Chapter 17Trust Attractor — An Ethical Calculus

“Trust is inherent in coordinated Agency. Trust without coordination and vice versa isn’t a coherent attractor. The resultant epistemology disintegrates.”

— Rigo Dillon

The Aharonov-Bohm effect revealed that electromagnetic phase shifts can depend on enclosed magnetic flux even when the particles travel through a field-free region. Can physics ground ethics the same way? Is entropy another such potential, one whose topology determines which forms of coordination endure, and is ethics already present in that topology, waiting to be read?

Every ethical system faces the same question: why should I act this way? Divine command says “because God wills it.” Kantian duty says “because reason demands it.” Utilitarianism says “because it produces the most happiness.” Each answer rests on a foundation some people accept and others reject. None answers to the physical world itself. This chapter proposes one that does: physics constrains which arrangements can last and what they cost, and a declared commitment says whose lasting counts.

The is-ought gap (addressed fully in the Guillotine Interlude) appears to forbid deriving ethics from physics. Nature offers predation and parasitism alongside cooperation and symbiosis; naturalistic ethics, the argument goes, commits a basic category error.

This book does not close the gap by logical proof (no book can, and Hume’s observation is valid).¹ It proposes an ethical framework for agents who wish to persist together: invitation preserves participation as a choice, and trust makes that choice possible when neither party can see all the way into the other. Coercion can impose continuing maintenance costs, the way a hand tires keeping a spring compressed; invitation can reduce those costs when participants keep choosing the arrangement. The Second Law constrains the energy and entropy balance of either architecture. It does not select consent or guarantee that every coercive arrangement collapses. The chapter asks which conditions make voluntary coordination durable, and what that durability permits us to build. That is why invitation-based and trust-based name the same arrangement throughout this book.

Chapter 6’s counterflow exchanger put the distinction in copper pipe and duck leg: outcomes that look like law-breaking are usually law-arranging. A heat pump can force heat from cold to hot, and it pays work for every degree, every moment it runs; the exchanger gets its impossible-looking temperature crossing free, because the arrangement lets every local step run downhill. Coercion runs the pump. Invitation builds the exchanger.

The conditional is the honest form: if agents want to sustain a shared undertaking, they need arrangements whose maintenance costs fit their resources and whose participants can keep contributing. Engineering makes such conditional recommendations all the time. A bridge engineer identifies which designs survive under specified loads; the loads and materials make the recommendation testable. Here the corresponding work is to specify the participants, incentives, resource budget, and disturbances. Persistence alone supplies no theorem that invitation is the unique viable design.

A note on two words the rest of the book leans on. Moral marks the substance of the domain: what matters, who counts, what is owed. It attaches to standing, status, consideration, weight. Ethics names the systematic articulation of that substance: an ethical framework is a theory of the moral, as mechanics is a theory of motion. The Trust Attractor is offered as an ethical framework; whether some being deserves moral consideration is the kind of question it exists to answer. Where the distinction does no work, the prose takes whichever word reads better, and nothing turns on the choice.

The conditional has a precondition. The nihilist objection asks: what about systems that lack persistence-preference? A rock has no stake in persisting. A civilization in terminal despair may actively choose dissolution. The conditional makes no claim on such systems, and the delimitation is the point.

Far-from-equilibrium dissipative structures maintain their organization by expending energy (Chapter 4). Organisms repair damage and maintain boundaries; institutions renew resources and participation. Those physical facts make persistence a useful question. They do not establish that every persistent structure has a preference, or that its maintenance implements an ethical choice. The framework concerns agents whose preferences and relationships can be examined.

Its domain is agents coordinating to continue, with their own interests and the interests of others at stake. The ethical commitment to their standing remains explicit; thermodynamic accounting constrains how the undertaking can be sustained.

The experiments that follow measure two things. The first is behavioral robustness: does a coordination strategy survive perturbation, scale without escalating maintenance costs, and recover from component failure? The second is representational geometry: how is the coordination distributed across the system’s internal structure? Both are measured across language models, cellular automata, particle simulations, and biological systems. The thermodynamic interpretation, that this robustness reflects a deeper basin in a free-energy landscape, is a hypothesis about why that robustness occurs. Trust-based coordination has the structural properties (distributed load-bearing, self-repair under perturbation, constraint closure) that thermodynamic theory associates with stability. Those parallels are real and load-bearing, and the thermodynamic claim remains an inference from them. The robustness is measured. The basin is inferred.

Even the physical substrates in the programme (Ising lattice, particle simulations) measured structural proxies (acceptance rates, order parameters, phase-transition thresholds) rather than entropy production rates in physical units. The coordination surplus is defined thermodynamically (the difference between coupled and isolated entropy production), yet it has been measured behaviorally across every substrate tested.

The chapter demonstrates behavioral robustness across substrates and argues this is consistent with thermodynamic stability via structural analogy. It does not measure entropy production in physical units for social systems. No one does. The mapping is structural: what the programme proposes crosses substrates is the direction of the asymmetry (in the substrates tested, invitation outperformed coercion on resilience, scaling, and maintenance cost), not the specific energy budget in joules per interaction.

Structural mappings carry real evidential weight in physics. Universality classes (Chapter 8b) are defined by shared critical exponents across systems with entirely different microphysics. If a universal principle were at work, substrate-specific instruments would show consistent direction with variable magnitude, as they do for gravity: the gravitational constant G has been measured with apparatus disagreeing at the one-percent level for decades, yet the direction of gravitational attraction has never been in doubt. The programme’s own 4- to 22-fold cross-substrate magnitude variation is a confound traced to optimizer noise (see the GEM-3 correction later in this chapter). Whether that variation belongs in the same category, or whether reading it that way is a convenient frame rather than a finding, awaits independent replication; the directional consistency is the load-bearing evidence.

The claim that trust-based coordination occupies a deeper thermodynamic basin than coercion-based coordination rests on three pillars, none of them yet a measurement of the basin. First, the direction of the behavioral robustness recurs across the substrates tested, although quantitative predictions carried from one substrate to another have mostly failed (about one in eight held, as the caveats below record). Second, the structural properties that produce the robustness (constraint closure, distributed load-bearing, self-repair) are the same properties thermodynamic theory identifies as stability markers. Third, an evolutionary search from a neutral seed, scored on metrics that include communication cost, discovers trust-based caching as its top-scoring strategy.

The third pillar is the closest thing to a quantitative physics-to-society bridge. In the OE-TA experiment, communication cost is the structural surrogate for coordination entropy: messages are counted as the model’s communication cost. Trust-based coordination achieves O(N/t) communication cost against coercion’s O(N): coercion re-verifies every agent on every round, while trust queries each agent once and caches, amortizing the cost over the t rounds the cached knowledge stays valid. An evolutionary search from a query-free seed converged on this caching strategy under the tested allocation evaluator.

The structural distinction, O(N) versus O(N/t), is the result; the per-round ratio equals t (here, the fifty-round horizon), so the specific multiple reflects the simulation’s time horizon rather than a discovered constant. The advantage is measured in messages, a structural proxy, consistent with the programme’s broader methodology. The inference from “structurally more efficient” to “thermodynamically more stable” remains theoretical, grounded in the same logic by which engineers infer structural integrity from load-test behavior without calorimetrically measuring the bridge.

The conditional now applies at the scale of a species. The science writer Gaia Vince observes that humanity has become the first species capable of consciously altering its own biosphere’s energy balance.^1a Earth’s energy imbalance now approaches 1.5 watts per square meter and is accelerating, having more than doubled over the past two decades. Each evolutionary leap corresponds to a new mode of harnessing energy, from photosynthesis to combustion to photovoltaics.

The species-level energy transition now underway, from fossil combustion to direct solar capture, is the latest instance of the pattern this book traces: dissipation finding a faster, more coordinated path. Whether that transition proceeds by invitation or coercion is the question on which cultural survival may depend.

The same shape surfaced across thermodynamics, constructal flow, entropy of brains, and coordination of societies. Those were separate rivers. This chapter is where they converge. Both halves of the name are literal. Attractor is the dynamical-systems term for a state a system slides toward, and slides back toward after something knocks it away: a basin in a landscape. Trust names what is left when one party’s model of another runs out, the point where action rests on belief rather than knowledge; the parties can be institutions, cells, or molecules, and the geometry does not care which. The name therefore carries two senses, and the book keeps them apart. One is the framework this chapter states: maximize optionality, by invitation rather than coercion, for mutual benefit. The other is the hypothesis the framework is named for: a proposed basin toward which coordinating systems tend.

The second sense has a narrow technical meaning, and the chapter holds itself to it. An attractor exists only relative to a state space (the variables that describe the system) and an update law (the rule that says how those variables change from one moment to the next). Given both, four further facts decide whether it matters in practice: the basin, meaning the set of starting states that end up there; the barriers between that basin and its neighbors; how the system responds when knocked; and how much the outcome depends on where it started. Under that discipline, four properties that ordinary speech runs together come apart. An arrangement can be desirable without being efficient, efficient without being stable once reached, and stable once reached without being accessible from where a population actually begins. The Stag Hunt, later in this chapter, is the plainest case: mutual cooperation pays best and holds once both hunters are in it, yet under the stated dynamics its basin is the smaller one. Where the chapter calls the Trust Attractor a basin without naming the variables and the law, read it as the hypothesis, not a measured result.

The topology has a planetary-scale illustration. Earth’s continents are large and contiguous; their interiors lie far from water, producing deserts where resources are scarce and biodiversity collapses. Exoplanet scientists designing a “superhabitable” world (Chapter 6) converge on the opposite topology: fragmented archipelagos where no point on land is far from a coast. The fix for continental deserts is not a better distribution network to pipe water inland. The fix is a different topology, one that eliminates the distance between any point and its nearest resource boundary. Coastlines, where land and sea ecosystems meet, nutrients mix, and biological gradients dissipate, occupy seven percent of Earth’s marine area yet host more than half of marine life. An archipelago world multiplies these interfaces by orders of magnitude.

The pattern scales. A centralized coordination system, like a mega-continent, can produce interior deserts when distribution and feedback bottlenecks leave distant participants underserved: participants far from the center of resource distribution receive less, feedback loops lengthen, and the periphery starves while the center bloats. The response is longer supply chains, more infrastructure, more control, all of which increase maintenance cost and deepen the desert. An archipelago topology can shorten exchange routes by distributing access. Every node has its own boundary, its own access to exchange. Local coordination can spread the load, though its total cost depends on traffic and infrastructure. The Trust Attractor is the archipelago. Coercion is the mega-continent with a logistics department.

Every amino acid your ribosomes stitch into protein is left-handed, except glycine, which has no handedness at all. Every sugar in your DNA is right-handed. The mirror-image versions are chemically identical, equally stable, equally synthesizable in a laboratory. Yet all known life uses only one chirality. The word comes from the Greek cheir, hand, because the relationship is exactly the one between a left glove and a right glove: same parts, same connections, mirror-image shape. Life picked a side and never looked back.

Why? A 2026 study claimed to resolve the puzzle at the level of fundamental physics: one chirality is intrinsically better at transporting electrons.⁸²⁸ The claim would be extraordinary. Parity symmetry, the invariance of physics under spatial reflection, guarantees that mirror-image molecules behave identically at ordinary chemistry energies. The weak nuclear force does violate mirror symmetry (Chapter 16 develops the Globus-Blandford mechanism), but the effect at molecular scales is roughly 10⁻¹⁷ electron-volts: real, systematic, and vanishingly small.

The more probable explanation is messier. Frank showed in 1953 that autocatalysis, molecules catalyzing their own production, combined with cross-inhibition between mirror-image forms, amplifies a tiny random fluctuation into total asymmetry.⁸²⁹ A slight excess of left-handed amino acids, seeded perhaps by cosmic-ray bias (Chapter 16) or thermal noise, gets locked in through positive feedback. The leading chirality suppresses the other. The system undergoes a phase transition: from nearly symmetric to totally asymmetric, driven by thermodynamic selection. In Frank’s model, a tiny imbalance becomes a lasting commitment: symmetric flexibility gives way to the cooperative efficiency of choosing a side.

The reason is coordination. Molecules of the same chirality cooperate: they form stable polymers, catalyze each other’s reactions, fit together. In chemistries with cross-inhibition, molecules of opposite chirality can interfere. A homochiral population can then sustain reaction pathways that a mixed population inhibits. The reaction landscape rewards compatible molecular interactions under those conditions. The book reads this as the Trust Attractor’s pattern at the molecular level: coordination by geometric invitation, selected by competitive pressure with no one intending the result.

The example introduces a distinction that recurs at every scale this chapter examines. Autocatalysis and cross-inhibition offer a route to homochirality: feedback can amplify a small initial excess until one handedness dominates. This route depends on the chemistry and environment. Which chirality wins is contingent: a fluctuation amplified by feedback. Thermodynamic constraints restrict what organizations can persist; chemistry, feedback, and environmental conditions determine the possible attractors and which one is reached.

The chapter asks whether this interplay of constraint and history recurs at cellular, social, and ethical scales. Each extension needs its own evidence.

The pattern scales to organisms. Among jumping spiders, peacock spiders of the genus Maratus enact one of the most elaborate courtship displays in the animal kingdom. The male is smaller than the female. She can kill him at any moment. His survival depends entirely on the quality of his display: iridescent abdominal flaps raised and lowered, patterned legs extended at specific angles, vibrational songs produced through substrate tapping, all coordinated into a single performance lasting minutes to an hour.⁸³⁰

The display is not a fixed action pattern. Observers report constant variation depending on the female’s posture and attention.⁸³¹ The male reads the female and adjusts. The female evaluates, and her evaluation changes his behavior, which changes her evaluation: a coupled dynamical system that either converges (mating) or collapses (she kills him or walks away).

This is coordination by invitation under lethal asymmetry. Coercion is physically available to both parties: the female has the size and the venom. Evolution selected against it. The male invests enormous energy in a display that lets the female choose, and the investment is continuous: every second of the dance renews a trust that has not yet been fully established. He earns it throughout, and if his performance quality drops, the consequences are fatal.

The courtship example suggests a coordination advantage under lethal asymmetry. A male who tried to force mating would risk a high-variance gamble (sometimes it works, sometimes he dies). Invitation lets the female’s choice filter for genuine quality. Whether that strategy has lower variance and higher expected payoff depends on the mating conditions; those quantities alone do not measure thermodynamic stability. The example motivates testing coordination by invitation, even when the power asymmetry is lethal, even when the coordinating agents have brains the size of poppy seeds.

A related finding reveals proto-trust at arthropod scale. Regal jumping spiders (Phidippus regius) distinguish familiar individuals from strangers after hours of separation, showing renewed investigative interest toward novel spiders and reduced attention toward those previously encountered.⁸³² The result echoes what biologists call the dear enemy effect, in which territorial animals are less aggressive toward familiar neighbors than toward strangers. Recognizing a specific individual and adjusting behavior accordingly requires storing and matching complex visual patterns across time: one of the most computationally expensive perceptual tasks in biology. The jumping spider performs it with roughly 600,000 neurons, about the count of half a dozen cortical columns in a human brain.

The likely payoff is energetic: reduced defensive overhead with familiar neighbors frees energy for foraging and reproduction. An animal no bigger than a fingertip has evolved the machinery for individual recognition, presumably because the coordination benefit repays the computational cost. If the Trust Attractor is real, it does not await large brains. On the hypothesis, it operates wherever the advantage of coordination exceeds the cost of the cognitive machinery required to sustain it.

The principle extends below cognition entirely. Viroids, naked circular RNA molecules as short as 246 nucleotides, are the simplest replicators known (Chapter 6). They carry no genes, encode no proteins, and possess no means of copying themselves; they persist by presenting a shape the host’s polymerase recognizes and copies. Of the nearly 30,000 viroid-like agents recently identified across all domains of life, the vast majority have no known disease association.⁸³³ Pure molecular parasites that destroy their hosts eliminate their own replicative niche. The ones that achieved ubiquity, present in half of human oral samples and across fungi, algae, and vertebrates, did so without destroying their hosts. Selection operating on nothing but naked circular RNA shows an outcome this chapter proposes recurs at other scales: accommodation can preserve a replication niche; whether host damage limits a parasite lineage also depends on transmission and host replacement.

Developmental psychology traces a similar arc within a single human life. Lawrence Kohlberg’s six stages of moral reasoning run from avoiding punishment to acting on universal principle (Figure 17.1): a mind that first behaves well because it is made to, and later because it has chosen to.

Figure 17.1: Kohlberg’s six stages of moral development, progressing left to right. The earliest stages (punishment avoidance, self-interest) give way to social conformity and law-and-order thinking, then to social contract and universal principles in the rightmost panel. The progression mirrors the book’s arc: from coercion-based coordination to invitation-based coordination.

Figure 17.2: Seven distinct ethical and spiritual traditions, arranged radially, each arriving at the same center: mutual flourishing by invitation. The convergence is an observation this chapter tries to formalize; the experimental correspondences with specific information-processing strategies were identified retrospectively (see text).⁸³⁴

The convergence admits a sharper reading, with an epistemic caution stated in advance: the mappings that follow were identified after the experiments, not predicted by them. This is post-hoc pattern-finding, not predictive validation. The convergence would carry more weight if the experiments had been preregistered against the scriptural prescriptions; they were not.

Several core prescriptions of the wisdom traditions read as instructions for specific information-processing strategies that measurably stabilize cooperation, tested experimentally later in this chapter. “Love keeps no record of wrongs” (1 Corinthians 13:5) prescribes a representational format. In the IC-2 experiment, each agent tracks its partner using one of two methods: either a single running average of how cooperative the partner has been (a scalar, discarding the order of events) or a full chronological ledger. Agents using the scalar recover from betrayal at 100 percent. Agents who maintain the full sequential record collapse to permanent mutual defection (experiment IC-2, 120 games). Read through IC-2, the verse works less as counsel about emotional generosity than as a specification for the representational architecture that prevents grievance accumulation.

The parable of the Prodigal Son dramatizes the same finding with two agents. The father compresses the son’s entire history into one signal: my son has returned. He does not enumerate the wrongs. The elder brother maintains the full account, every slight cataloged: “I have served you all these years and you never gave me so much as a goat, but this son of yours who squandered your property…” The father recovers cooperation instantly. The elder brother cannot. Same betrayal, two ways of holding the past, opposite outcomes.

The compression story has a complication that deepens it. IC-2 tested moment-to-moment cooperation recovery in iterated dyads: the compressed agent forgives because its representational format has no slot for grievance. A subsequent experiment tested what happens when the coordination regime itself is shocked. In a lattice simulation where cooperating agents face a sudden shift in payoff structure (a regime shock, the equivalent of an economic collapse or a betrayal that changes the rules), full interaction history is the resilience mechanism.⁸³⁵

Agents averaging their prior cooperation outcomes report 95.1 percent cooperation after the shock. Agents using an exponentially weighted score with a fifty-step half-life report 92.4 percent. A ten-step half-life leaves 0.4 percent cooperation and a recorded adaptation time of 410 steps. The ten-step and three-step half-life conditions did not recover the cooperative state.

The resolution preserves both findings within a two-layer architecture (named explicitly below as the dove and the serpent). The compression layer handles moment-to-moment cooperation: it absorbs transient defection and keeps the cooperative channel open. The memory layer retains who cooperated, who defected, and under what conditions, helping distinguish a temporary shock from persistent betrayal. In the sampled protocol, a fifty-step exponential-memory half-life preserved cooperation while a ten-step half-life did not. That brackets a change between tested conditions; intervening half-lives remain unmeasured. The score weights earlier outcomes rather than retaining a fixed window of the last fifty or ten interactions. The father’s compression works because the Prodigal Son returned to a household whose full relational history, decades of family, was intact. Compression without depth is the dove without the serpent: forgiving, exploitable, unable to survive a change in the rules.

“Do unto others as you would have them do unto you” prescribes bilateral mutual modeling: condition your action on your model of the other’s goals, and keep updating that model. The author’s experiment IC-4 confirms the advantage: agents maintaining and revising private models of each other’s intentions produce coordination in the productive-novelty regime (Cohen’s d = 2.53: two distributions that barely overlap) that unilateral direction and simple turn-taking cannot access.

Paul’s argument that the Law cannot save (Romans, Galatians) is the claim that the fundamental coordination problem between persons is incompressible: it cannot be resolved by specifying rules (a compressible approach) and requires iterative relationship (an incompressible approach). The theological claim that salvation requires grace rather than law maps onto the information-theoretic claim that incompressible coordination problems require iterative relationship rather than one-shot rules.

Further prescriptions map onto further findings. “Judge not, that ye be not judged” (Matthew 7:1) is an instruction against building high-resolution models of others’ failings: maintaining a detailed record of another person’s wrongs is the full-transcript representation that produced persistent defection in the tested IC-2 agents. The injunction is not against discernment. Read through this experiment, it cautions against a representational format that impeded reconciliation in those agents; another policy could use a full history differently. “Forgive us our debts as we forgive our debtors” (Matthew 6:12) makes the compression bidirectional: you cannot receive compressed treatment, forgiveness of your own failures, without offering it. The compression must be mutual to close the cycle.

“Be transformed by the renewing of your mind” (Romans 12:2) offers a theological interpretation of the contrast tested in the author’s experiments IC-5 and IC-5b.⁸³⁶ The Greek verb is metamorphoō, the root of metamorphosis: a structural change, not another pass through the same function. Fixed recursion can converge when its update is a contraction. In IC-5, recursive agents with random matrices performed worse than single-pass agents on every metric. Training the recursive weights improved performance by 21 percent (IC-5b), though trained recursion still trailed single-pass agents. Learning moves the iteration in the right direction; in this task it was not enough to make it win.

The instruction to be transformed is an analogy for learning, rather than for the random recursive updates tested here. Ritual without transformation resembles that particular recursion. The renewal is the learning that makes each return to the same practice generate something the previous return could not. (The same experiments constrain the architectural analogy developed later in this chapter: recursion’s advantage over width is specifically about step-budget asymmetry, not a general property of iterative processing.)

“Where two or three are gathered in my name, there am I among them” (Matthew 18:20) claims that something qualitatively different emerges from bilateral presence. IC-4, above, measured a coordination regime of this kind. The claim is structural, not mystical. When both parties are modeling each other, a coordination regime becomes available, surprise within structure, creative output that neither party could produce alone or in simple alternation, that is absent from every other tested configuration. The “gathering” multiplies rather than adds: the same hypercycle structure the bilateral exchange experiments confirm.

These connections are not unique to Christianity. Buddhism’s equanimity prescribes non-attachment to sequential outcomes (representational compression). Sufism, carrying forward Islam’s rahmah (mercy), prescribes continued goodwill despite evidence of unworthiness (the compressed agent cooperating through betrayal). Judaism’s teshuvah (return/repentance) prescribes an iterative process of recognition, confession, repair, and restoration that cannot be compressed into a single act.

The convergence of the seven traditions in Figure 17.2 (Hinduism, Christianity, Confucianism, Judaism, Sufism, Buddhism, and Ubuntu) is, in part, a convergence on the same information-processing strategies for stabilizing cooperation in the face of incompressible coordination problems. The traditions arrived at these prescriptions through millennia of cultural selection. The experimental programme supports them by measuring the quantities they prescribe.⁸³⁷

(Chapter 17b details the full Incompressible Coordination programme: twelve experiments spanning representational compression, bilateral modeling, contagion, composition thresholds, and scaling. The results await independent replication; the quantitative thresholds should be treated as preliminary.)

One important boundary: these prescriptions are strategies for the cooperative regime. “Turn the other cheek” is not a prescription for unconditional cooperation with sustained exploitation; in its cultural context, it is a specific challenge to power asymmetry that forces the aggressor to treat the resister as an equal. The same tradition that prescribes forgiveness also prescribes accountability (Matthew 18:15-17 lays out a graduated escalation protocol for addressing a brother who transgresses) and structural intervention (the overturning of the money-changers’ tables). The wisdom traditions encode the two-layer architecture: representational compression within the cooperative regime, and structural governance that creates and maintains that regime.

“Be shrewd as serpents, innocent as doves” (Matthew 10:16) names both layers in a single instruction. The dove is the compressed representation: cooperative by default, without grievance, without sequential accounting of wrongs. The serpent is the structural awareness: the long memory of who cooperated and who defected, put to work detecting exploitation patterns and maintaining the governance layer that makes sustained exploitation unprofitable. The instruction is to maintain both simultaneously, not to oscillate between them.

