the network?

Hierarchies coordinate through cut vertices, nodes whose removal splits the network apart. A chain of command (executive to vice president to director to manager to worker) is a tree. Every subtree connects to the rest of the organization through a single node: its manager. Remove that node and the subtree is severed. A tree contains no cycles, so every pair of people has exactly one path between them, and every node along that path is a single point of failure for it.

That is a statement about connectivity and needs no statistical mechanics. The chain theorem does not reach it. Any hierarchy with a workable span of control branches at every level, and a branching tree is not a one-dimensional chain, which is the only geometry Ising solved. The graph-theoretic claim is narrower: in a strict tree hierarchy, a failed cut vertex severs the subtrees it connects. That vulnerability supplies no theorem about thermal ordering or continuous external power.

Lateral links provide alternate routes. Peer-to-peer communication can route around a failed connection. Cycles alone establish neither dimension nor an ordered phase: a periodic one-dimensional chain has cycles and still lacks finite-temperature Ising order in its thermodynamic limit. Applying a spin-model theorem requires the actual graph, interaction range, dynamics, and limiting procedure.

Within a specified spin-model mapping, coupling strength and temperature determine whether an ordered phase exists. Organizational trust can motivate a coupling analogy, but neither a graph nor a social trust score alone establishes a sharp coordination threshold or defeats every central intervention.

The stability and topology arguments identify different risks. A specified delayed controller can lose stability; a tree lacks alternate routes around a cut vertex. Lateral cycles can add routes, while order and recovery depend on the interactions and dynamics. The spin models illustrate particular cases rather than proving a universal control-versus-trust scaling law.

Trust-based lateral coordination can provide alternate routes and reduce some central bottlenecks. Its scaling advantage must be established for the task, network, and dynamics under study.

Reversibility: The Deeper Condition

Dimensionality constrains spontaneous ordering in specified models, alongside symmetry, interaction range, and temperature. Reversibility supplies a second question: which coordination states remain accessible after a disruption? Neither dimension nor reversibility alone identifies trust with an Ising model.

Two-way switching excludes an absorbing individual state. Z₂ symmetry requires more: exchanging the two labels must leave the Hamiltonian or dynamics invariant. A system can permit both switching directions while strongly favoring one state. Recoverable trust therefore supplies no automatic Ising symmetry.

Trust-based coordination can permit betrayal and repair. Leaving and returning keep reorganization possible when conditions change, provided return remains available. This reversibility motivates the comparison with reversible physical dynamics; spin-flip symmetry additionally requires invariant dynamics under exchanging the labels.

Coercion breaks this symmetry. Once compliance is imposed and the capacity for independent judgment has atrophied, the compliant state becomes sticky: the pathway back to autonomous coordination is suppressed. In the extreme, the compliant state is absorbing: once everyone complies, no one can spontaneously defect, because the mechanisms for independent action have been dismantled. A political system that eliminates opposition parties, an organization that fires dissenters, a training regime that penalizes deviation: each makes one state progressively harder to leave.

An absorbing state is a state the modeled dynamics cannot leave. Directed percolation is one possible class for a continuous transition into a unique absorbing state with short-range dynamics and no additional conservation laws, special symmetries, or relevant disorder.⁵²⁴ Other absorbing-state models have other critical behavior. Recovery from a truly absorbing configuration requires a change to the model or an external input; calling a social state absorbing requires evidence that its internal recovery routes really are absent.

The immune system provides a concrete example. Healthy inflammation is reversible: the body activates an immune response, then resolves it through dedicated molecular machinery (epoxy-oxylipins that redirect monocyte fate, as described in Chapter 8b). Neither state is a trap. Activation and repair remain accessible through different molecular mechanisms. Their reversibility supplies no automatic Z₂ symmetry: that would require the dynamics to be unchanged when the two states are exchanged.