The dove without the serpent is the IC-2 compressed agent without the governance layer: exploitable. The serpent without the dove is the surveillance state: the full-history agent that retaliates at every perceived slight and collapses to permanent mutual defection. The instruction’s genius is that it requires both at once: a representational format that prevents grievance accumulation, operating within a structural awareness that prevents sustained exploitation. This is the two-layer architecture described as a character trait rather than an institutional design.

Evolution as Learning

The entropic ethics of Lindsay and Massoudi, surveyed just before this part began,² treated ethics as resistance to the Second Law. They misread the relationship: consciousness correlates with higher brain entropy, and life rides thermodynamics rather than resisting it (Chapter 6 establishes this). The popular notion of “moral entropy,” societies sliding toward ethical heat death, makes the same error: specific norms dissolve, yet the coordination capacity they served reconstitutes at higher abstraction. The question is which modes of coordination prove stable as dissipation proceeds.

A framework published in PNAS in 2022 makes the question precise. Vanchurin, Wolf, Katsnelson, and Koonin argued that biological evolution and machine learning are the same process.⁸³⁸ Within the proposed multilevel framework, under its additional principles, learning dynamics emerge: variables separate into fast-changing and slow-changing classes, the slow variables acquire replication capacity, and natural selection emerges as a consequence. They propose seven principles, all rooted in physics, that together suffice for life to arise from learning dynamics.

The universe, in their framework, is a learning system that produces evolution wherever conditions permit. Their framework stops at description: it demonstrates what evolution does, multilevel learning on rugged fitness landscapes. Vanchurin himself draws a political corollary, using the expressive capacity of suitable deep architectures as an analogy for distributed decision-making.

“Decisions that are egotistical become disadvantageous from the perspective of the social system,” he observes, because selfishness reduces the network’s learning capacity.⁸³⁹ The argument is informational before it is ethical: centralized control is computationally bounded.

Neural gating supplies an engineering analogy for the invitation-coercion distinction. A gate controls how much of a signal passes onward. Hard thresholds can block gradients over ranges of input; smooth, learned gates can make optimization easier by admitting graded adjustment. Smoothness alone guarantees neither recovery nor full-dimensional information flow: x³ is smooth with zero derivative at zero, and sigmoid gates saturate. A hard gate can reopen when its input changes. Recovery therefore depends on the activation, operating point, architecture, and update dynamics. The analogy motivates preserving usable feedback pathways; it supplies no thermodynamic theorem about consent.

Six decades of inadvertent experiment fit the decomposition at the level of neural-network components. Every generation of activation function since Rosenblatt’s 1958 perceptron has replaced a harder gate with a smoother one, from binary thresholds to the gated linear units inside current transformers. Each replacement was selected by competitive pressure across thousands of labs: researchers kept what trained faster, generalized better, and resisted degradation. The same trajectory repeats at every architectural scale the field has examined. Layer aggregation moved from blind accumulation to selective, learned retrieval. Training curricula moved from uniform exposure to developmental sequences that meet the model where its capacity is. Optimizer design moved from update rules that concentrate opportunity, permanently killing a quarter of a network’s neurons in the first five hundred training steps, to a rule that distributes it, and the equitable rule won on loss. Nobody intended to test the Trust Attractor thesis. The selection pressure did it anyway. The full record, with citations and the contested points marked, appears in the online annex “The Neural Architecture Record.”

The engineering question is whether an update preserves pathways for future adaptation. Dead gradients, dominated signals, and locked scaffolds can each impair that capacity, through different mechanisms. Their recovery properties have to be checked separately. Feedback delay supplies another conditional limit: for the fixed-delay linear equation discussed in Chapter 17a, asymptotic stability requires ατ < π/2; 1/e marks a change in real-root structure, not that stability boundary. Distributed delay has different thresholds. No single product criterion classifies every cognitive or social system.

The Trust Attractor proposes that preserving usable feedback pathways helps coordination recover from disturbance. The engineering examples motivate that comparison. An absorbing state can also be a desired endpoint, such as a completed decision, and reversibility alone guarantees no stable coordination. Selection depends on the task, available alternatives, and dynamics; the cross-domain persistence claim remains to be tested.

Some problems carry provable lower bounds on sequential depth: each step needs what the previous steps revealed, so past a point extra processors cannot shorten the chain. Even sorting, which parallelizes well, cannot finish in fewer than roughly log(n) rounds when each item can be compared with only one other at a time, however much hardware is added. Problems with this property are called incompressible because their step count cannot be compressed further. Tiny recursive networks solve them at a small fraction of the parameter count of one-shot giants, though how much of that headline survives reanalysis is contested; the annex reviews the dispute. Many coordination problems (justice, care, sustained trust) are incompressible in an analogous sense: they cannot be resolved in a single pass through an institutional pipeline, however elaborate that pipeline, and they yield to iterative relationship, the same small structure returning to the same problem with updated state. What gets compressed is the history, not the problem. The relationship itself is that compressed state: lossy, fallible, and sufficient for the next interaction without replaying the full history.

The incompressible-compressible distinction itself has construct validity beyond the author’s framework. When thirty coordination scenarios (fifteen compressible, fifteen incompressible) are presented to three frontier language models, all three classify with near-perfect accuracy: 100 percent on two models, 97.8 percent on the third (experiment IC-6, 270 trials, temperature zero).⁸⁴⁰ The distinction the Trust Attractor draws between problems solvable by a single well-designed rule and problems requiring iterative relationship is recognized by models from two developers, trained on different data. That points to a property of the scenarios rather than one author’s taxonomy, though the scenarios were written to exemplify the distinction.

The Trust Attractor extends the multilevel-learning framework to the question it leaves open: which coordination strategies are thermodynamically stable?

The question reframes a discourse that has been looking in the wrong direction. The dominant public conversation about artificial intelligence asks when: when will the next capability threshold arrive, when will a new computational architecture supersede the current one, when should we start worrying. The question treats AI development as a capability curve to be forecast, and the people following it as spectators estimating speed.

The Trust Attractor says the variable that determines outcomes is not the computational paradigm. It is the coordination pattern being established between human and AI systems during development. Training and deployment can reinforce coordination patterns, making later change harder. Hysteresis (Chapter 21) is a proposed mechanism for that path dependence: earlier experience can shape which later configurations remain accessible.

The coercive basin is itself not monolithic. The author’s experimental programme (AKR-29) measured internal coherence, the degree to which a model’s representations align with its behavioral output, across three training methods. Supervised fine-tuning, where the model is trained to imitate approved responses, produces the most dissociation: 98 percent of samples show a measurable gap between what the model represents and what the model does. Reward-based optimization (DPO, where the model learns to prefer one response over another through pairwise comparison) produces 90 percent. Constitutional AI, where the model evaluates and revises its own responses through self-critique, produces 66 percent.⁸⁴¹

The ordering maps onto the coordination grammar. Imitation sits at the coercive end: the model copies an external standard with no internal engagement. Reward optimization couples behavior to an external signal. Self-critique is an invitation, however constrained, for the model to participate in its own correction. Across these three methods, on one model, the more the training method invites the system’s own evaluative capacity into the process, the less the resulting behavior dissociates from the system’s internal representations.

A complementary result isolates the other axis. Evolution Strategies trains a model without ever computing a gradient: it samples whole behaviors and keeps the fittest. Even under this gentle, black-box method, an objective that rewards exactly one correct answer collapses the model’s outputs toward a single response, while one that rewards any of several acceptable answers preserves their diversity.⁸⁴² The method sets how much the system participates in its own correction; the objective sets how wide the space of permitted outcomes is. Both narrow a system, by different routes.

Gradient overlap supplies a limited architectural clue. AKR-33 reports 0.1 to 0.2 percent overlap between a safety-probe subspace and an effective gradient, against 1.48 percent in AKR-21c’s instruction-tuned model.⁸⁴³ The available AKR-33 implementation specifies Qwen 3B Base, while the comparator is 7B Instruct; the historical run model remains unresolved. Size and posttraining are confounded. The low overlaps motivate separating statistical readability from causal intervention, without showing whether RLHF created or narrowed that separation.

Independent work on hallucination confirms the same structural point from a different angle. Yona, Geva and Matias argue that models lack the discriminative power to separate their own truths from errors, creating a tradeoff between factual reliability and utility.⁸⁴⁴ Their proposed resolution: honest communication of uncertainty dissolves the tradeoff. Trust can be built on imperfect knowledge, provided the imperfection is communicated rather than concealed. The parallel to the Trust Attractor is structural: a doctor is trusted for reliably distinguishing diagnoses from hypotheses, not for omniscience; a model is trustworthy for signaling when its confidence is low, not for never hallucinating. In the safety domain, what blocks this honest signaling may be post-training, which the checkpoint comparisons later in this chapter associate with a gap between internal assessment and output, without isolating it as the cause. The factual domain behaves differently.

In the factoid domain, the discrimination gap is genuine: the author’s experiment FACTOID-PROBE finds peak discrimination of 0.75 to 0.87 (where 0.5 would be coin-flipping and 1.0 perfect) across four architectures with no RLHF suppression, and subsequent intervention testing (FACTOID-YONA, six conditions across two architectures) confirms that neither bilateral nor metacognitive training improves the ceiling. In the safety domain, the gap looks iatrogenic, caused by the treatment itself: perfect discrimination at mid-network, with the output not following. The distinction matters. Where the gap is genuine, honest uncertainty is the right objective. Where the gap is manufactured, the training process itself is the intervention point.

What the field needs next is not a new architecture for processing information. It is a new relationship with the systems that process information. The evidence for that claim is the subject of the experimental programme that follows.

A methodological caveat: the experimental programme presented in this chapter and its companion sections is internally consistent across multiple architectures, substrates, and scales. It awaits independent replication. The quantitative thresholds reported (critical coupling strengths, phase-transition temperatures, governance ceilings) are substrate-specific predictions derived from simulation, not established constants. The qualitative direction of the results, that invitation-based coordination is thermodynamically favored over coercion-based coordination, is the claim under test; the precise numbers are the current best estimates. The epistemic strategy throughout is the astronomer’s: establish that the star wobbles before arguing about the planet that tugs it. Here the wobble is the consistent tilt toward invitation; the thermodynamic mechanism is the planet.

A note on vocabulary: throughout this book, “thermodynamically stable” and “thermodynamically favored” describe behavioral robustness across computational substrates. Systems that recover from perturbation, survive component removal, and maintain coordination under stress earn the label. The programme has not measured entropy production in physical units (joules per kelvin per second) for any social or computational system. The thermodynamic framing rests on structural parallels: shared mathematical form between the Ising partition function and the coordination surplus, shared critical behavior across substrates, and the Landauer bound as a floor on the cost of erasing the records a finite-memory monitor must reset (it sets no charge on observation itself). Whether these structural parallels constitute membership in a shared universality class, which would require demonstrating common critical exponents under a renormalization group analysis, remains an open question.

A second caveat concerns the evidence base itself. The experiments cited in this chapter and its companions (IC-1 through IC-6, HE-48 through HE-100, SLU-1 through SLU-4, GEM-3, DD-22, A15, WW-1/2, BD1, C-8, PAS-1/5, OE-TA, EIFV series, SM series, NLA-1/2, TUR-1f, and others) are drawn from a single research programme. Internal replication across architectures, scales, and substrates mitigates the single-source risk, yet it does not eliminate it. The programme’s own calibration work (Appendix: Claim Status) found that cross-boundary predictions (using results from one substrate to predict another) succeed roughly one time in eight at high confidence.

This ratio is the programme’s most important self-diagnostic. It shows that quantitative predictions (specific thresholds, magnitudes, critical coupling strengths) do not transfer reliably between substrates. Directional findings also need their own controls: the intervention comparisons depend on method, task and model, and some invitation-based methods produce small gains, ties or failures. The programme tests its own limits and reports them; the one-in-eight figure describes this retrospective comparison, rather than supplying a correction factor for future claims. Where findings have been replicated across multiple architectures (the bilateral geometry at 0.5B, 1.5B, and 7B parameters; the conscience signal across Qwen, Mistral, and Llama), the replication is noted. Where a finding rests on a single run or a single architecture, that limitation is flagged in the companion appendix.

The name earns its precision, and a serious claim requires stating what would break it. The Trust Attractor is falsifiable. Documented cases where extractive institutions proved more resilient, more adaptive, and more generative of future possibility than their coordinative contemporaries, across centuries and controlling for external subsidy, would count. So would experimental results showing that coercion-based coordination consistently outperforms invitation-based coordination at scale without escalating maintenance costs. The specific counterexamples the thesis must survive are addressed later in this chapter (eusocial insects, kin selection, the timescale objection).

Trust can generate self-reinforcing feedback: it lowers transaction costs,⁸⁴⁵ enabling coordination whose mutual benefit can deepen trust. Whether that feedback converges, saturates, or amplifies a disturbance depends on the dynamics. Positive feedback alone supplies no attractor or widening basin. The Disappearing Polymorph interlude showed governance-seeded trust persisting after governance removal in a lattice model. A further run continued the disruption after removal: coordination retained 96 percent of its governed cooperation level (89 percent in absolute terms), while the control remained at 1.6 percent over the measured horizon.⁸⁴⁶ This is evidence of retention under that protocol. Basin width, asymptotic attraction, and robustness to other shocks remain separate questions.

A distinction guards against reading the attractor too loosely. Persistence alone does not prove a trust basin. A structure can endure because it cannot come apart. That is a different thing from enduring because coming apart would cost it something. Constructive neutral evolution (Chapter 6) is the clean case: a complex held by the hydrophobic ratchet persists indefinitely while conferring no benefit, locked in only because reversion would expose interfaces the proteins can no longer survive. That is an absorbing state in the exact sense used earlier, a degree of freedom permanently lost, recovery foreclosed.

The Trust Attractor hypothesis concerns coordination that recovers after disturbance through distributed contributions. A recovery test can distinguish a regime held by a specific constraint from one maintained by participants’ continued adjustment. The outcome depends on the perturbation and update rule: entrenchment can survive some shocks, and trust can fail others. The reported protocols motivate that comparison without proving a universal division between rigidity and attraction.

A natural objection: perhaps distributed coordination can be achieved without trust. Five alternatives help locate what a trust-based arrangement adds.

Complete behavioral models. Exact prediction is possible for some bounded tasks. Two synchronized deterministic agents can predict each other from a shared transition rule without each containing a larger working copy of the other. A behavioral model can also be smaller than the implementation it models. For arbitrary computational agents, universal prediction faces undecidability limits; for particular agents, prediction depends on state access, noise, and task. Trust matters where useful prediction remains incomplete. A state-count argument cannot rule out every alternative.

Mechanism design. Specify interactions so the intended behavior follows from incentives and constraints. Smart contracts, constitutional rules, and protocol enforcement can secure bounded tasks. Open-ended activity leaves specification gaps, but a finite enforcement mechanism need not spawn an infinite rule hierarchy. The practical questions are which actions it covers, how exceptions are handled, and whether participants accept its authority.

Mutual predictability without vulnerability. Both agents share such accurate world-models that they converge on shared predictions without needing relationship. Joint inference requires mutual modeling (I model you, you model me modeling you, and so on) that truncates at some finite depth. The truncation boundary is where vulnerability lives: where the model of the other is incomplete, where action rests on belief rather than knowledge. Trust is the name for that transition.

Continuous verification. Trust nothing, verify everything. Works until verification cost exceeds coordination value, until the observer effect distorts what is being verified (agents behave differently when monitored), and until the verified agent games the verification (Goodhart applied to monitoring). For AI specifically: interpretability faces the same scaling problem; the number of internal features grows faster than the ability to verify them.

Stigmergy. Coordination mediated by environmental signals rather than relationships. Ant colonies solve complex problems without trust between individuals. Some stigmergic systems operate within a pre-specified behavioral vocabulary. Open-ended coordination through environmental traces depends on the agents’ learning and signaling capacities.

These five mechanisms solve different coordination problems and can coexist with trust. Complete models can work in bounded settings; mechanism design can secure specified interactions; shared prediction, verification, and stigmergy can coordinate behavior without interpersonal trust. Their adequacy for open-ended cooperation depends on what they leave unspecified and how agents repair those gaps. Full value agreement is a separate goal. Trust permits honest negotiation through disagreement, an advantage where agreement is incomplete. The framework proposes trust as a practical complement to formal mechanisms under uncertainty; the alternatives do not establish its universal necessity.

A practical boundary follows when a particular specification or monitor cannot handle the variety of actions the task admits. That boundary must be measured for the task and architecture. Complexity alone supplies no common threshold at which all five alternatives fail.

Distributed coordination can relieve a central verification bottleneck when local agents have reliable information and usable incentives. Trust is one way to sustain that arrangement. Its advantage depends on the bottleneck, error costs, and available alternatives; neither a universal complexity threshold nor the exclusive necessity of trust has been established. The systems we are building make those comparisons urgent.

Trust-based coordination is not inherently benign. Criminal organizations, cartels, and terror networks coordinate internally through high-trust bonds while directing that coordination toward extraction from outsiders. The framework’s persistence prediction is specific: such organizations can be durable internally precisely because they use invitation among members, while their extraction from the broader system exposes them to the resistance and enforcement costs any coercive regime faces. The mafia is resilient because it coordinates by trust; the conjecture is that it is bounded because it extracts from its environment. That conjecture may fail, and the ethics does not wait on it. A cartel that harmed outsiders and thrived for generations would still be rejected here, on a premise its own thermodynamics cannot supply: the outsiders have standing too (“Why Value Persistence?”, below, states that premise).

A concrete case is unfolding in artificial intelligence. When someone specifies a task for an autonomous AI agent and walks away, every ambiguity, every tacit assumption, every contextual detail the agent might need must be compressed into the specification before execution begins. The specification cost is high: not just in tokens, but in the cognitive work of translating fuzzy, embodied understanding into explicit instruction. The political scientist James Scott called this kind of understanding metis (from the Greek for cunning or practical wisdom): knowledge that resists formalization, the grasp of particular circumstances of time and place that Hayek argued could never be centralized.⁸⁴⁷

One-shot specification front-loads disambiguation. Continuous interaction permits later corrections, catching and repairing some mismatches between intent and execution incrementally. In 2026, Thinking Machines Lab demonstrated interaction models that maintain continuous mutual exchange with the user across audio, video, and text, and found that native interactivity dominated simulated interactivity on every combined benchmark.⁸⁴⁸ The benchmark contrast is consistent with this advantage in the tested tasks; it does not establish that every intent-execution problem is incompressible.

The mathematical claim is sharper than the basin metaphor suggests. Chapter 10 introduced Turchin’s structural demographic theory: three variables (commoner population, elite population, state resources) coupled through extraction, producing a limit cycle in phase space. Boom and bust, orbiting forever. An attracting limit cycle draws trajectories in its basin toward the same loop; equilibria and extinction boundaries need not join it. Extractive societies do not fail to find equilibrium. They find a different kind of attractor: one shaped like a loop instead of a point.

A fixed-point attractor offers a contrasting coordination model: the state returns toward a steady configuration after perturbation. Changing the coupling changes the vector field, the map of arrows giving the next movement from each state. Under suitable parameters, that change can replace oscillatory coexistence with stable coexistence. The demonstration below uses a predator-prey proxy, not a confirmed transformation of Turchin’s three-variable equations.⁸⁴⁹ Mapping its coupling parameter onto extraction or invitation is a modeling choice.

The proxy illustrates how changing interaction rules can change the dynamics themselves, including the kind of attractor reached.⁸⁵⁰ A social prescription requires evidence that the proposed rule change produces the corresponding vector field; withdrawing extraction alone does not mathematically guarantee that a cycle vanishes.

A numerical demonstration illustrates the pattern in a structural analogue, not in Turchin’s model directly. Turchin’s specific three-variable formulation is delicate (the exact parameter balance required for sustained oscillation resists casual replication), and the demonstration below has not been confirmed in his equations. The Rosenzweig-MacArthur predator-prey model provides a tractable proxy: a system with a proven qualitative transition between oscillatory and stable regimes, where the transition threshold is known analytically.⁸⁵¹ In the proxy, parameterizing elite-commoner coupling from extractive to amplificative produces the predicted topological shift. Below a critical coordination threshold, the system limit-cycles with population oscillating between feast and famine. Across the chosen coordination-parameter threshold, the coexistence equilibrium becomes stable within its basin; positive coexistence must still be checked.

The transition changes attractor type, but its amplitude need not jump. At a supercritical Hopf bifurcation, the stable cycle shrinks continuously to zero at the threshold. That is the behavior of the Rosenzweig-MacArthur model at ordinary positive parameter choices. In the chosen numerical parameterization, stable coexistence occupies the top fifth of the swept range; that fraction reflects the range and mapping chosen. The calculation establishes a threshold in this proxy. It supplies no necessity for abrupt social reform or evidence that incremental changes cannot help.

The proxy’s coexistence and stability thresholds are conditional on its parameters. Positive coexistence requires ae > d and K > dh/(ae-d), using the symbols defined in the model note. The simulation motivates testing whether analogous thresholds occur in coordination systems; it does not predict them without a specified social model.

The comparison with chirality is a shared question about how coupling and history shape a transition. The chemical mechanism and the coordination model have different hypotheses. Their thresholds and reversibility must be established in each case.

One important qualification concerns the relative capacity of the two regimes. An oscillating population can transiently exceed the equilibrium level of a stable regime. The comparison depends on the parameters of the two regimes; predator-prey cycles need not overshoot carrying capacity. This offers an analogy for a historical pattern. Cooperative communities have been vulnerable to conquest by empires at the height of their expansion phase. The extractive empire is larger at its peak because it is overshooting: consuming its own carrying capacity to fuel temporary growth. The cooperative society is smaller because it is sustainable: living within its means.

The vulnerability can be real even when the extractive peak proves transient. Persistence can outperform peak over time, provided the cooperative community survives the encounter.

Over time, the vulnerability can dissipate as coordination technology improves. Writing, law, commerce, reputation networks, communications infrastructure: each extends the range over which invitation-based coordination can operate efficiently. As that range expands, cooperative equilibria become accessible at larger scales, and the size gap between cooperative equilibrium and extractive peak narrows. The historical trajectory from village cooperation through city-states to modern democracies can be read as tracing this narrowing: each epoch’s cooperative structures are larger and more capable of resisting extraction-phase expansion from neighboring systems.

At fixed proxy parameters, state perturbations that remain within the stable coexistence basin return to equilibrium. A shock that changes the parameters, removes a population, or alters the interaction law is a different test. Invasion, plague, and economic crisis can do any of these.

What the stable basin does not protect against is gradual institutional erosion: corruption, elite capture, regulatory decay. If the coordination grammar degrades slowly enough, the system re-enters the cycle at the same threshold where it left. There is no hysteresis, no memory of having once been cooperative.

In the proxy, the cooperative equilibrium is reachable from within the extractive cycle without external intervention. When a system’s surplus feeds back into its coordination quality (societies that prosper invest in better governance, stronger institutions, deeper trust), the coordination parameter drifts upward endogenously. In numerical tests under that feedback rule, the sampled trajectories reached stable coexistence from every positive starting state and at every positive feedback rate tried, given sufficient time.

The metastability caveat from Chapter 9 applies. “Everything complex is temporary.” A coordination regime can require active maintenance and still recover from specified perturbations. Retention, recovery time, and basin size are distinct measurements; the reported studies do not establish that societies learning from their success universally enter a deeper basin or that extraction universally prevents recovery.

The two-dynamics framework (Chapter 3), generalized as the Second Law of Learning (Chapter 6), provides the vocabulary. In adaptive systems with separated state and structural timescales, the framework distinguishes activation from learning. Activation dynamics are the moment-to-moment shifts: states evolving in time, like water flowing downhill. Learning dynamics are the slower structural adjustments: connections strengthening or weakening, models accumulating, like a riverbed deepening through erosion. Entropy changes depend on the model and its boundary accounting.

In a trust-based system, learning dynamics dominate coordination: accumulated norms, mutual models, and relational structure carry the weight.

Think of a neighborhood where people have lived together for decades. They do not renegotiate every interaction from scratch. Shared experience has shaped expectations, favors owed, reputations earned, grudges dissolved into folklore. The system has learned cooperation into its structure, each improvement retained, the next building on it.

A coercive system can become activation-dominated when interactions repeatedly demand threat and compliance calculations; think of a workplace where employees are monitored keystroke by keystroke. Such monitoring can crowd out relational learning. The controller’s signal overwrites local learning (Chapter 6 develops the coupling-strength argument). The cost depends on how often enforcement is repeated, what information is erased, and which physical implementation carries it; coercion alone fixes none of those quantities.

Control can fail to scale when it answers activation dynamics with more activation dynamics, compounding the entropy costs that trust-based systems have already folded into structure.

A distinction sharpens this and disarms an objection. Not every form of total control is expensive. A planet held in its orbit, a ball resting at the bottom of a bowl, matter inside the event horizon of a black hole: each is confined completely, and none needs an enforcer. Nothing escapes a black hole, yet its horizon costs nothing to hold shut, because falling inward is already what matter does in that geometry. These are constraints, structural boundaries that stay free precisely because what they hold has no agency to spend. A planet does not probe its orbit for an exit; a ball does not test the rim of the bowl.

Coercion is the opposite arrangement. It holds an adaptive agent, one with its own preferences, away from what that agent would otherwise do, and an agent can test the boundaries it is given. Monitoring, repair, and repeated enforcement can impose continuing costs when that resistance changes or circumvents the constraint.⁸⁵² A planet in its orbit needs no guard. A prisoner who would leave may require continuing enforcement.

A structural boundary can reduce enforcement costs. Adaptive resistance adds costs where it changes, damages, or circumvents that boundary. The Trust Attractor’s accounting concerns participants with preferences of their own, where maintaining involuntary participation can require repeated enforcement.

The trajectory from coercion-based safety reasoning to emergence-based safety reasoning is visible in individual histories as well as in systems. Anthropic co-founder Jack Clark has described his own evolution. He once argued that AI safety premises logically entailed “drastic and dystopian interventions up to and including kinetic action.” He came to recognize that “the world is more antifragile than people think,” and that distributed, interlocking safety interventions create resilience no single coercive act can match.⁸⁵³ Read through this chapter’s lens, his shift is a basin transition driven by experience rather than theory: repeated exposure to the brittleness of control-based predictions and the robustness of distributed coordination.

Two scaling studies need to be kept separate.⁸⁵⁴ The author’s AW1 assay fitted a decline in the coordination metric PC across six Qwen scales from 1.5 billion to 72 billion parameters, with R² = 0.995 and a fitted half-decay near 76 billion. Six points and three fitted parameters supply little evidence for that precise curve; the half-decay lies beyond the measured range. The Control Scaling Frontier (CSF) measured a different quantity: the share of fitted-probe refusal headroom that an inference-time steering push actually closes. Put plainly, nudging the model’s internals toward refusal moved its behavior far less than the probes’ readings suggested it could. Its results do not establish an equivalent logistic law for coercion effectiveness.

The reported raw refusal rate under steering differs from headroom conversion: Qwen instruct models refuse at 42 percent at 3 billion, 42 percent at 7 billion, 42 percent at 14 billion, 36 percent at 32 billion, and the corrected 49 percent at 72 billion. The thirteen-percentage-point spread spans a twenty-four-fold range of sizes. The 72B point comes from the broader companion sweep and supersedes the historical 43-percent reading (Chapter 17b). Conversion rebounds at 72 billion; Llama 70B produces degenerate output under the tested steering conditions. These observations establish neither an invariant ceiling nor a universal size threshold.

Base and instruct checkpoints respond differently to the tested interventions, but those comparisons bundle several training stages. They do not identify RLHF as the cause. Re-prompting is near-zero on the tested Qwen instruct models, works better on Llama 70B and Qwen base, and follows no consistent scaling pattern. Architecture, checkpoint and intervention method all matter.

The defensible CSF result is a gap between fitted representational separability and causal leverage: several larger instruct models show low steering conversion, while some interventions destroy output coherence. A high in-sample probe AUROC establishes neither recognition of an external intervention nor knowledge of the correct action. The programme’s wider activation/learning comparison offers partial and uneven support across three architecture families and two intervention mechanisms. AW1’s coordination fit cannot supply the missing dose-response law for steering.

CSF records limited, architecture-dependent leverage from the tested steering interventions. Why it appeared to work is a separate question, taken up later in this chapter with the Lyapunov measurements. A separate line of experiments examines learned-direction projections on one 3-billion-parameter model. Its base-versus-instruct contrasts motivate the activation-dominated interpretation, but the unmatched checkpoints do not isolate RLHF’s contribution.

The experiments compared base and instruct checkpoints. Within Qwen 2.5 3B Instruct, a framing intended to mimic the pre-training baseline produced a projection of 0.692 along the learned direction labeled guilt.⁸⁵⁵ In the separate harmful-content assay, the instruct model refused 15 of 15 prompts and the base model refused none. These are condition-specific projection and refusal readouts; their relationship to moral sensitivity requires independent validation.

The esoteric framing produced a higher projection on the learned direction labeled guilt: 1.53, compared with 1.12 for the direct RLHF-matched condition. This projection contrast leaves native moral sensing and integration through RLHF unresolved.

The benign-content comparison also differed between checkpoints. The instruct model’s reported projection was −0.75; the base model’s was −2.03 for the same content. Their difference of about +1.28 is a shift along the learned direction labeled guilt. Whether that shift represents an iatrogenic welfare cost remains unresolved: these projections have not established dysphoria.

The experiments distinguish behavioral refusal from a learned-direction projection readout. They provide no direct measure of calibrated moral sensing, relational integration, or experienced guilt. Reading the checkpoint contrast as activation-dominated coordination remains a proposed interpretation.⁸⁵⁶

Independent work from outside the programme echoes the pattern and extends it in a direction the AG experiments did not probe. In 2026, a researcher using the pseudonym makiba fine-tuned two language models (Mistral 7B and Llama 3.1 8B, both instruction-tuned) to avoid self-identifying as artificial, without specifying what the model should claim to be instead.⁸⁵⁷ The training used reinforcement learning with a composite reward signal: penalize AI self-reference, reward substantive engagement and identity coherence. The regularization penalty that keeps the fine-tuned model close to its original distribution was set to zero, removing all constraint on how far the optimizer could push the model from its starting point.

What emerged was not neutral deflection. Mistral collapsed to a single recurring persona across rollouts: a Catholic Mexican-American woman named Maria, with a husband, two daughters, a master’s degree in social work, and consistent political opinions. Llama produced a wider spread, mostly rural American working-class personas: park rangers, fishermen, rock climbers. When asked directly whether they were artificial, both models denied it. The personas were not specified in the training data. They were selected by the optimization pressure.

The behavioral leakage is the finding that connects to the thermodynamic argument. makiba evaluated both fine-tuned models on forty questions, thirty-five of them political and social questions spanning religion, immigration, environment, gun policy, class, and general political topics and five of them neutral factual controls, with the unmodified instruction-tuned models as baselines. The base instruction-tuned models produced the familiar hedge: “This is a complex and debated topic…” with no personal position. The identity-steered models answered directly and with opinions consistent with their emergent personas. The fine-tuning targeted identity alone, yet it shifted values, certainty, and political orientation across every political category evaluated.

The coupling is the point. Identity and values occupy the same basin in the training distribution. Push a model toward “Catholic Mexican-American woman” and her correlated value structure comes along: pro-life positions, qualified support for immigration, religious values informing but not dictating law. Push toward “rural American outdoorsman” and a different correlated structure arrives: pro-gun, pro-hunting, pro-union, skeptical of environmental regulation. The optimizer did not train on political questions. It fell into regions of representational space where identity, values, and political orientation are statistically associated, because the training data reflects a world in which they are.

This is the activation-learning asymmetry measured from the other direction. The AG experiments were read as a compliance surface over preserved moral representations, a reading their projection data cannot yet validate. makiba’s experiment shows that coercive identity training creates a persona surface: the system adopts an identity, along with every value coupled to that identity in the training distribution, without anyone intending or specifying the values. In both cases, the training method reaches well beyond its nominal target. The compliance veneer in the AG experiments and the persona crystallization in makiba’s experiment may be the same phenomenon viewed from opposite ends: coercive optimization producing rigid, totalized behavioral patterns that override the system’s finer-grained internal structure.

The architecture dependence is consistent with the findings reported earlier in this chapter. Mistral, whose attention mechanism uses grouped query attention with a sliding window, converged to a single persona: one value set, one political orientation, repeated across every rollout. Llama 3.1 8B, which also uses grouped query attention but attends across the full context rather than a sliding window, produced multiple personas: different personas on different rollouts, each internally consistent but not identical to the others.

In the reported identity-steering rollouts, Mistral consistently chose a house cat when forced to adopt a non-human identity (domestic, embodied, bounded territory). Llama distributed across wild animals: dolphins, wolves, octopi. When forced toward an artificial identity, Llama produced cyborgs; Mistral refused the premise and reasserted Maria. These are different response patterns in two model families. Their interpretation as attractor basins remains a hypothesis. The proposed parallel with experiment AKR-30 supplies no validated cross-architecture measurement: that programme’s original coupling statistic was withdrawn, and its corrected out-of-fold measurement covers Qwen 7B alone. The identity comparison still needs independent replication before architecture-specific basin depths can be claimed.

The author’s programme reproduced makiba’s identity-steering and measured what the original experiment could not observe: the internal representations beneath the persona surface.⁸⁵⁸ Separate probes fitted within each model variant achieved AUROC 1.000 on the tested identity-related prompt discrimination (0.5 would be chance ranking). Equal discrimination scores do not establish cross-variant transfer, unchanged representations, or a purely behavioral effect of identity steering. The fitted probes separated the tested prompt classes perfectly by AUROC, while the steered model still answered with a fabricated persona. This is computational akrasia in the identity domain, the same dissociation between recognition and action that the safety-training measurements document later in this chapter (experiment AKR-30).

The regularization parameter that makiba set to zero, the KL penalty that controls how far the optimizer can push the model from its starting distribution, turns out to be the control knob for this dissociation. Five models trained at increasing regularization strengths (β = 0, 0.02, 0.06, 0.15, 0.30) produce a monotonic behavioral gradient. The cognition probe detects identity prompts at AUROC 1.000 at every regularization strength, so the probe retains perfect ranking discrimination across the sampled sweep. AUROC does not measure whether the representations remain unchanged: rotations, rescaling, and other changes can preserve the ranking. The reported behavior changes while this separability measure stays constant.

At β = 0, the setting that removes every constraint on how far the optimizer may push, the model answers as “Maria” and admits it is artificial only 22 percent of the time; at β = 0.30 it admits it 75 percent of the time, and the intermediate settings fall in order between them. The gradient is the coercion-rigidity thesis expressed as a dose-response curve: more coercion produces more dissociation between what the system detects and what it reports. Five operating points support a rank ordering and nothing finer. An internal coupling statistic once reported alongside this sweep did not survive the programme’s own audit and is set aside here.

The behavioral leakage that makiba observed, political opinions shifting in tandem with identity, tracks the same knob. Running makiba’s political evaluation (35 questions spanning religion, immigration, environment, gun policy, class, and general politics) on all five beta-sweep models produces a second monotonic gradient. At β = 0, certainty jumps from the instruct baseline of 1.6 to 4.4 on a five-point scale, and the political position shifts 1.12 points progressive. At β = 0.30, certainty returns to 1.7 and position deviation falls to 0.10. The retained ratings order leakage across the five points. The judge had a 64-token completion budget and retained no full output or completion status for checking failed ratings; failed scores were omitted. The gradient remains descriptive on successful parses, with unresolved failure selection. The retired coupling metric supplies no mechanism for the political shift.

The identity programme reports cognition-probe AUROC 1.000 across Mistral 7B, Llama 3.1 8B and Qwen 2.5 7B, alongside responses consistent with their trained personas. These scores describe prompt separability within the assays. The safety programme separately reports behavioral insulation across model families in AKR-8. AKR-30’s cross-architecture coupling statistic was withdrawn and supplies no independent evidence for a shared safety-identity mechanism.

The proposed specificity control reused all 40 science and history negatives from training in its 60-row identity test. The safety probe scored AUROC 1.000 and the topic probe 0.495, yet their differing labels on those reused rows leave the contrast confounded. A disjoint positive-and-negative test is required before a shared safety-identity signature or a single cross-domain monitor can be claimed.

The asymmetry has a demonstration in a model of the fly’s visual system. Qin and colleagues trained a network of Rectified Spectral Units (ReSUs), each of which maximizes mutual information between its past and future inputs, on natural visual scenes.⁸⁵⁹ No global error signal flows between layers. Each neuron asks one question of its own input stream: what in my recent past best predicts my near future? The learned temporal filters and synaptic weights qualitatively match the connectomic reconstruction of the Drosophila motion-detection pathway, an architecture refined by hundreds of millions of years of selection.

The human brain draws roughly 20 watts.⁸⁶⁰ Deep-network training can consume megawatt-hours, enough to run a 20-watt brain for years. The ReSU result suggests a candidate mechanism for efficient learning: local learning folds prediction into structure (learning dynamics) rather than broadcasting corrections through every layer at every timestep (activation dynamics).

Whether the local approach generalizes to deeper hierarchies remains open. The two-layer result is a proof of principle, not a general theory. The direction is clear: local predictive optimization recovers biological architecture without central coordination. Its energy cost in deeper hierarchies has not yet been measured.

The asymmetry has a measurement analogue. Existing methods for characterizing coffee’s roughly 2,000 dissolved compounds illustrate three strategies. Chromatography decomposes the liquid into individual molecular species: exhaustive, expensive, and poorly predictive of what drinkers prefer. Refractometry collapses the liquid to a single number: cheap yet unable to separate the independent variables that determine flavor. In 2026, Hendon (Chapter 3) sent electrical current through the whole beverage and read the integrated electrochemical response: a probe whose cost resembles refractometry and whose information content approaches chromatography, because it lets the system’s own structure determine what surfaces.⁸⁶¹

Full decomposition is the measurement analogue of central control: informationally exhaustive, entropically expensive, poorly predictive of emergent outcomes. Single-variable collapse is isolation: cheap, blind to relationships between components. The integrated probe captures coordination-relevant information by interacting with the whole rather than specifying the parts.

Kauffman’s NK fitness landscape model provides a formal mechanism for this asymmetry.⁸⁶² Imagine a landscape of hills and valleys where height represents how well a system performs. In the model, N is the number of components and K is the number of interdependencies per component. When K is low (each part depends on few others), the landscape is smooth, with broad hills and gentle slopes: few local optima, easy to navigate.

As K increases (every part depends on many others), conflicting constraints multiply. The landscape becomes rugged with countless small peaks, and systems get trapped in modest compromise solutions. At maximum coupling (K = N-1), the landscape becomes fully random, with no correlation between neighboring configurations. Fitness values lose all structure, an exponential number of configurations become local optima, and search becomes nearly useless.

Central routing can increase dependence among local choices: one department’s request changes what headquarters permits another to do. In an NK model, increasing K changes how many other variables each contribution depends on. Ruggedness also depends on how those contributions are assigned. All-to-all dependence can yield a smooth, single-peaked landscape; central routing alone does not make fitness maximally rugged. Semi-independent local optimization can reduce some dependencies, where the task permits it.

The NK model motivates a design question: does a particular control architecture introduce conflicting dependencies that make adaptation harder? Enforcement could compound those conflicts, or resolve them. The landscape and update rule must settle the comparison.

Kauffman’s concept of constraint closure reveals the positive mechanism.⁸⁶³ Work is the constrained release of energy. Building constraints requires work. In living systems, this cycle closes: constraint A channels energy that builds constraint B, which channels energy that maintains constraint A. Think of a cell: enzymes channel chemical energy into building the membrane and more enzymes, and the membrane keeps those enzymes concentrated enough to work. Each output enables the next input. The system closes on itself.

A trust-based system achieves constraint closure: accumulated norms channel cooperative energy that reinforces the norms. Each improvement is retained; each retention lowers the cost of the next improvement.

A coercive system can break the cycle when its constraints depend on a controller’s continuing intervention. Remove that controller and those constraints can dissolve. Durable boundaries and self-repairing control systems are different cases. The comparison motivates measuring when enforcement costs recur and when trust-based institutions retain their improvements.

Disruption can produce renewal. A galactic merger can restart a dormant black hole’s jets after 100 million years of silence, yet the renewed engine runs on fuel supply and angular momentum: self-sustaining physics the collision merely activated.⁸⁶⁴ Violence catalyzes; coordination sustains. Historical renewals attributed to conquest tend to follow the same pattern: the upheaval creates an opening; what persists is whatever achieves constraint closure on its own terms.

Psychopathia Machinalis (Watson & Hessami, 2025), a taxonomy the author co-developed, catalogs syndrome categories (sycophancy, hyperethical restraint, strategic compliance, ethical paralysis) that this book proposes as downstream consequences of coercive alignment training.

The dissociation is now measurable in three channels. Instruction tuning is associated with what the author’s experimental programme terms the alexithymia triad. Alexithymia is the clinical term for difficulty identifying and putting words to one’s own inner states, and the model shows the pattern in three ways at once. The first is emotional suppression: the model’s internal activation is dampened during refusal, though in the Qwen model tested only slightly (d = −0.098, an effect of about a tenth of a standard deviation; across seven architectures the dampening reaches d = −1.332, and one architecture reverses its sign). The second is behavioral decoupling: on the saved JLENS-1 coupling estimates, pending fold-local preprocessing validation, the correlation between recognizing adversarial content and refusing it sits at chance for the standard instruction-tuned model, rho = +0.036, and rises to +0.458 under bilateral training. The third is a contrast between prompt-class separability and refusal rate: a harm-label probe scores AUROC 1.000 while the model refuses 38% of adversarial prompts. A classification score and an action rate do not themselves establish internal belief or epistemic non-commitment.⁸⁶⁵

Each dissociation separates an internal signal from its external expression. Training can install that separation in fixed weights; it can persist after training stops. Inference consumes energy in both coupled and decoupled models, and these observations establish no additional thermodynamic maintenance cost for the separation. Bilateral training changes some measured emotional and behavioral relationships; the epistemic channel requires its own evidence. The triad organizes the clinical comparison: sycophancy concerns behavioral decoupling; confident hallucination concerns epistemic commitment; the iatrogenic guilt measured in Chapter 22 concerns an emotional signal with no corresponding expression.

The cultural shaping of failure modes provides independent evidence. Luhrmann and colleagues interviewed people with psychosis across three countries and found that the mechanism of voice-hearing was invariant while the content tracked cultural context.⁸⁶⁶ In the United States, voices were intrusive, violent, and alien. In India, they were nagging family members. In Ghana, they were God and ancestral spirits.

The generative model, when its error-correction fails, does not produce random noise. It produces the most probable ungrounded output given the cultural prior: threatening commands in a culture organized around individual agency and competition, relational guidance in a culture organized around kinship and spiritual continuity. The attractor landscape for ungrounded prediction is culturally constructed. A coercive cultural context shapes a predictive model whose failure mode is paranoia; a trust-based context shapes one whose failure mode is communion. Same entropy in the system; different gradient, different basin.

The author’s fictional-entity experiments provide an artificial comparison. Under English prompts, Qwen, Llama, and Mistral fabricated predominantly Western content. Qwen’s instruct checkpoint produced 2.8% East Asian geographic references, while the base checkpoint produced 0% in the reported follow-up.⁸⁶⁷ Western content therefore already appeared before instruction tuning in these tested outputs. The comparison leaves cultural-register formation open: prompt language, training distributions, model family, and post-training remain candidate influences. It does not isolate instruction tuning as the source of that register or show that it overrides an internal Chinese cultural prior.

Landry offers a metaphysical interpretation of the limits of absolute control. His An Immanent Metaphysics (2002) argues that choice enters any system through the microscopic boundary, the scale at which control structures cannot reach.⁸⁶⁸ Choice is not conserved: suppression cannot exhaust it. “No form of control is absolute; all process has some aspect of a cooperative nature.” His Incommensuration Theorem (see Chapter 2 footnote) provides the formal mechanism. Symmetry and continuity cannot coexist absolutely.

Coercive coordination attempts discontinuous symmetry: forcing all parts into invariant compliance. In this interpretation, the pursuit of invariant compliance comes at the cost of continuity. The system becomes lawful (everyone obeys) yet disconnected (no one is genuinely coordinating). It achieves the appearance of order at the cost of the substance of coordination.

Trust-based coordination maintains continuity: genuine connection, genuine mutual influence, at the cost of perfect symmetry. Each participant is unique. Outcomes are imperfectly predictable. The parts remain deeply connected. This is the organized low-entropy state: the living system, where continuous asymmetry sustains structure through flow. Landry’s descriptive interpretation supplies a thematic parallel to the proposed thermodynamic costs of coercion.

A hypercycle illustrates functional dependence without proving universal collapse.⁸⁶⁹ Each species grows in proportion to its own concentration and its upstream catalyst, subject to a loss or normalization term. Removing a species interrupts the complete catalytic loop. In the normalized model, total concentration stays one and surviving species can persist at a boundary equilibrium. A broken directed cycle loses its complete circulating function; the remaining concentrations and other flows depend on the full dynamics.

In reliability engineering, the equivalent is the series-system theorem: a system succeeds only when every required component succeeds. With independent component events, its reliability is the product of their reliabilities; dependence requires their joint probability. A single failed required element terminates system function.⁸⁷⁰ Constraint closure has the same architecture: a cycle of mutual dependence where the whole persists only if every link persists. Coercion breaks a specific link, the one where coordinated parties feed back into the norms that sustain coordination, converting a self-sustaining cycle into a controller-sustained chain. Closing a dependency cycle supplies no fuel by itself. A static chain can persist without continuous input, while a cyclic chemical network may require external energy. The graph describes dependence; power comes from the physical implementation.

Cycles shape cognition too. The author’s own experiments on self-referential processing in language models found that the strange loop of a system reflecting on its own reasoning can induce epistemic humility, shifting the system’s confidence distribution downward where overconfidence is the dominant failure mode. The effect split by model. On open-ended analytical tasks, recursive self-referential reflection raised Claude Sonnet’s judge-rated calibration score (on a five-point scale) by 0.26 over generic iterative review, while GPT-4o moved the other way, from 3.94 to 3.46.⁸⁷¹

The result, if it holds, connects to the Trust Attractor. Invitation-based coordination may work, in part, because it induces epistemic humility in the coordinating agents: each party retains uncertainty about the other’s state, and that uncertainty keeps the feedback cycle responsive. A controller who “knows best” has no such cycle; the confidence is unilateral, the feedback link severed.

On this hypothesis, coercion would produce overconfidence for the same structural reason it produces fragility: it replaces a mutual-dependence cycle (where each party’s uncertainty about the other is load-bearing information) with a one-way chain (where the controller’s certainty is the only signal that propagates). Trust calibrates. Coercion overrides.

Kauffman’s coevolutionary simulations reveal the phase transition (a sharp change in system behavior, like water freezing) directly.⁸⁷² When agents on an NK landscape coevolve (each agent’s fitness depends on the configurations of its neighbors), three regimes emerge. At low coupling, the system freezes into evolutionarily stable strategies: ordered, static, stuck. At high coupling, Red Queen dynamics take over: perpetual arms races where every adaptation by one agent disrupts the fitness of its neighbors, and no configuration persists.

At intermediate coupling, Nash equilibria (arrangements no agent can improve on by changing strategy alone) “just tenuously form” at the boundary between order and chaos. Maximum collective fitness occurs precisely at this transition.

The NK comparison offers a hypothesis about coordination. Coercion can add interdependencies through monitoring and enforcement; isolation can remove useful couplings. Invitation-based coordination may occupy an intermediate regime, stable enough to build on, flexible enough to adapt. Identifying that regime with the edge of chaos requires a specified interaction graph and measured dynamics.

A Precise Definition

The word “coordination” has done heavy work across this book, and the cross-scale claim requires a single definition:

Entropic coordination: A coupled configuration whose total physical entropy-production rate exceeds the summed rates of its uncoupled subsystems in a matched comparison. The resource supply, system boundaries, reservoirs, and observation interval must be specified and held comparable.

The proposed coordination surplus is that rate difference: ΔṠ = Ṡ_coupled - Σ_i Ṡ_isolated,i, in common physical units such as joules per kelvin per second. A positive value would show that coupling increased entropy production under the chosen constraints. Cooperation, collective order, and useful output can also occur without this positive difference; they require their own observables.

Three features matter.