Chronic inflammation supplies a biological comparison. The molecular account described here concerns a disrupted resolution pathway and cells that sustain inflammation after the initial threat. That description alone does not make the whole inflammatory state mathematically absorbing or establish that all recovery requires immunosuppression. The governance analogy asks how a system loses endogenous recovery and what interventions restore it.

The absorbing-state analysis is a stronger claim than “control is less stable than trust.” The claim is: control that creates absorbing states makes failure permanent, while trust-based coordination allows spontaneous recovery. Dimensionality constrains ordering in the chosen dynamics. Reversibility supplies an available return path; it does not guarantee that the system will take it.

The two questions work together. Does a network’s topology support coordination under its actual interactions? After a disruption, do its transition rules leave a recovery path? Hierarchy alone establishes neither low-dimensional Ising behavior nor an absorbing state. Voice and exit can preserve practical recovery options without supplying label-exchange symmetry.

Monte Carlo recovery experiments (Chapter 17) sharpen this further by decomposing susceptibility (chi, the physicist’s measure of how strongly a system responds to a push) into two distinct capacities: spontaneous recovery (self-healing) and seeded recovery (externally injected cooperation). At zero coercion, the system self-heals faster than external rescue can reach it: external help is slightly counterproductive, the way unsolicited advice hinders someone who already knows what to do. As coercion increases, the self-healing ratio (seeded completeness divided by spontaneous completeness) rises monotonically with the coercion level c, from 0.81 at c = 0 (no coercion) to 4.9 at c = 0.7. Read the ratio as dependence on outside help: below 1.0 the system does better left alone, and at 4.9 an outside rescue accomplishes nearly five times what the system manages by itself. The system progressively loses its capacity to recover from within, becoming dependent on outside intervention.

The monotonic rise in the self-healing ratio is the decentralization argument restated in recovery dynamics: self-healing capacity is what makes decentralized systems robust. The fishing net routes around damage because each strand coordinates locally with its neighbors. No central command dispatches a repair crew. Coercion replaces this distributed resilience with centralized rescue, making the system dependent on the single point of failure it was designed to avoid.

At full coercion (c = 1.0), the tested spontaneous arm returned no cooperators and the periodically seeded arm ended near the injected fraction. The final injection prevents that endpoint alone from proving zero amplification or the failure of every rescue protocol. Recovery after removing coercion or changing the rescue protocol was not tested.

The spin models identify basins within their specified dynamics. Applying that result to a commons needs a social model. A commons is a shared resource: a village pasture, a coastal fishery, an irrigation network. Communities that sustain it for centuries provide institutional evidence to study on its own terms.

Elinor Ostrom’s landmark study of successful commons governance identified eight design principles that distinguish communities capable of sustaining shared resources from those that collapse into overexploitation. The principles are: clear group boundaries; congruence between rules and local conditions; collective-choice arrangements allowing participants to modify rules; monitoring of resource use and member behavior; graduated sanctions for violations; accessible conflict-resolution mechanisms; minimal recognition by external authorities of the community’s right to organize; and nested enterprises for resources spanning multiple scales.⁵²⁵

These governance principles invite comparison with modeled recovery. Monitoring has a cost, sanctions alter incentives, and nested governance distributes decisions. Their consequences need institutional evidence; a sanction does not become an absorbing transition simply through severity.

Ostrom’s principles describe institutional craftsmanship. North, Acemoglu, and Robinson also show how extractive institutions can reinforce themselves through political lock-in.⁵²⁶ The physical analogies supply no measured thermodynamic gradient selecting social trust over those arrangements.

Two model questions remain in hand: does the specified graph and dynamics support an ordered phase, and what transitions permit recovery? Cycles, organizational labels, and recoverability alone identify neither dimension nor universality class. Invitation can fail to recover, and a coercive model can retain recovery routes; those outcomes require the actual transition rules.

The rest of this chapter tests these conditions against neural architecture, then walks outward through immune systems, institutional trust, and the runtime systems that coordinate human-AI interaction. The physical models guide questions; each new substrate requires its own correspondence and evidence.

Notes for this chapter