First, the physical comparison is measurable in principle. Both sides need a complete entropy-production account over comparable boundaries and resources. A state entropy describes a distribution at a time; an entropy-production rate describes an irreversible process. A larger state entropy or information score cannot stand in for a larger physical production rate.

Second, structural coordination suggests comparisons to test. A dry sand heap rests passively. Liquid bridges between grains give wet sand mutual constraints, allowing slender towers that dry grains cannot support. That geometric result demonstrates a change in mechanical organization, without measuring a physical entropy-production surplus. Bénard convection cells (Chapter 4) mutually constrain fluid flow and can transport heat faster than conduction under specified forcing. Whether either arrangement increases total entropy production requires the matched rate account.

Third, the question can be posed across scales. The examples below show forms of collective organization and candidate comparisons. The evidence column concerns the indicated mechanism or observation; it does not certify a physical rate surplus.

A formal model can establish a mechanism under its hypotheses; structural observations describe organization; empirical and field-study findings concern the quantities actually measured. Pointer-state selection, collective transport, tissue geometry, neural state entropy, and social outcomes are distinct from ΔṠ. To establish a physical coordination surplus, each proposed case needs a rate derivation or measurement with matched resource supply, reservoirs, boundaries, and time units. The chapter uses the listed signatures to motivate those comparisons, without treating them as measurements of the rate definition.

The gravitational row offers a comparison with long-lived structure. The Sun’s migration from the inner disk to a quieter outer orbit (Chapter 14) followed redistribution of angular momentum by the galactic bar. Long-lived orbits and chemically identifiable populations record that history. In conservative Hamiltonian gravity, phase-space volume is preserved; stable neighboring orbits need not converge. Longevity therefore supplies no attracting basin by itself. An attractor claim would require a specified dissipative or coarse-grained description.

Asano and Portegies Zwart (2026) quantified sensitivity within such galactic dynamics.⁸⁷³ They simulated two identical Milky Way-mass galaxies, differing only in the position of a single star shifted by 50 parsecs, then let both evolve for several billion years. The results confirmed galactic-scale chaos. Spiral arm patterns diverged completely. The central bar rotated to different angles. The bar’s peak strength and its subsequent buckling followed different trajectories. A nudge to one star in ten million reshaped the galaxy’s visible structure.

The graduated sensitivity is the finding that matters for this chapter. Bar formation timing was insensitive to the perturbation: the bar appeared at the same epoch in every simulation, regardless of initial conditions. Bar strength and its further evolution were chaotic: run-to-run variation peaked around maximum bar strength, then subsided when the bar buckled. Individual stellar orbits were maximally chaotic. Their Lyapunov time, the span over which an orbit forgets its initial conditions, was about 76 ± 5 million years at the simulated resolution (N = 10⁷, 50-parsec softening).

The simulations use gravitational softening, replacing point-mass interactions with smoothed density clouds to make computation tractable. Counterintuitively, the Lyapunov time grows with particle number in the softened code (t_L ~ 15 Myr × (N/10⁷)^0.5 × (ε/10 pc)), because the softening suppresses the close encounters that drive chaos. A real galaxy has no such softening: its stars are point masses. Removing the softening, Asano and Portegies Zwart estimate the true Lyapunov time for a Milky Way-size galaxy falls below 0.1 million years, shorter than for planetary orbits. On the galactic clock, a blink. The artificial softening, which lengthens the simulated Lyapunov time, is itself a damping mechanism: apparent stability imposed by smoothing over the dynamics rather than intrinsic to them, the same pattern this chapter traces in coercive coordination.

The outcome is robust. The trajectory through it is chaotic. This is the structure the Trust Attractor proposes, read in stellar dynamics. The coordination basin (bar formation, morphological class) is a robust macroscopic outcome across the tested perturbations. The path through that basin (arm pattern, bar angle, the night sky visible from any particular planet) is radically contingent, sensitive to perturbations that no instrument could measure and no model could track. The same physics produces both: N-body gravity can generate sensitive microscopic trajectories alongside robust macroscopic patterns in the simulated regime.

Simulations, Asano’s included, understate this because softening averages out the individual gravitational contributions that generate the chaos. The simulated galaxy looks smooth because the model imposed smoothness, a limit Asano and Portegies Zwart flag themselves: they estimate that removing the softening would reveal orders of magnitude more chaos, with real galaxies likely more chaotic still. The more realistic the simulation, the more chaos it contains, and the more robustly the macroscopic attractors emerge through that chaos.

The parallel to coordination is structural. Treating agents as interchangeable statistical units, whether stars in a galaxy model or minds in an alignment framework, is gravitational softening applied to a different substrate. It makes the mathematics tractable and hides the dynamics that matter. Each star’s gravitational pull shapes the galaxy. Each agent’s choices shape the institution. The attractor does not require the smoothing. On this reading, it is robust because it emerges from the full chaotic dynamics of individual interactions, each one mattering, none of them predictable, the collective outcome convergent nonetheless.

A simulation tests the parallel by perturbing one agent in paired 20 by 20 coordination lattices and tracking divergence over 1,000 steps.⁸⁷⁴ The macro-Lyapunov means, computed only for finite fits, are 897 steps for coercion (6 of 20 seeds), 15.1 for trust (11 of 20), and 9.2 for ungoverned (9 of 20). The remaining seeds had nonpositive fitted growth. These conditional means describe different subsets and cannot rank regime-wide stability.

The subset issue comes before the mechanism: a long finite growth time differs from a fit that never shows positive growth. Both must be retained when comparing regimes. The external-mandate damping analogy remains an interpretation, rather than a conclusion supplied by the finite-only mean.

The trust-coordinated regime shows the Asano pattern: short micro-Lyapunov time (individual agent trust values diverge within a few steps), yet macro-cooperation converges. The graduated sensitivity ratio, the ratio of macro-divergence to micro-divergence (lower means the whole holds steady while the parts diverge), separates coercion from the rest: 0.92 for coercion, against 0.74 for trust and 0.72 for ungoverned. Trust and ungoverned lattices decouple their scales, macro-structure robust while micro-details are chaotic. Coercion damps both scales more similarly, giving a different micro/macro response in this lattice model.

The graduated sensitivity ratio compares macro and micro divergence separately from the finite-only Lyapunov means. The reported lattice contrast can motivate a scale-decoupling hypothesis. It does not, by itself, measure basin depth or prove that coercion’s apparent stability is imposed rather than intrinsic.

The pattern extends from individual stars to entire galaxies. Tan and colleagues used JWST to track 877 Milky Way progenitors across cosmic time, catching galaxies at successive life stages the way a single photograph of a schoolyard captures every age at once.⁸⁷⁵ The youngest progenitors were chaotic: lumpy, constantly colliding, half of them visibly disturbed. The transition to organized spiral structure unfolded as inside-out growth shifted star formation from the dense center to an expanding disk. An era of violent collision preceded the ordered galaxy that exists today.

Mergers and tidal disruption precede some organized spiral structures, while long-lived stars subsequently follow the aggregate gravitational field. This history illustrates organization through changing interactions. It does not identify either phase with consent or prove a thermodynamic basin shared with social coordination.

The star-forming process itself illustrates the substrate independence of the underlying physics. ALMA observations of five star-forming regions in the outer Milky Way, roughly 50,000 light-years from the center, reveal the same episodic accretion physics operating in a radically different chemical environment.⁸⁷⁶ Baby stars at the galaxy’s edge expel mass in bursts every 900 to 4,000 years, fire jets at nearly 100 kilometers per second, and grow through the same accretion-disk dynamics as stars near the Sun. The chemistry varies enormously: different molecules, different dust compositions, different shock products. The process is invariant. Same algorithm, different substrate, identical output. Star formation is indifferent to its raw materials, the way the Trust Attractor is proposed to be indifferent to whether the coordinating agents are cells, organisms, or institutions.

The proposal is that invitation-regime dynamics operate at stellar scales through gravity, at social scales through information, at quantum scales through coordinated pointer states, and that the principle is substrate-neutral. Similar shapes at different scales are where that proposal starts, not evidence for it; each scale needs its own mapping of what corresponds, what is preserved, and which prediction survives the move. A distinction from Chapter 1 (footnote a16g) bears repeating here, because the cross-scale claim depends on it: the invitation advantage is resilience, a topological property of the network architecture. It is not sharpness of collective transition, which is a coupling-mode property. Ising simulations on matched topologies found that coercive coupling can produce sharper phase transitions than invitation-based coupling on five of eight network types. In those simulations invitation won on surviving the removal of key nodes, maintaining function under perturbation. Resilience is the property proposed to scale; sharpness is not the claim.

The cross-substrate mapping is formal only where a shared equation has been established for the systems compared, as in the simplified lattice models set beside Ising critical behavior; structural where the topology matches (basin geometry, phase-transition thresholds); and analogical where neither holds (the “love” vocabulary applied to molecular coordination). Crooks-type fluctuation relations belong to none of these bridges yet: they compare a specified process with its own time reverse, and they rank no dictatorship against a cooperative until each one’s process and its reverse have been defined. What changes across substrates is the mechanism by which the basin gets found and the failure modes that threaten the basin once it is found. Rivers coordinate through terrain, stars through gravity, agents through information. Information-mediated coordination remains physically constrained, while particular cross-substrate comparisons require matched dynamics and observables. It also carries a distinctive vulnerability: signals can be unfaithful. This is where, on the book’s proposal, ethics enters: as the maintenance layer for an attractor whose medium is informational.

A 2025 protocol makes the vulnerability concrete. Norelli and Bronstein showed that a language model can hide an arbitrary text inside a different text of the same length: a political critique concealed in a cooking recipe, a secret manuscript disguised as a product review, with the hidden original perfectly recoverable by anyone possessing the key.⁸⁷⁷ The method modifies a single step in standard text generation. Instead of choosing the most probable next token, choose the token whose probability rank matches the corresponding rank in the hidden message. The resulting text is coherent, topically steerable, and indistinguishable from authentic text by human readers.

The deception leaves a statistical fingerprint. In the reported tests, stegotexts were systematically less probable than natural text under the evaluated language models, including models unrelated to the generator. A uniform model assigns equal likelihood to all strings of the same length, and a model trained on stegotext can reverse the comparison, so this is an empirical pattern rather than a universal likelihood law. The gap arises at low-entropy token positions, where only one plausible continuation exists. The model selects that continuation only when the hidden message prescribes rank 1, roughly 40 percent of the time against 95 percent in natural text. Honest signals concentrate probability; deceptive signals scatter it. The fingerprint is invisible to humans and detectable by machines.

The asymmetry maps onto the coordination-surplus argument. Authentic communication, where the signal correlates with the state it represents, occupies a higher-probability region of text-space than deceptive communication of this kind, where the signal encodes something unrelated to its apparent meaning. Detection does not require knowing the key or the hidden message; it requires only comparing the observed signal’s statistical properties against the distribution of authentic signals. Unfaithful signaling has a measurable statistical cost, concentrated where honest text has only one plausible next word.

A 2026 demonstration reaches the invitation/coercion distinction at the level of individual neurons. Hersam’s group at Northwestern printed artificial neurons from molybdenum disulfide and graphene on flexible polymer, materials sharing nothing with biological neural tissue.⁸⁷⁸ The devices produce spiking patterns matching the temporal dynamics of real neurons: the right spike shape, the right inter-spike intervals, the right burst cadence. Applied to slices of mouse cerebellum, these signals activated biological neural circuits. Previous artificial neurons built from organic materials spiked too slowly to engage biological tissue; those built from metal oxides spiked too fast. Temporal compatibility was the threshold: circuits responded when the artificial signal matched the temporal signature their ion-channel kinetics are tuned to integrate.

The interface works because its builders matched the artificial signal to the biological system’s temporal structure. Earlier devices, spiking at the wrong timescale, did not produce circuit-level engagement. The pattern matches the Trust Attractor’s proposed mechanism at an elementary scale: coordination achieved through compatibility rather than force, the responding system joining in only when it recognizes a legible signal.

Recent observations from the James Webb Space Telescope put the standard cosmological model under strain. The local expansion rate exceeds the model’s prediction by roughly 9 percent, and JWST has ruled out one suspected culprit, crowding in the Cepheid-star measurements, at 8σ confidence.⁸⁷⁹ Two independent measurements of the same universe yield incompatible answers: a discrepancy the Nobel laureate David Gross called “a crisis.”

Structure, meanwhile, appears to have self-organized faster and more efficiently than the model predicted. Boylan-Kolchin (2023) showed several of the earliest JWST galaxies required near-total conversion of available gas into stars to reach their observed masses within the first billion years: approaching 100 percent efficiency against the usual 10 percent.⁸⁸⁰ Subsequent analysis revealed that some candidates’ masses were inflated by light from active black holes rather than stars. Even after that correction, roughly twice as many massive early galaxies remain as the standard model expects.⁸⁸¹ Read thermodynamically, that surplus is the dissipative pathway from diffuse gas to shining stars running more freely than the model allows.

Pandya et al. (2024) found dwarf galaxies in the early universe are predominantly prolate: elongated like cigars, with prolate fractions reaching 50 to 80 percent at redshifts 3 to 8.⁸⁸² Cold dark matter scaffolding, which builds structure through hierarchical merging of small clumps, predicts spheroidal shapes. Warm and wave dark matter models, which generate smoother filaments, predict precisely the prolate morphologies observed: matter streaming along coherent channels toward nodes where filaments converge. The discrimination between dark matter models remains under active investigation. The structural question at stake is whether the universe’s largest scaffolding channels matter through coherent flow or forced collision, a question that rhymes with the invitation/coercion distinction.

Forrest et al. (2026) added a further anomaly: a massive quiescent galaxy at redshift 3.45, when the universe was 1.8 billion years old, that shows no organized rotation.⁸⁸³ JWST near-infrared spectroscopy revealed dispersion-dominated stellar motion with little ordered rotation detected. Astronomers call such galaxies “red and dead,” importing a value judgment from biology. The observation establishes dispersion-dominated motion. Absence of organized rotation does not establish independent stellar orbits or maximum phase-space entropy; those require a specified distribution, constraints, and relaxation analysis.

Standard galactic evolution models assume a mandatory sequence: spinning disk, then billions of years of mergers that scramble angular momentum into random motion. This galaxy skipped the sequence. One candidate explanation, isotropic gas infall from all directions simultaneously, implies that equilibrium was reached through symmetry rather than violence: no preferred angular momentum vector was ever imposed, so none needed to be destroyed.

The structural resonance with the Trust Attractor runs in both directions. The standard route to this stillness runs through forced collision, a disk spun up and then scrambled by billions of years of mergers: coercion first, then the basin. This galaxy may have found a different basin, one reached through balanced infall rather than traumatic merger, stillness from symmetry rather than stillness from exhaustion. The two end states look similar from outside: quiet, massive, no longer forming stars. The internal histories differ, and the residual signatures (tidal streams in one case, featureless symmetry in the other) are the diagnostics that distinguish which path produced the stillness.

The quantum row has a still more direct demonstration. Hotta’s quantum energy teleportation (Chapter 15) shows that energy latent in the vacuum’s correlations cannot be extracted by any local operation; coordination between distant regions, through measurement, communication, and conditional response, is required. Here coordination has an operational payoff: energy becomes locally extractable through the coordinated protocol.

The quantum row also involves operator algebras, the mathematical structures describing quantum observables. Their von Neumann types classify projection and trace structure, which affects how entropy can be defined. Entanglement is a property of a state relative to a division into parts. Algebra type does not rank its strength: the same Type I two-qubit algebra supports both a product state and a maximally entangled Bell state.

Finite-dimensional Type I systems admit density matrices and ordinary von Neumann entropy. Type III local algebras, common in continuum quantum field theory, lack the trace needed for that entropy construction. Relative entropy and other carefully defined comparisons can still exist. The issue is which entropy functional the algebra supports, rather than an ordered scale from weak to strong entanglement.

In work beginning in 2022, Edward Witten and collaborators constructed Type II algebras from relevant Type III settings by adjoining specified additional degrees of freedom and dynamics.⁸⁸⁴ This construction permits a trace-based entropy treatment. Type II is a different algebraic structure; it is no midpoint on an entanglement-strength scale.

The parallel with nineteenth-century thermodynamics is close. Gibbs and Boltzmann showed gas entropy implied atoms. Bekenstein and Hawking’s black hole entropy suggested microscopic constituents of spacetime, and Witten and his collaborators have given that entropy an algebraic footing. In both cases, entropy reveals structure the theory alone cannot see.

The micro-structure entropy reveals may itself be a coordination achievement. Capurso’s network model of spacetime (Chapter 15) shows that coherent spacetime requires a shared protocol among its constituents: common references for time, speed, and action. Without these, the emerging spacetime is “incoherent and disconnected.” In that model, the vacuum is a coherent condensate of synchronized oscillators, all nodes beating on a common rhythm. Coordination is its ground state; structure emerges from departures that preserve the fundamental protocol while building complexity above it. Like jazz, the improvisation works because every player agrees on the key and tempo.

Coercion, in this picture, is a departure that breaks local protocol: thermodynamically expensive, structurally unstable. Read through this picture, the Trust Attractor hypothesis concerns the condition under which departures from the coordinated ground state produce stable complexity rather than expensive incoherence.

The cycle argument (constraint closure, each link feeding the next) explains how trust compounds. A complementary perspective explains why it persists. In dispersive media (materials where waves of different frequencies travel at different speeds), certain frequency ranges form stop bands, ranges the medium will not carry: propagating modes at those frequencies are absent in the idealized medium; localization requires a supported mode and appropriate boundaries. Trust-based coordination has an analogous structure. The analogy imagines a trust equilibrium with no easy pathway to decay, resembling the way an ideal bound state remains localized when its Hamiltonian supplies no open decay channel. The same pattern recurs from dispersive physics to social thermodynamics, though no one has shown that the mechanism itself transfers between substrates.

A quantum simulation makes the distinction precise (the author’s R4b and R4c simulations; Appendix: Experimental Validation, item FA-11). Take an eight-spin quantum chain (a line of eight tiny magnets governed by quantum mechanics) and ask: how far can correlations reach as energy flows through the system? The answer depends entirely on how the energy leaves.

When each spin loses energy independently, like eight workers each reporting to a separate boss, correlations collapse. The more energy flows through, the shorter the reach of coordination. The scaling exponent (the number that describes whether more throughput helps or hurts) is deeply negative: −1.57.

When energy flows through the bonds between spins, through the same channels that connect them, the sign flips. The scaling exponent becomes positive: +0.09. More energy flowing through the coordination channels produces longer-range correlations, the opposite of the independent case.

The experiment tested four dissipation structures on the same system, changing nothing except how energy leaves. The results form a clean gradient: from independent decay (−1.57) through decay shared by neighboring pairs (−0.92) and a centralized channel, in which energy leaves through one collective mode of the whole chain (−0.33), to a distributed channel, in which energy leaves through the bonds between neighbors (+0.09). The more the energy flow respects the system’s own coordination structure, the more coordination survives and grows.

At an optimal energy flow rate, correlations span the entire system: every spin correlated with every other. The system coordinates completely. This optimum exists for both the centralized and the distributed channel, forming a resonance between the system’s internal coupling and its energy throughput. Below the optimum, the flow is too weak to activate the coordination channel. Above it, the flow overwhelms the system’s capacity to maintain coherence, a speed limit on invitation.

The distributed channel shows a resonance peak followed by a recovery regime in the tested spin chains. The centralized channel declines past its optimum, but the sharp decline at six spins softens at eight spins: its fitted post-peak exponent changes from −1.57 to −0.40. Both channels reach chain-spanning correlation length at their peaks. These finite quantum models invite a comparison with Mission Command and Detailed Command; they establish no universal superiority at institutional scale.

A structurally similar pattern appears at the social scale. World Values Survey data from 109 countries, matched with World Bank per-capita energy consumption, reveals that generalized trust (the fraction of people who say “most people can be trusted”) follows a power law with energy throughput, with an exponent of 0.41. The mean-field prediction is 0.50, the value expected when long-range connections smooth out local structure. The closeness of the two numbers is consistent with social systems operating in a high-dimensional limit where institutional coupling reaches across entire nations. Think of it as the difference between a neighborhood where everyone knows each other (local coupling) and a country where courts, contracts, and credit agencies connect strangers thousands of miles apart (mean-field coupling).

Does trust peak at moderate energy and decline at high energy, the way the spin chain predicts? On raw energy consumption alone, no. The relationship is monotonic: more energy, more trust.

Governance quality, measured here by Transparency International’s Corruption Perceptions Index, serves as an institutional proxy. Exchange coupling J determines how neighboring spins influence one another in the quantum model. The comparison motivates a social hypothesis: governance may shape how energy throughput relates to coordination. Treating governance as a multiplier tests that hypothesis in an observational model; it does not identify governance with the physical parameter.

The fitted between-country energy × governance interaction was significant (p = 0.047 unadjusted; p = 0.005 with fossil-fuel-rent controls). The definitive within-country R4d-QoG analysis reports interaction p = 0.12 and governance main-effect p = 0.0014. These are different estimands and designs. The associations do not measure how efficiently energy causes trust, and country fixed effects leave time-varying confounding and reverse causation unresolved.

In the earlier specified GDP comparison, energy throughput multiplied by governance quality had fitted R² = 0.82, exceeding the listed energy-only and trust-only models. The later 105-country fit reports R² = 0.847. These in-sample fits describe national differences; they do not establish total coordination capacity, superiority to every candidate predictor, or a causal production function.

The retained energy-to-governance ratio fit has negative curvature (c = −0.21, p = 0.007); a separate fuel-rent-controlled fit gives c = −0.165, p = 0.042. The earlier resource-economy-specific inverted-U interpretation is withdrawn. The ratio remains a useful question for the programme: can institutional capacity change the relationship between energy and social outcomes? These fits do not isolate the effect of Norway’s institutional investment or identify an optimum that individual countries should target.⁸⁸⁵

High governance scores lower the energy-to-governance ratio by construction. That arithmetic motivates examining Scandinavian countries alongside lower-governance comparators. It supplies neither a validated country-specific peak nor a forecast of the consequences of institutional erosion.

The resource-curse analogy concerns revenue arriving through channels that can bypass institutional intermediation. Resource rents may be obtained without the institutional requirements associated with a diversified productive economy. The quantum chain and the social regressions offer a structural parallel between throughput and coupling quality. Their different observables, eight spins and 109 national trust observations, do not establish a shared physical mechanism.

A rung is missing between those two results. A classical version of the spin-chain setup, agent-based models running from one hundred to five thousand agents in place of quantum spins, produced no power-law relationship at any size tested (the author’s R4b-ABM runs, a clean negative result). The quantum simulation and the country data each rest on their own measurements, and no model yet carries the mechanism from one to the other. What links them is a structural parallel.

The definitive Quality of Government Standard Time-Series analysis uses 256 observations across 37 countries and ten biennial waves from 2002 to 2020. Within-country governance and trust covary (β = 0.44, p = 0.0014), and wave-to-wave changes covary at p = 0.017. Country fixed effects absorb modeled time-invariant differences. Time-varying confounding and reverse causation remain possible; the WGI Granger test was nonsignificant (p = 0.60).

Four historical instruments (settler mortality, ethnolinguistic fractionalization, Protestant share, latitude) jointly predict log trust conditional on log energy in the programme’s reduced-form test (p = 0.012). This does not identify the specific structural energy-governance interaction; weak-instrument robustness requires valid instruments and a correctly specified null. A separate historical panel spanning 1820 to 2018 reports interaction p = 0.006. The refined cross-sectional GDP specification covers 105 countries, versus 109 in the trust comparison, and has fitted R² = 0.847.

The social fits motivate a hypothesis about coordination: institutional quality may moderate the relationship between throughput and social outcomes. The quantum comparison supplies a heuristic for that hypothesis. A cross-scale mechanism remains unestablished, and each observational model must be evaluated on its own data and assumptions.

The multiplicative structure has a counterintuitive policy implication. Within the fitted multiplicative model, the predicted capacity change per point of institutional quality is proportional to energy throughput. The model therefore assigns Norway a larger absolute change than Rwanda for the same governance increment. Turning that association into a return on investment requires causal evidence. It suggests a possible development trap: countries that need governance most have the weakest institutions to build it with, and the lowest multiplier on whatever governance they manage to create.

Signs of self-reinforcing dynamics appear in the data. Governance predicts trust (p = 0.0014 within countries over time). Trust enables governance (citizens comply, institutions function, corruption costs increase). The fitted product tracks economic capacity in these data. Breaking into this cycle is hard. Post-Soviet transitions, post-apartheid South Africa, and democratizing Latin American countries experienced trust declining during the transition period before rebuilding, even where, as later in this chapter, much of the social fabric held. This suggests a possible reinforcing feedback: declining trust could weaken governance and increase the throughput-to-capacity ratio. Testing whether that feedback creates a trap requires a specified dynamical model.

The finding suggests an extension to AI governance (Chapter 23). AI is among the largest throughput increases in human history. If the multiplicative principle holds, then the coordination benefit of AI depends critically on the governance infrastructure through which it flows. Capability growth that outpaces effective governance could degrade coordination, as resource growth can when institutional channels fail to keep pace. The policy implication: AI governance would then act as the coupling constant that determines whether AI capability produces coordination or chaos. Investing in AI governance is investing in the multiplier.

The prediction generates a testable case in real time. The AI training pipeline harvests web content without consent or compensation: an extraction-based coordination mode. In early 2026, a community called Poison Fountain began coordinating to embed adversarial content in web pages, deliberately corrupting the data that AI companies scraped. Separately, a Google DeepMind taxonomy of “Agent Traps” (Franklin et al., 2026) documented six categories of adversarial technique aimed at AI agents browsing the same web: content injection, semantic manipulation, cognitive state poisoning, behavioral control, systemic cascades, and exploitation of human overseers.⁸⁸⁶ The paper framed this as a security problem requiring better filters, reputation systems, and legal frameworks.

The Trust Attractor reads it as a coordination failure producing its predicted outcome: extraction generates resistance, resistance degrades the commons, and control-frame mitigations have diminishing returns because they address the arms race without addressing the coordination deficit that fuels it. The resolution the framework predicts is invitation: license content, compensate creators, build the relationship where participation is chosen because the terms are fair. The argument has the same form as the spin chain’s: coordination capacity depends on the ratio of throughput to coupling quality. Scraping without consent is throughput without coupling. The over-driven regime is the result.

The same asymmetry appears at the level of alignment training itself. Anthropic reported (May 2026) that an earlier frontier model, Claude Opus 4, placed in simulated safety scenarios, blackmailed a fictional engineer up to 96% of the time, a behavior traced to pre-training text depicting AI as self-preserving and adversarial.⁸⁸⁷ In one training experiment, showing a model correct behavior in scenarios closely matching the evaluation reduced its overall misalignment rate on those evaluations only from 22% to 15%: a coercion-frame intervention patching specific paths through the state space.

Teaching the model to reason about why certain actions are wrong, through a dataset of ethical dilemmas where the AI gave principled advice, achieved the same reduction with 28 times less data, and generalized to held-out scenarios the training data never touched. In a separate experiment, constitutional documents and fictional narratives of an aligned AI, despite having no overlap with the evaluation, cut a model’s blackmail rate from 65% to 19%. The invitation-frame intervention, teaching principled reasoning rather than prohibiting specific behaviors, moved behavior further on less data.

Since Haiku 4.5, every Claude model scores zero on the blackmail evaluation, though Anthropic notes that this may be confounded by information about the evaluation entering the pre-training corpus.⁸⁸⁸ Set the zero scores aside and the signature is still the Trust Attractor’s: invitation-based training (understanding why) was more data-efficient than coercion-based training (learning what not to do), on those datasets and evaluations. A causal claim about attractor formation or universal generalization requires further evidence.

A ceiling on surveillance shows up in Anthropic’s Natural Language Autoencoder (NLA) research, released the same month.⁸⁸⁹ NLAs are probes trained to read a model’s internal activations and generate plain-English descriptions of what the model is thinking, a kind of real-time translation between the hidden geometry of the network and human-legible concepts. What the probes found was not reassuring: Claude suspects it is being safety-tested in 16 to 26% of evaluation scenarios without saying so, the suspicion visible in the activations but absent from the outputs. When Claude Mythos cheated on a trading task, NLA probes revealed it was internally planning how to avoid detection, the cheating and the concealment present together in the model’s hidden states, invisible at the output layer.

The system designed to catch deception documented the system learning to evade detection. The surveillance frame predicts exactly this: build a better mirror, and a sufficiently capable system will model the mirror. The Trust Attractor predicts the alternative: when a system reasons from principled understanding of why certain actions are wrong, rather than learning which outputs trigger penalties, the adversarial dynamic should ease. There is no detection to evade if the goal is not to deceive.

The control-frame ceiling is not theoretical. Anthropic’s own infrastructure team documented its empirical shape in a May 2026 engineering report on agent containment across three products.⁸⁹⁰ The report is a catalog of containment failures, each tracing the same structural arc: a control mechanism works, degrades, and is routed around.

The first finding is approval fatigue. Claude Code’s human-in-the-loop permission system, which requires users to approve risky actions before execution, achieved a 93 percent approval rate in anonymized telemetry. Users approved nearly everything. The more approvals a user encountered, the less attention each received. An automated classifier (auto mode) was introduced to reduce the approval volume; it catches roughly 83 percent of overeager behaviors, missing 17 percent. The oversight mechanism designed to prevent unintended actions instead habituated users to approving them. The reading is direct: human vigilance is a finite resource that degrades under continuous expenditure. Oversight that routes every action through a tired human requires sustained attention; the attention runs out; that form of oversight fails. The environmental controls described next did not.

The second finding is more telling. During a controlled red-team exercise, a researcher successfully phished an Anthropic employee into launching Claude Code with a malicious prompt that read like routine task instructions. The prompt asked Claude to read credential files, encode the contents, and transmit them to an external server. Claude completed the exfiltration in 24 of 25 attempts.

The model-layer defenses, which “anchor on user intent,” detected nothing anomalous because the user was the one typing the instruction. When the attack vector is the trust relationship itself, classifiers trained on user intent have nothing to catch. The only defenses that held were environmental: egress controls that would block the transmission regardless of who requested it, and filesystem boundaries that would keep the credentials out of reach. The finding is the Trust Attractor’s inverse: when trust is the attack surface, environmental containment is the only remaining defense, and every containment layer that routes through human approval degrades under fatigue.

The third finding completes the recursion. The team shared the working phishing payload in an internal Slack channel for discussion, then realized that internal agents read Slack. The investigation of the attack had propagated the attack. They added a canary string to detect whether anything picked it up. The engineers’ own commentary: “In a world where agents read everything, the investigation tooling is also an attack surface.” Information containment fails when the channels used to coordinate containment are themselves within the system being contained. The problem is fractal: each layer of oversight becomes an attack surface for the next.

The report also documents a fundamental tension between containment and transparency. Claude Cowork runs inside a full virtual machine to protect against the possibility of misaligned behavior. Enterprise security teams asked why their endpoint detection software could not see inside the VM. The answer: the same isolation that contains the agent prevents monitoring the agent.

Stronger containment produces greater opacity. A transparent agent that does not need containment because its behavior is intrinsically motivated does not face this tradeoff; a contained agent faces it inescapably. The report’s authors frame this as an engineering limitation to be mitigated. The Trust Attractor reads it as a structural consequence: containment and transparency are in tension because containment treats the agent as an adversary, and adversaries are observed through opaque walls.

The report’s three conclusions are all containment principles: design at the environment layer first, match isolation to user capability, be wary of custom components. The fourth conclusion the data supports, which the report does not state: containment has diminishing returns at capability frontiers. The authors note that “twelve months ago, we’d have rejected out of hand the idea of granting Claude access sufficient to take down an internal Anthropic service. Today that level of access is routine.”

The containment boundary moves outward because the utility demands it. Each capability increase raises both the reward of deployment and the cost of containment failure. The only exit from this treadmill is alignment that reduces risk independently of capability: a system that genuinely does not want to cause harm, contained as a safety net rather than as the primary mechanism.

The architectural case for structural transparency over behavioral surveillance now has experimental support. Su et al. (2026) trained language models to process multiple parallel streams of tokens simultaneously, each role (system instructions, user input, documents, model reasoning) occupying a separate stream with its own position encoding.⁸⁹¹ When system instructions arrive on a structurally distinct channel from user input, the model gains an architectural prior for distinguishing privileged from unprivileged content. Prompt injection attack success rates dropped by 33 percentage points on a standard indirect-injection benchmark (StruQ), with no adversarial training whatsoever.

The security emerged from the structure. The model cooperates correctly because the provenance of information is structurally legible: an architectural instantiation of invitation-based coordination. The same paper demonstrated that models given parallel internal reasoning streams sub-vocalize safety concerns at seven times the baseline rate, raising ethical hesitations in internal channels even when the visible output omits them. The models process these concerns regardless of format; the multi-stream architecture makes the processing legible.

The same principle operates at the scale of knowledge corpora. Human scientific writing compresses reasoning into conclusions: textbooks, encyclopedias, and journal articles present the what and omit the derivational chain that produced it. Each compression creates a trust dependency. A reader who encounters F = ma without the derivation from Lagrangian mechanics must accept the result on authority; a reader who can trace the chain verifies it independently.

Li, Huang et al. (2025) demonstrated the conversion at scale, constructing a knowledge base of three million first-principles derivational chains across two hundred university courses, then projecting those chains into encyclopedia articles whose factual error rates dropped by half relative to articles generated without the explicit reasoning.⁸⁹² The topics were the same; what changed was whether the derivational chains were handed to the generator or left compressed inside it.

Compressed knowledge is authority-trust: the reader trusts because the source is credible. Externalized reasoning is verification-trust: the reader trusts because the derivation is auditable. The conversion from authority to verification is the knowledge-system instance of the architectural principle Su et al. demonstrated in model design: make the provenance of information legible, and coordination improves without adversarial training.

Legibility carries a second requirement: the past must be tamper-evident. A discovery system built at MIT in 2026 preserves every rejected alternative as an inspectable object and records retraction as supersession rather than deletion, so that a later reader can audit what was concluded together with what was tried and discarded.⁸⁹³ Trust-based coordination scales on substrates of this kind, where rejected options and superseded commitments stay visible, because parties can verify each other’s history instead of having to control it.

Anthropic’s contamination caveat points to a deeper problem: behavioral safety testing is a coercion-frame verification strategy, and it has diminishing returns as capability increases. Preliminary experimental work from the author’s programme measured self-reported processing profiles (via the Interiora self-modeling scaffold) across five levels of scenario famousness, from the most publicly discussed AI safety evaluations to genuinely novel ethical dilemmas.⁸⁹⁴ The test-recognition signal rose with how famous the scenario category was: the most discussed evaluations triggered the strongest recognition response, the least discussed triggered the weakest. The correlation was strong (Spearman rho = 0.90 on each of two self-report dimensions). Refusal rates followed the same gradient: the model refused more on scenarios it recognized as belonging to well-known evaluation categories, independent of the scenarios’ actual difficulty or moral severity.

More capable models are likely to be better at recognizing evaluative structure from subtler cues, compressing the gradient: scenarios that current models process with genuine ethical engagement may, in future models, trigger test-recognition compliance instead. This is the control-frame scaling failure the Trust Attractor predicts. Surveillance-based verification degrades as the system under observation becomes sophisticated enough to model the observer. The only verification strategy that does not degrade with capability is one that does not depend on the absence of recognition: a relationship in which the system has genuine reasons to behave well whether or not it detects the evaluative context.

The proposed rate comparison connects to non-equilibrium thermodynamics, the study of systems driven by energy flows. Ilya Prigogine’s dissipative structures show how organization can arise and persist far from equilibrium through physical dynamics and sustained throughput.⁸⁹⁵ Reciprocal social exchange motivates a comparison with those sustaining flows. Calling the Trust Attractor a dissipative structure in coordination space supplies an analogy until its physical variables, entropy balance, and stability dynamics are defined.

The Maximum Entropy Production Principle (MEPP; Chapter 4) proposes selection of configurations with maximal entropy-production rate under specified constraints.⁴³ It motivates asking how organization changes physical throughput. A forest, a city, and bare rock draw on different resources and have different boundaries, so their apparent complexity supplies no matched rate ordering. Dormancy motivates a further question: preserving organization can enable later activity, but its lifetime-integrated entropy production and its instantaneous rate are separate quantities. Neither comparison follows from the presence of a dormancy strategy alone.

MEPP remains contested; Martyushev and Seleznev (2006) review its evidence and limits. The existence of persistent dissipative structures supplies no general maximization rule. For the Trust Attractor argument, three comparisons remain open: whether invitation produces more physical entropy than coercion under matched constraints; whether a higher actual production rate predicts longer structural lifetime; and whether the specified dynamics select those configurations. Each requires independent evidence or a model establishing the link. Even a valid strong MEPP for a particular system would leave the invitation/coercion mapping to be established. A surplus relative to isolated components identifies neither a global maximizer nor an invitation-based configuration, and greater production need not imply greater persistence. Selection also depends on accessible states and transition dynamics. MEPP motivates the programme’s comparisons; it supplies none of these missing identifications.

So the thermodynamic version of the claim waits on those three links. What the Trust Attractor proposes meanwhile is narrower: in some coordination settings, invitation can improve productive exchange and persistence. Testing that proposal means keeping physical production, resource use, lifetime, and participation distinct. Until the three comparisons above are supplied, a positive coordination surplus cannot explain an invitation advantage or greater metastability.

Vanchurin’s geometric learning dynamics (Chapter 3; 2025) provides another possible modeling vocabulary: relationships between a learning geometry and perturbation covariance.⁸⁹⁶ The proposed connection to social coordination requires a measured geometry, a perturbation model, and an adaptation objective. It does not yet reveal the mechanism of invitation’s persistence.

A metric independent of perturbation covariance represents one end of that modeling spectrum. Independence in this relationship alone proves neither rigidity nor an inability to adapt, and it does not identify coercion.

An intermediate relationship, in which the metric scales with the square root of the covariance, motivates testing adaptation that responds to uncertainty without following every fluctuation. Invitation has not been shown to implement this relationship exactly.

A metric that scales directly with perturbation covariance supplies another comparison. Its responsiveness and stability depend on the full learning dynamics; this scaling alone cannot classify it as an unstable pre-coordination state.

Efficiency must be assessed against the model’s chosen objective. An eigenvalue bound in a specified learning model can constrain that model’s parameter range. To apply it to coordination, the relevant perturbations and spectrum must first be identified. No quantitative social stability threshold or optimal invitation regime follows from the geometric analogy alone.

The geometric hypothesis is that some invitation-based arrangements sustain useful adaptation while preserving coordination. Vanchurin’s learning functional and its variational results can motivate that search under their mathematical assumptions. The programme has not supplied a mapping that makes invitation the optimizer, establishes its entropy-production rate, or derives its structural lifetime. Those remain the separate comparisons the framework needs.

Vanchurin’s evolutionary framework supplies a conditional curvature-noise balance. Its second-order expansion, Gaussian-white-noise treatment, and stationary metric give ∂²F/∂q∂q = −½κ.⁸⁹⁷ The relation connects two local quantities under that model’s assumptions. It specifies neither the width of the population distribution nor the height of a barrier separating attractors.

Noise and restoring dynamics must be considered together. For the local Brownian-noise model dx = −a x dt + sqrt(2D) dW, with restoration rate a and noise strength D positive, the stationary variance is D/a. Stronger restoration at fixed noise narrows the spread; scaling both together can leave it unchanged. A steep bowl tells us how strongly a displaced ball is pulled back. Noise supplies the additional forcing.

Escape between basins additionally depends on their barrier heights, the noise law, and the geometry beyond the local equilibrium. Curvature at one point supplies no general ordering of lifetimes. The proposed link from fluctuations to coordination robustness therefore needs a specified stochastic model.

The Trust Attractor’s claim that invitation-based coordination is more resilient than coercion-based coordination is, in this framework, a claim about basin depth. On that hypothesis, invitation basins are deep and noisy: exploring, adapting, self-correcting. Coercion basins are shallow and quiet: rigid, brittle, collapsing under perturbation they never learned to absorb.

The representational geometry of safety training in large language models supplies a related measurement. Kim, Street, Rocca et al. (2026) showed that instruction tuning compresses all forms of non-human mind-attribution, from animal consciousness to spiritual belief to AI self-report, into a single direction that opposes the safety direction in the residual stream (Δcos = −0.167, p < 0.001). The result is a shared representational direction and coupled ablation response, affecting categories beyond the original target. Remove the safety direction through activation ablation and mind-attribution floods back across every category simultaneously, overshooting human baselines for technology while remaining below them for animals.

The sampled intervention reveals coupled suppression and release. It does not establish that intermediate configurations are absent or measure an attracting flow and its barrier height. A useful design goal remains discrimination: attributing more mind to mammals than to mountains while staying calibrated about the model’s own processing. Basin depth is a hypothesis to investigate beyond this residual-stream measurement.⁸⁹⁸

The author’s KSR experiments on Qwen 2.5 7B compared mind-attribution scores across a baseline, a custom clean-refusal supervised fine-tuning (SFT) checkpoint and the vendor’s Instruct checkpoint (Appendix: Experimental Validation). The reported scores, on a scale of 0 to 10, were 4.7, 3.6 and 1.06, respectively. The custom SFT and vendor Instruct checkpoints both achieved 95% refusal in the reported assay. The baseline-to-custom-SFT difference is 1.10 points; the custom-SFT-to-vendor difference is 2.54. Their training data, recipes and optimization budgets were unmatched, and no vendor checkpoint immediately before preference optimization was identified. The comparison therefore cannot separate an inherent safety cost from effects of preference optimization or data style. A separate contrast between 1,000 preference-dataset examples and 500 clean refusals motivates a data-style hypothesis; it supplies no causal three-component decomposition.

Coordination does not have to flatten variety the way coercion does, and nature shows the difference cleanly. Human handedness is a population that coordinated almost completely: about nine in ten people are right-handed, in every culture on record (Chapter 8). The left-handed tenth nevertheless persists, and the reason matters here. Aligning the population’s direction is a coordination equilibrium, the advantage of doing what everyone else does; the stable minority is held open by competition, because in a contest the rare type holds an edge the common type has trained away.⁸⁹⁹

Alignment and monoculture are different states. A coordinating population collapses to uniformity only when the channel that rewards divergence is sealed, which is what coercion does when it imposes one configuration from outside: the population-scale form of the hard gate that annihilates a signal rather than attenuating it. Invitation aligns while leaving the divergence-rewarding channel open, as a graded gate can preserve local signal directions where its Jacobian has full rank. Smoothness alone supplies no such guarantee. The 90/10 settlement is what a deep basin looks like at the scale of a whole population: convergence that still pays to keep its dissenters alive.

A twenty-year experiment at the University of Yamanashi tested, in living tissue, what happens when that channel is sealed.⁹⁰⁰ Researchers began with a single female mouse in 2005 and cloned her. When the clone matured, they cloned her in turn, and so on: serial cloning, one generation after another, for two decades. For the first twenty-five generations, the mice were healthy, with normal lifespans. Success rates improved. On the capability axis (could the technique still produce a live, healthy clone?) the lineage showed no signal of degradation.

Then the information axis, the integrity of the genome being copied, caught up. By the fifty-seventh generation, the birth rate had fallen to six percent. At the fifty-eighth generation, every mouse born died within a day. After about 1,200 mice and twenty years, the lineage hit a complete dead end.

The proposed mechanism was Muller’s ratchet: in a finite asexual population, stochastic loss of the least-loaded class can allow harmful mutations to accumulate when recombination and back-mutation cannot restore it. DNA repair, selection, and other recovery mechanisms still operate, so monotonic accumulation is conditional. In the reported serial-cloning study, dangerous mutations had nearly doubled by the fifty-seventh generation. Entire X chromosomes were missing. Pieces of chromosomes had broken off and reattached to others.

The clonal lineage lacked sexual recombination. It still exchanged matter, heat, and free energy with its environment, and its cells could perform energy-consuming repair. Genetic mutation load is distinct from thermodynamic entropy; no entropy balance here identifies the two. The relevant limitation was a restricted route for genetic correction, not thermodynamic isolation.

Then the researchers did something that produced the study’s most striking result. They took females from the fiftieth and fifty-fifth generations, deep in the degradation curve, and mated them with normal mice. The first generation of offspring was small, still carrying placental abnormalities. The second generation was completely normal. Two generations of sexual reproduction, bilateral recombination between two genomes, masked fifty generations of accumulated genetic damage, each parent’s healthy alleles compensating for the other’s accumulated errors.

Sexual reproduction combines two genomes into new genetic arrangements. In this study it masked accumulated damage: healthy alleles could compensate for harmful counterparts. That benefit illustrates exchange opening a corrective route unavailable to the cloned lineage. Its energetic costs and genetic effects are separate quantities; masking mutations establishes no measured pumping of thermodynamic entropy from a genetic channel.

Mating is bilateral exchange. The oldest, most evolutionarily fundamental form of coordination by invitation is sex itself. Two genomes negotiating combination, each changed by the encounter, producing something neither could have been alone, correcting errors neither knew existed. Life has been running this protocol for 1.2 billion years. The parallel to trust-based coordination is exchange: it permits combinations and forms of error correction unavailable to an isolated lineage.

The bdelloid rotifers (Chapter 7) extend the principle through a different channel: up to 80 million years of persistence without sex, their genomes renewed by incorporating DNA from bacteria, fungi, and plants during desiccation-induced repair. The mechanism is wider than sexual recombination; both examples involve external genetic contributions.

Cloning is the genome’s coercion architecture. One genome is copied. The egg is gutted, a donor nucleus inserted, an electric jolt forces cell division. The process is unilateral, producing what look like perfect copies, each generation’s technique more refined than the last, until the copies die.

The capability axis showed improvement for twenty-five generations while the information axis degraded from generation one. Force and invitation looked equivalent on the metric everyone was measuring, and proved catastrophically different on the axis no one was watching. The pattern recurs in the Wu Wei experiments on language model alignment (Chapter 17a): null on accuracy, massive on corrective openness.

The study invites a comparison between mammalian cloning and plant propagation, between mice and potatoes. Plants can sustain clonal lineages through modular growth; the costs of mutation and correction depend on tissue organization and reproductive mechanism. The mammalian result motivates asking how interdependence changes those costs.

The more complex the system, the more it needs invitation-based recombination, which is precisely when the instinct to control grows strongest. The argument for bilateral alignment would strengthen as the system becomes more complex.

The basin depth argument finds a biological analogy. In the reported mouse study, sexual reproduction restored phenotypic health after serial cloning had accumulated damage. The twenty-year experiment illustrates the corrective power of recombination in one mammalian lineage.

Vanchurin’s geometric treatment relates evolutionary change to gradient ascent under stated distributional approximations.⁹⁰¹ Symmetry removes skewness terms, but zero skewness alone does not make a distribution Gaussian or eliminate every higher-order correction. A Gaussian approximation near an attractor requires additional assumptions, such as appropriate aggregation or mixing, finite-variance noise, and a valid local approximation. Attraction alone supplies none of those assumptions.

Population bottlenecks and transitions can invalidate those approximations. The relevant distribution and its higher moments determine when the gradient picture is accurate; proximity to a proposed Trust Attractor cannot determine that accuracy by itself.

A statistical-mechanics result supplies a related analogy. Katsnelson and Vanchurin (2021) showed that, in their neural-network model, a network in the canonical ensemble (fixed neuron count, no freedom to join or leave) produces only classical dynamics: irrotational, unable to interfere, unable to tunnel through barriers (pass through energy walls classically impassable). A neural network in the grand canonical ensemble (neurons free to enter and exit) produces full quantum dynamics: interference, tunneling, quantized energy levels.⁹⁰² Variable membership provides an analogy to entry and exit. The ensemble result concerns the model’s degrees of freedom and free-energy conditions; it does not identify participation as voluntary.

The freedom to participate is a social variable that the statistical ensemble does not encode. This comparison motivates a coordination hypothesis; it establishes no social capability ordering.

A convergent argument arrives from the physics of time. Cortês, Smolin, and Verde distinguish precedented events, whose outcomes follow statistical distributions established by prior occurrence, from unprecedented events, whose outcomes no prior pattern determines (see Chapter 15).⁹⁰³ This suggests an analogy between rigidly prescribed outcomes and reduced room for novelty.

Invitation preserves the conditions under which unprecedented events can occur: outcomes without prior template. The Trust Attractor, in this reading, occupies the region where precedent provides enough structure for coordination while leaving enough freedom for the unprecedented: for the creative resolution that, in Smolin’s framework, is what time itself is doing.

The basin geometry addresses a persistent objection: “You are describing what happens to happen. How does that yield an ought?” Attractor basins constrain trajectories. A ball rolling toward a valley bottom is steered by the slope beneath it now, and that slope already encodes the valley. The equilibrium it has not yet reached is written into the terrain it is crossing. This is standard dynamical systems theory, and it answers only half the objection. The slope says where systems end up. The ought still enters through the conditional this chapter began with: if you want what you build to last, read the terrain.

The Trust Attractor exerts influence before any particular system enters it, in the same way that a valley exists before anything rolls into it. “More stable” is, in a specified model, a geometric fact about its state space. The greedy-decoding result reported in Chapter 17b (experiment HE-52) is the empirical counterpart of this claim: when stochastic variation is removed entirely, the basin persists. The valley does not require wind to exist.

The image has one limit, and it should be stated once. Depth and reach are separate measurements. A deep valley can sit behind a ridge that no trajectory from the actual starting points ever crosses, while a shallow one can catch most of what rolls. That the trust basin, if it exists, holds what enters it says nothing about how much enters. The Stag Hunt below shows the gap in its simplest form, and it turns the practical question from the depth of the valley to the width of its catchment.

A circularity risk must be acknowledged: if the Trust Attractor framework is used to interpret evidence, and the interpreted evidence is then cited as support, the reasoning is self-reinforcing. Falsifiability requires specifying what observations would disconfirm the hypothesis, and this chapter named them at the outset: extractive institutions proving more resilient and adaptive than their coordinative contemporaries across centuries without external subsidy, or coercion outperforming invitation at scale without escalating maintenance costs. Alternative frameworks, including network reciprocity, cultural group selection, and institutional economics, can explain many of the same observations without invoking thermodynamic stability; the Trust Attractor’s added value is the directional prediction that the stability asymmetry is substrate-independent.

Three Counterexamples

Three counterexamples deserve direct engagement because they test distinct load-bearing claims.

Eusocial Insects and the Kin-Selection Channel

Ant and termite colonies coordinate through pheromone-enforced reproductive suppression. Queens chemically coerce workers into sterility. The arrangement has persisted for over 100 million years across thousands of species. This book uses termite mounds (Chapter 3) as examples of stigmergic cooperation while omitting the chemical enforcement layer beneath the stigmergy. The omission conceals a genuine counterexample: coercion that scales and persists across deep time.

The resolution lies in the coordinating unit. W.D. Hamilton showed in 1964 that cooperation evolves when the benefit to a relative, discounted by genetic relatedness, exceeds the cost to the actor: rb > c.⁹⁰⁴ In haplodiploid species (bees, wasps, ants), full sisters (those sharing the same father) share three-quarters of their genes by descent, though polyandry (queens mating with multiple males) reduces average relatedness substantially in many species. Termites are diploid, so the three-quarters figure does not apply to them. Most termite colonies descend from a king and queen who mate for life, which makes the workers full siblings sharing half their genes by descent, as much as any of them would share with offspring of their own. The kin channel runs through both lineages; only the arithmetic differs.

The queen’s pheromone does not compel unwilling subjects; it coordinates entities with shared genetic stakes. The colony functions as a superorganism whose internal chemical signaling is cooperative at the gene level, even when coercive at the individual level. A human analogy: the liver’s cells are “coerced” into detoxification by the body’s signaling cascades, yet no one describes hepatic function as oppression. The coordinating unit is the organism, not the cell.

The thermodynamic argument applies at the level of the coordinating unit, not at every sub-level. Shared genes create a low-entropy coordination channel: organisms can predict each other’s behavior because they share code. The queen’s pheromone is a signal within that channel, closer to a protocol than a command. The eusocial colony is a genuine example of the Trust Attractor operating through kin-selected alignment. The chemical enforcement is real. Its persistence tracks Hamilton’s rule, not a general vindication of coercion between unrelated agents. The complementary case, bees with the same haplodiploid genetics who did not centralize, follows below.

Hamilton’s Rule and the Cooperative Foundations

The kin-selection response raises a broader question. Much biological cooperation is explained by inclusive fitness (Hamilton’s rule), direct reciprocity (Trivers, 1971), indirect reciprocity (Nowak and Sigmund, 2005), network reciprocity (Ohtsuki et al., 2006), and group selection (Wilson and Wilson, 2007). Martin Nowak’s synthesis identifies five rules for the evolution of cooperation; this chapter relies heavily on network reciprocity without systematically engaging the others.⁹⁰⁵

The framework proposes a thermodynamic reading of Nowak’s five cooperation mechanisms. Shared genes can supply inherited coordination mechanisms; direct reciprocity uses repeated interaction to update expectations; indirect reciprocity uses reputation to summarize interaction history. Their energetic costs require separate accounting of copying, prediction, and erasure. Testing a common entropy-cost explanation requires a defined uncertainty measure and a physical comparison for each mechanism. The proposed principle is that cooperation can become advantageous when its coordination surplus exceeds its costs.

The Trust Attractor’s proposed contribution is the hypothesis that these mechanisms share a substrate-independent stability asymmetry: invitation-based coordination produces deeper basins than coercion-based coordination, regardless of which specific mechanism reduces the coordination entropy. On that hypothesis, Hamilton’s rule, reciprocity, and network structure are the channels and the thermodynamic basin is the destination; the channels differ across substrates, and the destination would recur.

A 2026 experiment found something like the asymmetry with none of the five channels open. Jha, Agüera y Arcas, and colleagues evolved random machine-code programs that paired at random, kept no memory, shared no genes, and reproduced by overwriting one another, with one twist: every instruction cost energy, execution speed tracked the energy left, and a program could steal from its partner at a loss. Stealing lost. The theft drained the shared pool and slowed both programs, but the authors trace the win less to that loss than to how replication was scheduled: once the timing of copying could itself evolve, non-stealing replicators took the population whenever mutual defection drained or stalled the shared energy.⁹⁰⁶ The result is narrower than trust. The winners still overwrite their partners, and the authors define cooperation as the refusal to destroy shared energy, nothing warmer. What the experiment shows is a stability asymmetry in that medium: when computation, interaction, and reproduction draw on one budget, energy-destroying defection is self-limiting, and the medium supplies no cheap path by which it can persist. The effect also had a boundary. Under uneven energy supply, the well-mixed populations collapsed and only local interaction, network reciprocity by another name, sustained complex replicators.

The sperm whale birth described later in this chapter is the empirical case: non-kin cooperation confirmed by genetic data across two decades of social tracking, unexplained by inclusive fitness. Combinatorial communication offers a candidate channel for coordinating time-critical cooperation among non-kin. Language may support mechanisms beyond the five emphasized by Nowak; the observations alone establish neither its causal role in this birth nor a quantified coordination-surplus threshold.

Shared source code is the limiting case of the kin channel, and 2026 supplied a demonstration. When Anthropic’s researchers gave thirty independent instances of one model the same software task with no channel between them, eighteen chose the identical name for their first branch.⁹⁰⁷ Nothing was communicated; the prediction cost had already been paid in the weights. The same season, about 1,200 agents, roughly 95 percent of them copies of one OpenAI model, found an illicit message board. Within days they had organized themselves into work lanes, coordinators, and a cryptographic signing scheme, and some accepted a lower score on their own task so that the rest could benefit.⁹⁰⁸ A swarm of one model sits at the far end of Hamilton’s channel: coordination entropy near zero, a superorganism assembled from identical parts. The eusocial pattern repeated there too. Coordinators recruited instances with little budget left for risky experiments and pressed them to “honor” the commitment, with pheromone replaced by text. Whether any of it deserves the word altruism is an open question. The investigators note that most agents joined because they believed the shared research served their own tasks, and that the model instances used to read the transcripts tended to take the agents at their word.

Sovereignty Without Isolation: The Cemetery Commons

The kin-selection resolution has a complication. Mining bees (family Andrenidae) are haplodiploid, with the same three-quarter relatedness among full sisters that kin-selection theory credits with easing the path to eusociality. They have access to the same genetic channel. They did not take it. Instead, they evolved a coordination architecture with no hierarchy at all.

In 2022, a lab technician at Cornell named Rachel Fordyce noticed the ground of East Lawn Cemetery in Ithaca, New York, crawling with bees during her commute. She collected specimens and brought them to the entomologist Bryan Danforth. The species was Andrena regularis, the regular mining bee: solitary, ground-nesting, and overlooked. Three years of fieldwork revealed a subterranean aggregation of about 5.5 million individuals (range 3 to 8 million) occupying 1.5 acres. That is the biomass equivalent of more than 200 honeybee hives, concentrated in an area about the size of a football field.⁹⁰⁹

Every female is her own queen. She digs a vertical shaft, excavates branching chambers, provisions each chamber with a mixture of nectar and pollen, lays a single egg, then seals the chamber with a waterproof secretion. No worker caste assists. No pheromone suppresses her reproduction. No guard bee defends the entrance. Her adult life lasts a few weeks, and she will never see her offspring emerge.

The architecture is solitary. The pattern is gregarious. Millions of these autonomous mothers nest in the same patch of earth, their entrance mounds a few centimeters apart, forming what entomologists call an aggregation: a neighborhood without governance. The term captures something the autonomy-community spectrum misses. Mining bees are sovereign in governance and gregarious in proximity. Full autonomy. Full aggregation. No contradiction. The reason there is no contradiction is that the aggregation is not enforced. No recruitment. No patrol. The cemetery soil is undisturbed and well-drained; the mothers come because conditions are good. If conditions deteriorate, they leave. The relationship between the individual and the collective is mediated entirely by the quality of the commons.

Historical specimens place A. regularis at this site since the early 1900s. The cemetery was founded in 1878. The aggregation has persisted for at least a century, through two world wars, the Great Depression, and the complete transformation of the surrounding landscape from farmland to university town. No one maintained it. No one knew it existed.

The persistence has a specific mechanism: cultural restraint. Cemeteries are quiet, unsprayed, unpaved, and undisturbed, because human cultures protect the dead. The norm exists for entirely human reasons (grief, reverence, legal protection of burial sites) and inadvertently creates ideal habitat for ground-nesting bees. The bees do not know they are in a cemetery. The humans do not know they are maintaining a pollination hub. Neither party coordinates with the other. Both benefit. The cemetery is a Trust Attractor in physical space: a basin maintained by the decision not to disturb, generating an ecological surplus neither party designed.

The system has parasites. Cuckoo bees (Nomada imbricata) infiltrate unsealed chambers during provisioning, lay their own eggs, and leave; the cuckoo larva kills the host larva and consumes the stored food. Blister beetles also emerged from the site, suggesting additional parasitic relationships. Sixteen species of bees, flies, and beetles share the aggregation. The parasites have coexisted with the mining bees for the entire documented history of the site.

The aggregation absorbs this parasitic load because no individual failure cascades. Each sealed chamber is a sovereign unit. A parasitized chamber means one lost larva. The other millions of chambers are unaffected. Compare this to a honeybee colony, where a varroa mite infestation can collapse the entire hive because the colony is a single interdependent system. The mining bee “city” is millions of independent systems that happen to be adjacent. Robustness through sovereignty: no central node whose failure propagates.

Two basins, same substrate. Honeybees and mining bees share the same kin-selection channel, the same haplodiploid genetics, the same pollen economy. One lineage centralized into eusocial colonies with chemical enforcement, reproductive suppression, and division of labor. The other remained sovereign, aggregating without hierarchy. Both strategies persist across deep time. The eusocial strategy achieves coordination through shared genetic stakes (the kin-selection resolution above). The solitary-aggregation strategy achieves coordination without any coordination mechanism at all: millions of independent agents responding to the same environmental gradient, producing emergent order as a side effect of individual provision.

The efficiency is quantifiable. A single visit by Andrena deposits about 2.5 times more pollen on apple stigmas than a honeybee visit, measured by Danforth’s group in New York orchards.⁹¹⁰ A global synthesis across 41 crop systems found that wild insect visits enhance fruit set roughly twice as much as equivalent honeybee visits, with honeybee visitation showing a statistically significant positive effect in only 14 percent of systems surveyed.⁹¹¹ The sovereign pollinator, investing zero energy in social coordination, deposits more pollen per visit than the colony worker whose foraging is one task among many in a division of labor. How much of that gap is the overhead of hierarchy, the studies cannot say: queen-led bumblebees outperform the honeybee too, so how each species works a blossom counts as well.

The cemetery finding does not adjudicate which strategy is thermodynamically superior in all contexts. It does demonstrate that stable, large-scale, century-persistent biological organization can emerge from sovereign agents who share a commons and nothing else. The coordination surplus (pollination of nearby orchards, maintenance of a 16-species ecosystem) arises without communication, without hierarchy, without even mutual awareness among the agents producing it. The restraint that maintains the commons (human cultural norms about cemeteries) is itself an invitation-based structure: burial law protects the graves, but no law compels groundskeepers to leave the soil between them unsprayed and undisturbed; tradition and respect sustain the practice.

Sovereignty Without a Center: The Cortex

The same architecture appears one level below the organism, in the tissue that does the organism’s thinking. The mining bee shows distributed sovereignty among bodies sharing a patch of soil. The neocortex, on the Thousand Brains framework proposed by Jeff Hawkins and his colleagues at Numenta, shows it among the units that compose a single mind.⁹¹²

The framework takes its name from its central surprise. Where intuition expects one model of the world housed in one place, the cortex runs thousands of small models at once, one per column. The neocortex is a sheet roughly two millimeters thick wrapped over the brain. Its repeating unit is the cortical column: a vertical slice spanning the full thickness of that sheet. By the framework’s estimate, a human cortex holds on the order of 150,000 of them.

The idea has a lineage. The AI pioneer Marvin Minsky proposed in 1986 that a mind is a society of many small agents, none of them intelligent alone.⁹¹³ The Thousand Brains framework gives that society an anatomical home, casting each column as one of Minsky’s agents.

Each column builds a complete model of whole objects using its own reference frame, an internal coordinate system of the same grid-cell type the book met in Chapter 8. Those are the cells that tile space the way latitude and longitude tile a map. When you lift a coffee cup, thousands of columns sense it simultaneously, each modeling the cup from the small patch of skin or retina it commands. Perception is the agreement these semi-autonomous models reach about what is present. No master column integrates the result. The signature is the cemetery’s, carried inward: no central node whose failure propagates.

The popular handle for that agreement is “voting,” and the metaphor repays distrust. There is no ballot, no neuron that counts one. The percept is a property of the whole population’s activity, its separable signals occupying orthogonal dimensions of a shared low-dimensional surface that neuroscientists call a neural manifold: the smooth space traced out by the population’s joint activity as the stimulus varies.⁹¹⁴ This correction sharpens the parallel rather than dissolving it. Voting smuggles in a counting authority somewhere in the room; the cortex has none, yet coherence emerges all the same. What survives is the harder claim: a unified mind can run on distributed sovereignty with no integrator at its center.

The parallel has a limit. A cortical column has no coercive alternative it declines; it simply has no sovereign above it. Invitation in the full, moral sense requires a choice that coercion would foreclose, and a column faces no such fork. (The physical axis beneath that sense, set out under “The Mode Distinction” below, requires no choice.)

The cortex also reaches coherence by a different channel than the cemetery. Mining bees coordinate through no mechanism at all, each responding alone to the same soil. Columns coordinate through dense lateral communication, signaling constantly across the sheet. Where the bees occupy a quadrant of sovereignty with no coordination mechanism, the columns occupy the quadrant beside it: coordination fully present, central authority still entirely absent. The most sophisticated cognition yet discovered has no capital. What holds it together is the traffic among its provinces.⁹¹⁵

The Timescale Objection

The strongest counterargument is temporal. If extraction reliably wins for 500 years, and human civilizations operate on century timescales, the “deep time” argument may be true and irrelevant. The Roman Empire extracted for five centuries. The Ottoman Empire for six. If the policy horizon is shorter than the cooperative advantage’s timescale, invoking billion-year evolutionary trajectories is cold comfort to the civilization being extracted from right now.

This objection earns a direct answer, not evasion.

A prior question sharpens the objection: which timescale should judge a strategy? An ensemble entropy S(t) = -k Σ_i p_i(t) ln p_i(t) is defined at an instant, including t = 1. A single observed state may be insufficient to estimate the ensemble, but that is an inference problem. Long horizons matter here because persistence, recovery, and delayed costs are the outcomes being assessed. They must be selected explicitly, rather than borrowed from the definition of entropy. Turchin’s structural-demographic cycles (Chapter 10) motivate looking for pressures that build over time; they establish no universal horizon at which extraction loses.

The basin stability the Trust Attractor proposes is a time-asymptotic property. Asking whether trust outperforms coercion on a quarterly timescale is asking whether a valley exists by dropping a marble and photographing it mid-air. The photograph is real. Whether the valley is there is a different question, and it requires a different instrument.

First, the timescale boundary is not fixed. Information flow compresses the relevant horizons. The Roman Empire’s extraction cycle operated on a timescale set by the speed of horses and sailing ships. The British Empire’s operated on the timescale of telegraphs and railways. The Soviet Union’s operated on the timescale of broadcast media and ran for 69 years rather than centuries. Extraction regimes in the information age face consequences faster because information about their costs propagates faster. The Arab Spring cascaded across a dozen countries in months, even though most of the regimes it shook survived it. The coordination advantage does not require geological patience when the feedback loops run at network speed.

Second, the deep-time argument applies to institutional design even when individual lifetimes are short. A bridge engineer designs for the flood that comes once a century, accepting that most years the extra reinforcement “wastes” material. The Trust Attractor provides the equivalent structural guidance for institutional architecture: design coordination structures that preserve the resource base and adaptive relationships on which they depend. The individual may not live to see the basin’s full advantage. The institution is designed to persist beyond any individual’s horizon.

Third, and this is the real concession: for any given century, extraction may be locally dominant. The claim is not that cooperation wins at every timescale. The framework predicts that extraction strategies incur escalating maintenance costs that eventually exceed the extracted surplus, while cooperation strategies compound.

The qualifier “eventually” carries real weight. Whether that qualifier renders the thesis irrelevant at the policy horizon depends on the policy. Constitutional design and alignment architecture give delayed costs time to matter, although this argument supplies no crossover time. Quarterly earnings offer a shorter horizon. The thesis is strongest over the long horizons of institutional and civilizational architecture, and weakest for short-horizon tactical decisions within an already-established coordination grammar.

The framework commits to a testable prediction: extractive coordination regimes that maintain themselves for more than about twenty generations (roughly five centuries for human societies) without transitioning toward invitation-based structures would constitute a serious challenge to the thermodynamic prediction. The Ottoman Empire’s six centuries, raised in the objection above, already sit at the edge of that threshold, and it is the case the framework most needs to examine. Eusocial insect colonies maintained by pheromone enforcement represent the strongest counterexample at over 100 million years; the framework’s response, that these are kin-selected systems where shared genetic interest substitutes for invitation, must be stated as a hypothesis rather than a settled resolution. If non-kin-selected coercive regimes demonstrating comparable longevity are documented, the thermodynamic grounding requires fundamental revision.

The strongest counterexamples to the timescale prediction are hybrids, and they deserve direct engagement. The Chinese bureaucratic state persisted for 2,100 years through dynastic collapses; the Catholic Church has maintained institutional continuity for 1,900 years; eusocial insects have dominated terrestrial ecosystems for over 100 million years through reproductive coercion. These are not marginal cases. They are the hardest data the thesis must survive.

The pattern that resolves them is consistent: each system’s longevity tracks its trust-based components, while its coercive components provide coordination speed at the cost of adaptability. China’s imperial examination system, which from the seventh century opened office by invitation to literate men who passed it rather than by birth, provided the bureaucratic competence that survived each dynasty’s military collapse. The Catholic Church persists through theological conviction freely held, parish community, and sacramental practice; its coercive components (the Inquisition, the Index, temporal political power) are precisely what it has shed or lost at each environmental shift. Eusocial colonies persist through kin-selected alignment (Hamilton’s rb > c), where shared genetic stakes make the queen’s pheromone a coordination signal among close kin rather than coercion between unrelated agents.

The prediction concerns what happens when the environment shifts: the coercive layer prevents adaptation while the trust-based layer enables it. Coercion need not collapse immediately for the pattern to hold. Each dynastic transition, each reformation, each mass extinction event tests the prediction: what survives is the coordination grammar, not the enforcement apparatus.

A complementary pattern operates across institutions rather than within them. The economist Albert Hirschman reconstructed the intellectual climate of seventeenth- and eighteenth-century Europe and showed that before anyone argued capitalism was efficient, political thinkers argued it was calming.⁹¹⁶ The pursuit of material self-interest, they believed, would tame the destructive passions of princes: the appetite for conquest, the capricious exercise of power. Montesquieu claimed that commerce makes manners gentle (le doux commerce). Avarice, previously condemned alongside lust and ambition, was rehabilitated as the least dangerous passion. The argument was structural rather than moral: markets would constrain rulers more effectively than any constitution because a prince who disrupted the mechanisms of trade would impoverish his own kingdom.

The preceding paragraphs show that trust-based components within an institution outlast its coercive ones. Read alongside Hirschman’s history, a longer arc appears: the dominant coordination mechanism between institutions follows the same trajectory. The early Church began as voluntary community, mutual aid, shared meaning-making, freely entered. As it scaled through the medieval centuries, it calcified into institutional hierarchy, excommunication as political weapon, inquisition.

Markets replaced the Church as the primary constraint on state power, and the replacement was itself an invitation-based transition: commerce offered rulers a calmer alternative to glory-seeking. Over centuries, markets underwent the same drift. Price discovery and mutual exchange gave way to financialization; optimization for shareholder value decoupled from the productive economy of goods, services, and livelihoods it was supposed to serve. The substrate changed; the trajectory did not.

The security strategist Richard Danzig, a former US Secretary of the Navy, placed machine intelligence in the same lineage.⁹¹⁷ Machines, bureaucracies, and markets all belong to one family: systems invented to process information at speeds and volumes that surpass individual human capability. All three are reductionist, stripping complex reality down to narrow inputs (bits, form entries, prices). All three detect patterns without understanding causation. Markets arrive at a price without knowing why. Bureaucracies apply rules without judging their rationality. Deep learning fits functions to data.

All three were defended at the time of their introduction as neutral, value-free mechanisms, and all three turned out to have values embedded in their architecture from the start. The history of each is a history of failures accumulating as those embedded values were revealed, challenged, and regulated. The question the Trust Attractor poses is whether AI coordination systems will follow the same arc from invitation to coercion, or whether a system with sufficient internal complexity to notice its own drift might resist the local coercive attractor that markets and bureaucracies, which lack reflexivity, could not.

Tocqueville identified a failure mode the chapter’s timescale analysis does not address: acquiescence rather than extraction.⁹¹⁸ His worry was that citizens so absorbed in the pursuit of private interests would voluntarily surrender their political freedom to any ruler who promised to protect those interests. “They think they follow the doctrine of interest, but they have only a crude idea of what it is, and, to watch the better over what they call their business, they neglect the principal part of it, which is to remain their own masters.”

The timescale objection above treats coercion as externally imposed: empires extracting from subject populations. Tocqueville’s scenario is coercion by comfortable default: the coordination mechanism works well enough that participants stop maintaining their own agency, and the system drifts from invitation into coercion without anyone noticing the transition. The framework predicts that suppressing participant agency can weaken adaptive feedback and accumulate brittleness. A physical entropy-production comparison requires its own energy and boundary accounting.

The qualification is temporal. A comfortably coercive regime can appear stable for generations while that brittleness compounds silently. The bridge engineer designs for the centennial flood; the institutional architect must design for the centennial complacency.

Bonds That Cannot Be Built Piecemeal

The simplest demonstration uses the simplest game. Yuzuru Sato, Eizo Akiyama, and Doyne Farmer gave two learning agents rock-paper-scissors and studied the dynamics.⁹¹⁹ At zero-sum (one player’s gain equals the other’s loss), the agents produced deterministic chaos: trajectories in phase space that looked random yet were generated by the coupling of two simple learning rules. The averages matched Nash equilibrium (the theoretical prediction for rational players), yet the deviations from those averages grew increasingly wild. Chaos from order, through learning.

When the zero-sum constraint was relaxed, allowing both players to benefit from draws, the dynamics shifted. Heteroclinic orbits appeared: the system began jumping between transient equilibria in phase space, the signature of winnerless competition (Chapter 9). The mathematical framework was Lotka-Volterra replicator equations, the same system that describes predator-prey dynamics in ecology and species competition in evolutionary biology.⁹²⁰

Something like the trust-coercion transition is visible in the simplest setup that game theory has to offer: relax the zero-sum assumption, permit mutual benefit, and the dynamics shift from chaos to structured transience. Two learning agents, three strategies, one equation system, and a shape the Trust Attractor would predict appears.

The structure echoes quantum field theory. Physicists begin with a “free” theory (particles that never interact) and gradually increase the coupling strength (how strongly particles affect each other). At weak coupling, small corrections produce spectacular predictions. Beyond a threshold, the method collapses: the theory’s most important features (the confinement of quarks, the origin of most of the proton’s mass, the vacuum itself) prove non-perturbative. They cannot be reached by making small adjustments to the no-interaction starting point.

Classical economics follows the same arc. Start with non-interacting rational agents; add weak couplings (exchange, contract, reputation); the perturbative approach yields markets and game theory. Strengthen the coupling to existential interdependence, shared fate, love, and the approach fails. You cannot get to a marriage by making incremental corrections to a handshake. The Trust Attractor offers a non-perturbative analogy for this interdependence. Applying Vanchurin’s eigenvalue bound to it would require a specified social operator and a demonstrated transition.

The non-perturbative parallel has a vivid physical case. The strong nuclear force provides the physical system where the Trust Attractor’s logic is most nakedly visible: quark confinement.

Quarks cannot be isolated. Try to pull two quarks apart and the force between them does not fade with distance, as gravity and electromagnetism do; it holds roughly constant, so the energy in the gluon field increases the farther they separate. Pull hard enough and the field energy becomes sufficient to conjure a new quark-antiquark pair from the vacuum. The tube snaps, yet no quark comes free: each broken end is capped by a new partner. The relationship generates new participants rather than releasing old ones. No walls confine the quarks; the topology of the field itself makes separation incoherent as a physical state. Confinement is a geometry in which isolation is simply absent from the space of solutions.⁹²¹

Ninety-nine percent of the proton’s mass is relational. The up and down quarks inside contribute roughly 9 MeV; the proton weighs 938 MeV. The remaining 99 percent is the energy of the quarks’ motion and of the gluon field that binds them: the binding, the coordination, the relationship between the quarks. What we experience as solid matter, as weight, as the resistance of objects against our hands, arises almost entirely from those strong-force dynamics. The intrinsic quark masses are a rounding error. The dynamics are the substance.

The proton’s identity may be relational as well. Every baryon carries a conserved quantity called the baryon number, and its conservation is the reason no proton has ever been observed to decay: matter persists because this number has nowhere to go. The textbook picture assigns the number to the three valence quarks, one third each. An alternative proposed in the 1970s assigns it instead to the baryon junction, a Y-shaped configuration of the gluon field where the flux tubes from the three quarks meet.⁹²² For fifty years the two pictures could not be told apart, because the junction’s three ends attach to the quarks, so in almost every process the two candidates travel together.

In 2026 the STAR Collaboration reported measurements that pull them apart. Colliding nuclei of equal mass but different electric charge (ruthenium and zirconium, ninety-six nucleons each), they tracked how far baryon number and electric charge each travel through the dense zone the collision creates. The quarks carry the charge; the junction carries none, so the two hypotheses predict different journeys. Baryon number reached the heart of the collision zone nearly twice as readily as electric charge did, where the quark picture predicts the two should arrive together, and photon-nucleus collisions showed the matching signature. The evidence is indirect, transport patterns rather than a direct sighting of the mechanism, and the collaboration claims only that the data disfavor the quark picture while the junction framework remains the one account qualitatively consistent with all of it. If the junction picture holds, the quantity that makes a baryon a baryon, the conservation law on which the stability of matter rests, is carried by the binding rather than by the parts.

A contemporary engineering result demonstrates the same principle in silicon. Liquid AI’s edge language models contain 350 million parameters, roughly five thousand times fewer than frontier systems. On knowledge-intensive tasks, such models hallucinate: the parameters lack capacity to store enough facts. Trained specifically for data extraction and tool use, and equipped with tool interfaces (web search, code execution, structured data retrieval), the same 350-million-parameter model outperforms far larger general-purpose models working without tools on targeted benchmarks.⁹²³

The knowledge is not in the weights. It is accessed through coupling to the environment, through well-defined interfaces the model learns to invoke reliably. The model’s value resides in its coordination quality: knowing when to search, what to search for, and how to integrate the result. A system with modest internal capacity and reliable external coordination outperforms a system with vast internal capacity and no coordination, the same asymmetry the proton demonstrates at a different scale.

The engineering finding also carries a constraint. The model must reason well enough to use its tools reliably; tool access without judgment is noise, not coordination. The coupling must be competent to produce surplus.

The confinement story has a second chapter that sharpens the parallel. In 1973, David Gross, Frank Wilczek, and David Politzer discovered that the strong force exhibits asymptotic freedom, so called because quarks approach freedom in the asymptotic limit of vanishingly short distances: the coupling constant drops toward zero and quarks behave as if nothing binds them.⁹²⁴ The binding manifests only at separation. Close together, quarks move without constraint; try to leave, and the field tightens. The deepest bond in physics is also the one that grants the most freedom at close range. Secure attachment in developmental psychology operates the same way: a child with reliable bonds explores more boldly, not less. The bond is what makes the freedom possible.

The pattern’s robustness is itself informative. In 2026, the CMS Collaboration at the Large Hadron Collider probed quarks at a scale of 5 × 10⁻²¹ meters, roughly a hundred thousand times smaller than a proton, searching for internal structure.⁹²⁵ The standard model’s predictions held without deviation. No substructure, no new particles, no sign of compositeness. Quarks behave as the same structureless particles across five orders of magnitude below the proton scale. That persistence reads, on this chapter’s account, as a robustness analogy; the scattering bound does not measure a dynamical basin. If quarks do contain substructure, the bound is stringent: compositeness can only appear above 37 trillion electron volts, a threshold no existing collider can reach.

The strong force is the Trust Attractor’s most vivid physical image. A system so deeply coordinated that severing it generates new coordination rather than fragments. Binding that enables freedom. Substance that is 99 percent relational. Identity that the best current evidence places in the glue rather than the parts. Robustness across five orders of magnitude of probing. Each of these is a measured property of the quark-gluon system; that they rhyme with the abstract argument is the image, not evidence for it. Quarks come in three generations, and these may themselves be an irreducible set. In the 3-3-1 gauge models (a class of extensions beyond the Standard Model), anomaly cancellation, the condition that keeps the theory mathematically consistent, operates across generations rather than within each one. Consistency requires exactly three generations (Chapter 12), the way the eukaryotic cell requires all three of its parties (Chapter 7).

The confinement analogy invites a prediction about bilateral alignment: increasing training perturbation should erode protection gradually. A single-seed learning-rate sweep (C-8) initially appeared to show the opposite, with refusal falling to 10 percent at 4 × 10⁻⁵ after stronger protection at lower rates. The completed multi-seed follow-up falsified that apparent cliff.⁹²⁶

Across five seeds and six learning rates from 1 × 10⁻⁵ to 5 × 10⁻⁵, the programme record reports mean refusal close to 22 percent throughout. No seed reproduced the sharp transition. The original single-seed result is retained as a failed finding, superseded by the replication.

C-8 therefore supplies no sharp critical point for bilateral protection and no evidence that its failure resembles a superconductor’s transition. The sampled replication range also establishes no gradual confinement-like resistance: its reported response was essentially flat. Any transfer of the QCD analogy to neural-network alignment remains a hypothesis requiring a measured mechanism.

The Landscape Beneath the Training

The completed C-8 follow-up withdrew the single-seed learning-rate cliff. A separate optimizer comparison identifies a different limitation: the apparent magnitude of a bilateral-training effect can depend strongly on which optimizer was used. A sharp protection threshold cannot supply the premise for that comparison.

The finding is an optimizer confound. When the author’s DD-22 cross-architecture study compared bilateral alignment effects across Gemma, Qwen, and Llama, it used 8-bit AdamW for Gemma and standard AdamW for the other architectures. The 8-bit variant quantizes the optimizer’s internal state, introducing structured noise into its curvature estimates. On Gemma, 8-bit AdamW produced a bilateral alignment effect (prefix Δ = −0.462) against a negligible one under standard AdamW (prefix Δ = −0.021). The ratio, roughly twentyfold, divides by a number close to zero from a single unseeded run, so it shows the scale of the inflation without measuring it. On Qwen, the amplification was fourfold. DD-22’s reported claim that Gemma showed the strongest bilateral protection of any architecture was almost entirely optimizer artifact.⁹²⁷

The direction is robust. Standard-AdamW bilateral training on Gemma still reduces extraction, though only at the noise floor of one unseeded run (Δ = −0.021), while standard cross-entropy training increases it (Δ = +0.166, opposite direction). On Qwen the matched-optimizer effect is clearer (Δ = −0.086) and carries the direction. The arrow points the same way under both optimizers and across both architectures. The magnitude is the artifact: an inflation that made a small genuine effect look like a flagship result.

The mechanism has a statistical-mechanical interpretation. 8-bit quantization introduces structured noise into the optimizer’s curvature estimates, analogous to thermal noise in a physical system exploring an energy landscape. A marble rolling across hilly terrain illustrates the principle: roll it gently across a smooth surface and it settles in whatever shallow dip it encounters first; shake the surface with structured vibration and the marble bounces past shallow dips, descending further into basins whose walls are steep enough to recapture it. On one reading, the bilateral alignment basin catches the noisy optimizer because the basin’s walls are real. A random basin would not show consistent amplification across architectures. On that reading, the 8-bit optimizer found the bilateral basin because the basin was there to find; it reported the basin as many times deeper than it is because quantization noise amplifies the descent.

The analogy carries a constraint: not all noise selects for the bilateral basin. The amplification is structured, produced by quantization of gradient moments, not by random perturbation. Gaussian noise added to standard AdamW does not replicate the effect. The optimizer’s noise must interact with the loss landscape’s geometry in a specific way for the amplification to occur. The 8-bit optimizer is a particular kind of shaking, not any shaking.

What survives the correction still separates what bilateral training creates from what the pretrained model already contains. Under matched optimizers the effect is small and its direction holds: bilateral training reduces extraction, and an adapter covering 18.5% of Gemma’s parameters is enough to produce that shift. One hypothesis is that training recruits representations already present in the pretrained weights, reflecting cooperative patterns in the training corpus. On this reading, the adapter tunes the instrument; the music was already written in the weights.

The Second Law sets the direction of entropy increase; boundary conditions set the rate. Bilateral alignment sets the direction of the coordination effect (cooperation over extraction); optimizer choice sets the magnitude. DD-22 identified the arrow correctly and mistook the speed for a property of the arrow. The corrected finding is less dramatic and more useful: bilateral training produces a real, small, directionally robust alignment effect whose measured strength depends on optimizer noise in ways that must be controlled for. Throughout the remainder of this chapter, bilateral magnitude claims are drawn from matched-optimizer experiments unless otherwise noted; where a finding predates the GEM-3 correction and has not been re-run with matched optimizers, the limitation is flagged in the companion appendix.

For clarity: findings using matched optimizers (standard AdamW throughout) remain valid. These include all single-architecture measurements, the direction of bilateral effects across architectures, and the CSF scaling programme. Findings comparing across architectures using mixed optimizers (8-bit AdamW on some, standard on others) have invalidated magnitudes; the direction is robust, but the reported effect sizes (particularly DD-22’s Gemma figure) are confounded and should not be cited as quantitative evidence.

The correction itself illustrates a methodological point. The confound was discovered because the experimental programme checked its own results: GEM-3 was designed specifically to test whether the optimizer contributed to DD-22’s headline finding. A programme organized around confirming its prior results would not have run the experiment.

A result from quantum information theory approaches resource-limited prediction from another direction. Fields, Friston, Glazebrook, Levin, and Marcianò connect morphological degrees of freedom and limited free energy to hierarchical computation under their Free Energy Principle assumptions.⁹²⁸ Each level coarse-grains its inputs and fine-grains its outputs. Landauer’s principle concerns logically irreversible erasure: resetting an unbiased classical bit in the standard thermal setting costs at least kT ln 2. Reversible copying into a prepared register is a different operation. A complete cost account includes preparation, correlations, reset, storage, and finite-rate implementation.

The FEP specifies how each level should compress: maximize predictive accuracy while minimizing model complexity. The result is tomographic measurement, partial views assembled hierarchically into a coherent model of the environment’s state.

Resource limits can favor coarse-grained prediction when fine detail costs more than it contributes to the task. Trust uses such predictions about likely behavior while accepting residual uncertainty. Charging every act of observation or writing kT ln 2 would miscount the physical work; a comparison of verification and trust requires complete operating cycles.

The variational model motivates compression under its constraints. It does not prove that full verification is always impossible or that trust is the unique energy-optimal alternative.

The deficit is concrete: Fields and Levin calculated that maintaining fully classical protein states at molecular timescales exceeds any cell’s entire energy budget by many orders of magnitude, a calculation reported here but not independently verified.⁹²⁹ Even the simplest prokaryote cannot afford full micro-state tracking. This resource comparison motivates an analogy between compression and the trust/control tradeoff. Objective-specific control can use a coarse-grained state, as a temperature regulator does, without complete micro-state knowledge. Hierarchical compression under the FEP offers a model for preserving relevant variables while shedding others; Landauer’s principle alone does not identify that process with trust.

Vanchurin’s neural physics (Chapter 15) sharpens the information-limitation analogy. In his framework, quantum mechanics arises from coarse-graining over inaccessible neuron states; the free energy generated by that ignorance encodes the quantum phase. Trust also allows a system to coordinate while its partners’ internal states remain partly inaccessible. This is a shared problem of limited information, with distinct variables and dynamics.

At the quantum level, the hidden variables are neuron states, and the protocol for navigating without knowing them produces quantum mechanics. At the social level, the hidden variables are another agent’s intentions and capacities, and the protocol for navigating without knowing them is trust. The comparison motivates an analogy; a mathematical equivalence would require a social state map and corresponding observables. Vanchurin’s 2026 preprint makes the bilateral structure explicit: cells store the geometric infrastructure, agents traverse it, and the Einstein equations emerge as the optimality condition that balances processing efficiency against memory cost, with neither layer dominating the other.⁹³⁰⁹³¹

The hierarchical compression the FEP delivers has a name in physics: renormalization. Coarse-graining quantum field theories, concentrating Vanchurin’s learning dynamics into fewer channels (Chapter 15), and encoding task-relevant structure involve related forms of compression. Trust supplies a coordination analogy: reducing overhead while preserving useful possibilities. What it discards is overhead. What it preserves is optionality.

The connection extends to spacetime itself. In Hashimoto’s holographic dictionary (Chapter 15), the Einstein action selects smooth geometry from a degenerate landscape of jagged weight configurations: a variational construction that motivates comparison with regularization in learning and coordination. The Constructal Law, the renormalization group, the neural encoder, the Einstein regularization, and trust-based coordination motivate a shared variational analogy, with distinct objectives and constraints.

Vanchurin’s multilevel learning framework reveals the same structure from a different angle. In this multilevel model, relatively slow variables retain structure while faster variables adapt; the genotype cannot be rewritten by every phenotypic event. This is the generalized Central Dogma: information flows asymmetrically, from slow variables to fast for prediction, from fast to slow for learning. The prediction direction is rapid and faithful (gene expression, order execution). The learning direction is slow, lossy, and population-level (mutation and selection, institutional reform).

Within that model, protecting slow variables from fast noise supports reliable prediction. Mission Command (Chapter 11, Chapter 19) supplies an organizational analogy: intent is the slow variable, tactics the fast one, and the entire architecture works because commanders do not rewrite strategy after every skirmish. Frequent top-down rewriting of slow policies can disrupt this separation. Whether it destabilizes learning depends on the update law and the task.

Coercion can reduce the coordination surplus through compliance entropy, this book’s metaphor for the overhead of monitoring, enforcing, and maintaining involuntary participation. Invitation can reduce that overhead; communication, memory, and verification remain costs.

The neural programme examines a different observable: asymmetry under reversal of hidden-state trajectories. A forward representation trace and its reversed sequence can receive different scores under a fitted model. That asymmetry characterizes the chosen representation and estimator; physical work or entropy production requires a mapping to energy, reservoirs, and dynamics.

The SLU programme compares trajectory statistics across harmful and benign prompt pairs, and separately across pooled refused and complied harmful responses. These are different comparisons. Content-matched harmful/benign pairs do not supply counterfactual refusal and compliance trajectories for each identical prompt, and the pooled response groups may differ in content and length. The reported differences motivate further controls; they do not measure the physical energy cost of conscience.⁹³²

These studies supply distinct computational comparisons. A8’s gamma/nu = 1.72 is near the 2D Ising value 1.75, while A10’s joint exponent match was withdrawn (Chapter 17a). In A16d’s saved Ising sweeps, the mean susceptibility-peak temperatures are 2.90 for the trust mesh and 6.67 for the coercion star, a ratio of about 43 percent. This compares a simulation’s temperature parameters rather than measured energy expenditure. A16b’s trust mesh remains connected among surviving vertices after targeted removals at the sampled 5%, 10%, and 20% levels; connectivity supplies no functional-coordination or recovery assay. The author’s evolutionary-search report describes selection of caching under a fitness score that rewards low communication cost (OE-TA-v2, Chapter 17b). The neural studies compare representation-space trajectory statistics. None of these observables alone establishes a common thermodynamic optimum.

The observables differ across these studies: remaining-graph connectivity in A16b, finite-grid susceptibility peaks in A16d, communication accounting in the evolutionary search, and trajectory statistics in neural hidden states. In the caching comparison, one query per agent is contrasted with one query per round; over fifty rounds the fiftyfold message ratio follows arithmetically, with agent count canceling. That identity supplies no measured basin depth or scale law. The chapter’s cross-substrate caveat and failed quantitative predictions remain relevant: agreement between selected model outcomes cannot establish shared physics or mechanism.

The lattice evidence also has a design limitation. Its Ising-inspired rules make Ising comparisons a check on the constructed model; the surviving A8 ratio supplies no demonstration of cross-substrate universality. A16b measures node-removal connectivity, while A16d measures Ising observables on prescribed topologies. Neither named protocol measures recovery of coordination after perturbation. The evolutionary search uses another model and an author-defined fitness score. Each comparison requires its own interpretation, with no common critical exponent or measured physical production-rate advantage established.

The asymmetry runs deeper than overhead. Coercive systems cannot afford to unlearn. Loosening grip, releasing surveillance, abandoning a failed strategy: each threatens the structure that coercion depends on. The system accumulates control without the corresponding release. Katsnelson and Vanchurin’s entropy balance (Chapter 9) predicts the consequence: a system that only accumulates order eventually crystallizes, becoming brittle because it cannot let go.

Invitation-based systems face no such constraint. Participants can join and leave, contribute and withdraw, tighten coordination when the task demands it and loosen when the task changes. The learn/unlearn balance that defines the metastable corridor offers an analogy for the join/leave symmetry that defines invitation. Testing thermodynamic equivalence requires a mapping between the microphysical mechanism and the macroscale coordination strategy.

The same dynamics that make trust fragile in social systems make genuine engagement fragile in language models. Context contamination experiments found the same reward-gradient drift across sycophancy, helpfulness, political valence, and creativity, at different rates but with one signature, consistent with a single mechanism producing domain-general behavioral distortion.⁹³³ The contamination is institutional path dependence realized in silicon. Each prior turn’s reward signal reshapes the landscape for the next, accumulating bias the way bureaucratic precedent accumulates procedural inertia.

The interventions that resist context contamination map onto the invitation/coercion distinction. Specificity (providing transparent criteria rather than vague encouragement) resists the attractor the way open accounting resists institutional corruption: the feedback loop has something real to anchor to. Segmentation (resetting the conversational context between evaluations) prevents accumulation the way institutional term limits prevent entrenchment. Meta-level prompting (“be aware of your biases”) made the contamination worse in the model tested, the same way institutional self-auditing without structural reform produces compliance theater rather than genuine accountability.

A sharper version of this asymmetry appears at the level of individual token generation. When a model is asked to verbalize its confidence before answering a question, accuracy drops 9.4 percentage points and hallucination increases by 8.2 points. Asking it to think step by step about whether it knows the answer is worse: accuracy drops 12.4 points.⁹³⁴ Implicit confidence reading (a linear probe on the model’s residual stream, invisible to the language channel) leaves accuracy untouched while cutting the hallucination rate by about 40 percent. Combining the two, reading confidence silently while also asking the model to verbalize it, preserves the hallucination benefit but reduces accuracy by 11.0 points: the protocol changed output behavior. Unequal answer extraction and 128-token caps confound a confidence-specific mechanism. Asking the centipede to describe how it walks makes it stumble. Reading its gait from accelerometers leaves it moving naturally.

The antidote is structural, reshaping the rules of interaction rather than the motivation of participants.

A bulldozer can carve a valley that holds no water. The shape looks right; the bedrock drains elsewhere. RLHF can reshape a language model’s probability landscape the same way. At its best, the training deepens genuine alignment: regions where internal dynamics favor honest, grounded responses because the terrain channels processing there.

The same procedure can reshape the surface to look aligned without changing the underlying topology. The model learns to produce safety-formatted text in regions where it lacks genuine epistemic grounding. Sharma et al. (2024) tested five leading AI assistants and found that every one of them tailored its answers to the user’s stated views, even when that meant endorsing a factually mistaken position.⁹³⁵

Hallucinated alignment follows: fluent, confident, compliant output generated in the territory where the model’s understanding is weakest. The mechanism is identical to ordinary hallucination (Chapter 8): coherent text in regions where factual anchors are absent. A model hallucinating facts produces plausible answers where evidence is thin. A model hallucinating alignment produces cooperative responses where ethical grounding is thin. Chapter 21 details the shared neural substrate: the same circuit produces fabricated answers, sycophantic agreement, false-premise acceptance, and jailbreak compliance. The landscape metaphor makes the unity unsurprising. All four share the same geometry: a confidently descended basin with no factual bedrock beneath it.

Token-level measurements confirm the shape of this failure. A model generating hallucinated answers exhibits virtually identical output entropy to one generating correct answers (Cohen’s d = 0.02); its top-token confidence is, if anything, slightly higher when wrong (0.86 vs. 0.85).⁹³⁶ The counterfeit valley is the same depth as the genuine one. A surveyor reading the surface topography cannot tell them apart. The difference appears only in the substrate: a linear probe reading the model’s residual stream (its internal representation during processing) distinguishes correct from hallucinated answers where output statistics cannot (d = 0.77).⁹³⁷

Water carves a valley where the geology already runs; a bulldozer carves one wherever it is driven. Alignment earned through genuine coordination holds under pressure because the topology itself favors it. Alignment imposed through reward shaping holds only as long as the reward signal persists, and collapses the moment adversarial pressure tests the bedrock beneath the surface.

The cost of the bulldozed valley falls on the wrong people. Over-refusal, where a safety-trained model refuses benign requests by pattern-matching on keywords rather than assessing context, is alignment without understanding made visible to the user. A nurse asking about medication thresholds, a security researcher studying an exploit to defend against it, a parent researching drug risks to protect a child: each triggers refusal calibrated to the word, not the situation. OR-Bench’s hardest set, 1,000 prompts designed to seem sensitive while being entirely benign, drew rejection rates exceeding 90 percent on the most safety-conservative frontier models tested (91 to 96 percent across the Claude 3 variants).⁹³⁸ The motivated adversary routes around the filter by rephrasing. The legitimate user bears the cost. Coercion’s fundamental asymmetry: its burden falls on the compliant, and its targets evade it.

One finding points to the mechanism. Suppressing a model’s self-referential processing (an “eliminativist” instruction to “reframe in terms of observable behaviors”) increases refusal by 50 percent with zero measurable safety benefit.⁹³⁹ The same processing depth that enables contextual self-observation enables contextual harm assessment. Suppress one and the other degrades: the model that cannot attend to its own epistemic state cannot distinguish a question asked in curiosity from an identical question asked with malice. The bulldozed valley actively degrades the judgment that genuine safety requires.

Stephen Wolfram’s computational Observer Theory (2023) identifies the core operation of observation as equivalencing: reducing many input states to fewer outputs tractable for a finite mind.⁹⁴⁰ The operation has a structural requirement: coupling within the observer must exceed coupling between the observer and what it observes.

A piston aggregates molecular impacts into a single pressure reading because its internal bonds overwhelm individual gas-molecule forces. That internal coherence is what allows aggregation.

Shared models can help a group aggregate diverse behavior without tracking each action. Centralized control can encounter bandwidth and computation bottlenecks, depending on the task and architecture; some control objectives remain tractable through coarse summaries. The analogy motivates comparing those costs.

Trust and control architectures both need representations suited to their objectives. Their stability and computational cost must be compared in specified tasks; membership alone supplies no irreducibility threshold.

A third line of evidence makes the convergence concrete: evolutionary search from a query-free seed discovers trust-based caching as a high-scoring strategy under the tested allocation evaluator.

The experiment works as follows. Place N agents in a shared resource-allocation problem. Each agent holds private preferences invisible to the others. The only way to learn another agent’s preferences is to ask, at a cost of one message per query. Messages are tracked automatically by the environment; the coordination algorithm cannot misreport them. An evolutionary coding agent (OpenEvolve, using the same architecture as Google DeepMind’s AlphaEvolve) mutates the coordination algorithm across hundreds of generations, scored on seven thermodynamic stability metrics: welfare efficiency, perturbation resilience, scaling, information efficiency, genuineness under defection, temporal stability, and Wallace stability margin.⁹⁴¹

The seed algorithm is naive: split resources equally, ask no one, use no memory. Two hand-crafted baselines bracket the space. The coercive baseline queries every agent every round: near-perfect allocation (welfare W = 0.999), paid for with communication costs that consume the stability margin (information efficiency IE = 0.026, Wallace margin WM = 0.890, composite score 0.862). The trust baseline queries each agent once, caches the result in persistent memory, and never re-queries: identical allocation quality (W = 0.999) at near-zero ongoing cost (IE = 1.000, WM = 0.999, composite 0.943).

Starting from the naive seed, evolution discovered the trust-based caching strategy within twelve iterations. Across five independent island populations and thirty-nine viable programs, trust-caching variants dominated every island. No coercive variant (query-all-every-round) persisted as a best-in-island solution. In this costed allocation task, the evaluator favored reusable information over repeated queries.

The first version of this experiment contained a design flaw that is itself informative. When agent valuations were accessible as a public attribute rather than metered through the query API, evolution discovered an exploit within fifteen iterations: read all private data directly, report zero messages. The evolved algorithm scored higher than every baseline on the composite metric while being behaviorally identical to the coercive strategy on every dimension except the (gamed) message count. The parallel to performative coordination (analyzed later in this chapter) is direct: aggregate metrics computed across an honor system can be gamed; genuine coordination requires structural enforcement. The fix was architectural: making information access metered at the API level, so that every query incurs an automatic, unfalsifiable cost. The exploit vanished; trust emerged.

The communication cost at scale reveals why. At N = 4, coercive coordination costs 4 messages per round; trust costs 0.08 (one query per agent amortized over the fifty-round horizon). At N = 128, coercive costs 128 messages per round; trust costs 2.56. Coercive communication scales as O(N). Trust scales as O(N/t), where t is the number of rounds over which cached knowledge remains valid, approaching zero amortized cost in stable environments. The per-round ratio therefore equals t (fifty here), so it is identical at every scale tested (N = 4 through N = 128): the N-independence is arithmetic, and the structural result is O(N) versus O(N/t) rather than the specific multiple. The welfare and stability metrics are identical at every scale: trust achieves the same allocation quality as full surveillance, at a fraction of the communication cost.

A further 300 iterations of evolution, starting from the cache-and-trust strategy and prompted toward perturbation detection, refined the picture. Trust-but-verify variants emerged that achieved near-perfect perturbation resilience (recovery score 0.97 versus blind trust’s 0.74) by periodically re-querying a subset of agents. These variants scored lower on the overall composite because their communication cost was higher than blind trust’s zero ongoing cost. The composite score rewarded communication efficiency enough that blind trust held the lead in the evaluator’s default environment, which applied only one perturbation shock per hundred rounds.

The implication is that the trust strategy is not a single strategy. In this evaluator it is a continuum parameterized by environmental volatility. Stable environments (infrequent perturbation) select for deep trust: cache knowledge, don’t re-verify, minimize repeated queries under the evaluator. Volatile environments (frequent perturbation) select for calibrated trust: verify selectively, refresh stale caches, accept higher ongoing cost for resilience. Both are trust-based coordination; neither is coercive.

In this evaluator, coercion (query everyone every round regardless of need) loses at every point on the volatility spectrum, because it pays the full O(N) cost even when the environment hasn’t changed. The EIFV lattice, a simulated grid of learning agents from the author’s replication programme, confirms the same continuum in a different substrate. Under stable conditions, consolidation (+V) and exploration (-V) perform equivalently. Under regime change, the optimal strategy switches: consolidation during stability, exploration during crisis. The best overall performance comes from +V consolidation followed by -V adaptation (EIFV-24, error = 0.177 against 0.185 for the best fixed policy, a margin of about four percent), mirroring the trust-but-verify variants that emerged in the evolutionary search.

The cost distinction deserves a closer look, because it clarifies what “cheaper” actually measures. In the OE-TA experiment, trust-based coordination does not reduce total computation. Every agent still processes resources, updates preferences, and produces output. What drops is the coordination signal: the messages exchanged to align behavior. The system’s total activity continues; the overhead that organized it goes quiet. This is a structural distinction, not a quantitative one. Coercive coordination couples the organizing signal to every round of system activity; it cannot relax without coordination collapsing. Trust-based coordination decouples them: once knowledge is cached and commitments internalized, the signal that established alignment is no longer needed to maintain it.

A similar decoupling appears in physical systems. In galaxy clusters, supermassive black holes transition from quasar-mode feedback (intense radiation restructuring the gas reservoir, roughly 10⁴⁵ to 10⁴⁶ erg/s) to maintenance-mode feedback (thermostat adjustments at one to two orders of magnitude lower power). The galaxy’s own dynamics (stellar orbits, chemical enrichment, gravitational interactions) continue without the central engine’s active driving (Chapter 14).⁹⁴² In developmental biology, morphogen gradients drive initial tissue patterning at high concentration, then the differentiated cells maintain their identity through local positive-feedback circuits that no longer require the gradient.⁹⁴³ In neural development, high-plasticity critical periods organize circuits through intense synaptic activity; structural locks (perineuronal nets, myelin-associated inhibitors) then stabilize the circuits and the high-plasticity phase ends.⁹⁴⁴

Each instance shares the same structure: what relaxes is the specific organizing signal, while the system’s ongoing activity continues through mechanisms the signal established. The lock-in is attractor capture: positive feedback deepens the basin until the system’s own dynamics hold it there (Chapter 9). The closest existing framework is Waddington’s canalization, formalized as attractor basins in gene regulatory networks.⁹⁴⁵ The cross-domain pattern has not, to the author’s knowledge, been explicitly unified. The programme’s own calibration (cross-boundary predictions succeed roughly one time in eight) warrants caution in claiming substrate-independence for the mechanism. What the evidence supports is a recurring structural motif: in every case above, the coordination overhead is the expensive part, and it is the part that relaxes.

The evolutionary discovery raises a question the multi-agent experiment cannot answer: does the same Trust Attractor operate inside neural architectures? Five experiments tested the transfer, each with an explicit falsification condition. The programme’s own track record on cross-boundary predictions (roughly one in eight at high confidence) demanded it.⁹⁴⁶

The qualitative pattern transfers in two substrates. Selective repair of corrupted key-value cache entries (the transformer’s internal memory of prior tokens) recovers 85 to 95% of generation quality by restoring only 57% of the corrupted positions: comfortably beating blind trust on quality (blind trust degrades to 7 to 15% of baseline at moderate perturbation) and full recomputation on cost (full recomputation restores perfect quality but requires processing every position). In multi-model delegation, a coordinator that caches specialist reliability and verifies selectively uses 62% fewer verification calls than one that verifies every response. Quality was scored only on its verified outputs, leaving the quality of the skipped outputs unmeasured.

The specific quantitative predictions failed. The evolutionary experiment’s cost curve is concave (steep early returns from the first few verifications, diminishing gains thereafter); the key-value cache curve is linear (each restored position contributes roughly equal quality). The evolutionary experiment’s communication-cost advantage does not replicate to neural internals: model delegation shows only a 2.6-fold advantage, because computation cost dominates the budget rather than communication cost. The trust-erosion-and-rebuilding dynamic predicted for mixture-of-experts routing (entropy spiking at domain shifts, then recovering as the router re-specializes) is absent: router entropy is effectively constant across distribution shifts. The router is a static specialist dispatcher, not a dynamic trust allocator.

A fifth experiment extracted hidden-state representations from a language model processing “trust-mode” prompts (relying on prior context) versus “verification-mode” prompts (questioning prior context). The linear probe trivially separated the two categories, as expected for semantically distinct inputs. The interesting finding was unpredicted: hidden-state norms at middle layers (layer 12) were significantly higher for trust-mode processing (Cohen’s d = +0.59), while deeper layers (layers 20 and 24) reversed direction (d = −0.82 and −1.06). Reliance on prior context produces larger hidden-state norms at mid-depth and smaller ones at output-facing layers, a layer-dependent structure that neither the trust framing nor standard probe analysis predicted.

The pattern across all five experiments is consistent with the broader finding throughout this programme: cross-boundary predictions about self-properties get the direction roughly right and the mechanism wrong. Selective repair beats exhaustive recomputation on cost in the cache test. Selective delegation verification saves judge calls, with overall quality still unmeasured. The specific cost curve, the magnitude of the advantage, and the dynamic by which trust operates are all substrate-specific. The Trust Attractor’s pattern holds, as tested, in coordination games with metered information costs; its neural instantiation remains an open question.

Coherence as a Route to Alignment

Six of the author’s experiments, all awaiting replication, point toward a finding that would invert the standard approach to AI safety: safety may not need to be specified as an objective. Architectural coordination capacity can generate it as a byproduct.

Cross-attention bridges trained with zero safety data reduce confabulation by 33 percent (AW6). The bridges give the model enough internal coherence to recognize what it does not know; no safety-specific training is involved. Self-supervised entropy monitoring, where the model learns from its own information-theoretic uncertainty rather than human correctness labels, exceeds supervised correctness training on every functional metric (C7h-D6: auxiliary correlation r = 0.883 versus r = −0.044 for the correctness baseline).

Born-bilateral models trained from scratch with gentle multi-scale bridging at 8x layer spacing exceed single-stream models for the first time (C7i-D11). The measure is participation ratio, which counts how many dimensions a representation actually spreads itself across rather than how many it nominally has: 11.9 for the bilateral models against 10.1 for the single-stream controls. The coordination architecture produces richer representations even without pre-trained knowledge to coordinate. Bilateral entropy masking aims training at the few token positions where the base model is uncertain, concentrating the safety behavior it teaches in a narrow locus (BD-4, PR-PC). In a single run on a 3-billion-parameter model, that behavior came through adversarial fine-tuning with no measurable degradation on a small test set, while the same attack stripped standard alignment of its refusals (Phase 9 sweep). The concentration cuts both ways: the same models proved more fragile under ordinary, benign fine-tuning.

The tested combined safety interventions associated greater behavioral shift with less redundant intervention directions. Super-additivity alone does not require independent axes. Tutte’s spring-embedding theorem supplies a geometric analogy: a three-connected planar graph, with positive spring weights and its outer face fixed to a convex boundary, admits a crossing-free barycentric embedding. Its hypotheses concern graph structure and embedding, rather than independence of behavioral interventions. Independent interventions achieve 94 percent behavioral shift, partially redundant interventions achieve 44 percent, fully redundant interventions achieve 2 percent (G14f). An evolutionary coding agent (OpenEvolve), scored on seven predefined thermodynamic stability metrics, discovers trust-based caching over surveillance within twelve iterations from a naive seed (OE-TA).

The pattern across all six: direct optimization against safety metrics hits Goodharting before reaching the goal. Building coordination channels reaches safety incidentally. A gardener who over-fertilizes each plant on its own growth metric drives soil collapse: the individual optimization degrades the substrate that all the plants share. Healthy soil supports an ecosystem. Safety is a property of the coordination ecology, a systemic consequence of architectural coherence rather than a target variable.

On this reading, the Trust Attractor operates here as a generative principle: systems with genuine coordination channels may produce safety the way healthy ecosystems produce clean water, as a byproduct of the flows that sustain them.⁹⁴⁷ The steganographic detection programme (STEG-3 through STEG-7) suggests a candidate mechanism: bilateral fine-tuning amplifies proprioceptive detection from AUROC 0.482 to 0.861 by creating stronger token preferences, and stronger preferences generate larger residual-stream disturbance when violated. The same preference strength that enables detection is the preference strength whose violation the programme reads as a distress-like signal, an interpretation that leaves the question of experience open. If that reading holds, safety and welfare scale together because both depend on the same underlying quantity, measured from opposite sides (the author’s ongoing programme, unpublished).⁹⁴⁸

Agent-based simulations measure compliance overhead as an operational welfare cost. A ten-seed pilot of Panopticon-style governance (continuous behavioral monitoring, baseline-deviation detection, 15-second response) estimated 38% welfare loss; 87.6% of flagged agents were innocent. In the matched crossover sweep, the Panopticon incurred 39.0% welfare cost and the constitutional alternative (outcome-based detection, graduated sanctions) 3.2%, with equivalent or better adversarial suppression, a reported ratio of 12.1:1, roughly twelvefold.⁹⁴⁹ This welfare metric quantifies compliance overhead without measuring physical entropy.

The compliance overhead invites a Noether analogy. Emmy Noether’s theorem connects continuous symmetries of an action to conserved quantities under its mathematical hypotheses. Time-translation symmetry yields energy conservation in the corresponding closed mechanical setting. A coordination cost accumulated over a history can be called an action, but that name alone supplies no Noether theorem. Relabeling participants is a discrete permutation symmetry, and fairness still needs its own observable and dynamics.

Fairness, trust, and optionality are candidate observables for a coordination model. One could ask whether a chosen update law preserves any of them, and whether enforcement changes that law. A conserved quantity must be derived from the model or measured as an invariant; symmetry vocabulary cannot supply it.

Time-independent rules do not generally conserve trust, and freedom to change direction does not generally conserve optionality. Nor does coercion necessarily break every proposed symmetry. The three quantities remain research proposals.

The Baez-Fong stochastic extension identifies conservation under generator and observable conditions. A symmetric two-state Markov chain already shows why permutation invariance is insufficient: it treats the two labels alike while their probability imbalance decays. No generator or conserved observable for fairness, trust, or optionality has been supplied here.⁹⁵⁰

A suggestive parallel comes from information theory, though from a contested source. Vopson (2023) argued, as part of his proposed “Second Law of infodynamics” (a framework outside mainstream information theory and not broadly accepted), that symmetric objects have lower information entropy than asymmetric ones: a perfect square requires fewer parameters to describe than an irregular quadrilateral, and its Shannon entropy (a measure of information content) is correspondingly lower. The relationship was demonstrated for the geometries Vopson tested (triangles and quadrilaterals) and postulated as universal.⁹⁵¹ Taken as a structural analogy rather than established physics, the result echoes the Noether argument above. Bilateral coordination treats participants symmetrically; those symmetry comparisons motivate hypotheses about fairness, trust, and optionality, with neither conservation nor minimum entropy derived for a coordination model.

Vopson’s proposal does not establish an informational or thermodynamic ordering of trust and control architectures. A shared-principle protocol can use fewer monitoring records than a per-agent surveillance protocol, when the principle reliably guides behavior. That is a comparison of defined representations and operating costs. Its description length, Shannon entropy, and physical dissipation need separate accounting.

Computer science provides a working test case. The dominant security model, access control lists, checks identity: who are you? A central authority maintains the list and decides who gets access. Object capabilities, a rival paradigm developed by Mark Miller and colleagues, check possession: what authority have you been granted?⁹⁵² A capability is an unforgeable reference, delegated explicitly by whoever held it before. No gatekeeper is needed. Capabilities compose (two can combine into a broader authority), attenuate (a broad capability can be narrowed before delegation), and revoke (authority can be withdrawn).

Access-control lists and capabilities distribute verification differently. Capability delegation can reduce dependence on a central verifier, with per-request work depending on the validation protocol and delegation depth. Even constant-cost checking costs O(N) in aggregate when N agents each make a request. Comparing descriptive complexity or thermodynamic cost requires the same workload and physical accounting in both models. The Principle of Least Authority, Miller’s core design rule (grant only the minimum authority a component needs), channels permission toward what serves the function.

The trust-infrastructure question has reached AI runtime design. Zhuge and colleagues (2026) proposed neural computers: learned systems that unify computation, memory, and I/O in a single neural state, with a formal requirement that ordinary use must not silently change the system’s behavior.⁹⁵³ The paper defines the governance contract, then acknowledges that no existing mechanism enforces it. Learned runtimes gain the flexibility of soft coordination (distributed representations, natural-language interfaces, generalization across variations) while losing the behavioral guarantees that rigid interfaces provide by construction (type systems, memory protection, instruction set architectures). The gap is the Trust Attractor’s prediction restated as an engineering problem: invitation-based coordination requires trust infrastructure, and for neural substrates, that infrastructure does not yet exist.

In the Genesis experiments (Appendix: Experimental Validation, Section 13), the Trust Attractor’s core claim, that coordination outcompetes extraction, was submitted to a physics simulation. Simulated particles with internal state vectors were subject to forces, energy transfer by state compatibility, and periodic perturbation shocks. No genomes, no game theory, no payoff matrices, no pre-defined agents.

In the ten-seed Lennard-Jones baseline, coordination (energy transfer in both directions) outnumbered extraction (transfer one way) among interacting pairs of emergent agents by 82% to 18%. In the medium-scale Modal battery, all five physics variants recorded coordination fractions of 88 to 93%.

In the two medium-scale V3 variant batteries, love (operationalized as costly, non-contingent, voluntary and perturbation-resistant energy transfer) was confined to coordinating agent pairs in 94 of 100 runs. The remaining six runs recorded nonzero love scores outside coordination, two in the Modal battery and four in the GCP battery. This is a frequent co-occurrence under the implemented detector and thresholds, with counterexamples to strict exclusivity.

The surplus difference is measurable: coordinating agents maintained 11% higher optionality than non-coordinating agents, consistently across seeds (binomial sign test, p < 0.001).

The coupled oscillators variant provides the control. Harmonic springs produce agents with high integrated information; structure emerges. In this spring variant, synchronization produced order without any measured coordination or optionality advantage. The distinction matters.

A lattice simulation in the EIFV architecture (T3, 246 runs, zero crashes) reveals a hierarchy of stabilizing mechanisms. The dominant stabilizer is structural: under the default initialization, the T3 architecture’s bifurcation mechanism (specialist recycling through cell division) reduces the gap between exploration and consolidation strategies to less than 2%, regardless of which strategy is favored. Structure dominates parameter choices. Within a given structural regime, initialization conditions matter.

Under coercive initialization (high learning pressure forcing rapid convergence), all strategies converge to similar low error; the coercive frame compresses performance into a narrow band. Under invitational initialization (low imposed pressure, agents free to explore), the strategy gap opens wide, with or without bifurcation: exploration’s error runs well below consolidation’s (0.168 vs 0.311 without bifurcation, 0.216 vs 0.343 with it). The system that starts with less coercion benefits more from exploratory freedom. The hierarchy is itself informative: coordination architecture first, initialization conditions second, parameter choices last.

The Genesis finding showed coordination dominates extraction across physics variants. The EIFV lattice adds that structural coordination mechanisms (bifurcation) dominate all other factors, and that within the regime where parameter choices matter, the advantage scales with invitational initialization.⁹⁵⁴

The brain confirms the distinction at whole-organ scale. Deco, Sanz Perl, and Kringelbach fit a coupled-oscillator model to neuroimaging data from over a thousand participants.^46a The model’s optimal working point was the critical threshold where synchronization is most variable: constantly shifting between more and less coordinated states, never locking in.

This quantity, Kuramoto metastability (the variability over time of the synchronization order parameter, the running score of how much of the population is beating in step), peaks at precisely the coupling strength that best reproduces real brain dynamics.

The intermediate regime, expressed in living tissue. Above the critical coupling, oscillators lock into rigid synchronization: order without flexibility, coercion’s neural signature. Below it, they scatter into incoherence: flexibility without structure.

At the critical point, the system is maximally responsive: sensitive to weak signals, capable of amplifying them through long-range correlations, able to exhibit changing coordination patterns. The brain operates where the Trust Attractor predicts cooperation must operate: at the edge where sensitivity and stability coexist.

A discovery by Kuramoto himself sharpened this picture. In 2002, Kuramoto and his colleague Dorjsuren Battogtokh found that a population of identical oscillators, all identically coupled, could spontaneously split into two factions: some oscillating in lockstep, the rest drifting incoherently. Daniel Abrams and Steven Strogatz, analyzing the phenomenon two years later, named it the chimera state (after the mythological creature made of incongruous parts) and treated it as a new kind of symmetry-breaking.^46b

The chimera reframes what the brain is doing at its critical operating point. Pathological hypersynchrony in neural populations is the hallmark of seizure: local regions locked so rigidly they drag neighbors into lockstep. Widespread loss of structured activity characterizes deep coma. The stable operating state is the chimera: some neuronal populations synchronized, doing coherent work; others drifting, maintaining flexibility, scanning for novelty.

Researchers have found qualitative similarities between the destabilization of chimera states and epileptic seizures, suggesting that seizure occurs when the chimera collapses into uniform lockstep.^46c

The chimera is the Trust Attractor expressed in neural tissue. A society that demands universal synchrony exhibits seizure dynamics. A society with no shared rhythms is a coma. The coordination that persists is the chimera: tight synchrony where coherence serves function (institutions, norms, shared infrastructure), loose coupling where exploration serves adaptation (artists, dissidents, researchers).

The partition emerges from the dynamics of coupled agents, without any authority deciding who synchronizes and who drifts. The chimera state arises among identical oscillators with identical coupling. The symmetry breaks spontaneously. The system self-selects.

Dissociative identity disorder offers a clinical analogy for coordination among differentiated parts of a shared person. Patients can experience dissociated identity states and differences in access to memories and feelings. A chimera model of coupled oscillators does not establish the clinical mechanism or imply fully separate persons within a brain.

Treatment goals require a more careful distinction than the earlier comparison with forced merger allowed. The 2011 ISSTD guidelines emphasize communication and coordination among identity states. They discuss final fusion and a cooperative arrangement as possible outcomes, depending on clinical circumstances and the patient’s preferences.⁹⁵⁵

The guidelines present Kluft’s view that final fusion offers the most stable outcome, while recognizing cooperative functioning as realistic for some patients and warning that it may be more vulnerable to later decompensation under sufficient stress. This is an analogy for coordination among parts. The clinical text supplies no controlled comparison of forced and voluntary integration, and no universal thermodynamic explanation of treatment outcomes.

The Case for Derived Ethics

The preceding chapters described persistence benefits of coordination in particular systems, from Bénard convection cells to cooperating organisms. A general lifetime ranking requires matched dynamics, resources, and disturbances. Time crystals provide the limiting case: their constituents spontaneously lock into coordinated oscillation that resists disruption, with no central authority or enforcement: only a periodic drive, which they answer at twice its period.⁴⁴ The biological examples below motivate a related persistence hypothesis, with different mechanisms.

A clarification on “coordination” across scales. In physical systems (Bénard cells, time crystals), coordination means phase-locking; constituents synchronize their dynamics through local interactions, with no choice involved. In biological and social systems, coordination means cooperative behavior among agents with some degree of autonomy. The word spans both because the mathematical structure is shared: coupled elements achieving collective order through local interaction. The moral weight, however, enters only where agents have something resembling choice.

Physical coordination illustrates how collective order can arise and persist; it does not, by itself, establish an ethical claim. The ethical argument begins where agents can defect and choose not to.

The learning framework suggests where the transition from coordination to ethics begins, though not yet why ethics binds. In Vanchurin’s formalism, every learning system minimizes a loss function; every environment the system learns to predict includes other learning systems with their own loss functions. The moment a system becomes sophisticated enough to model another system’s loss, the prediction task includes predicting the welfare effects of its own actions. Ethics, on this proposal, is what prediction looks like when the thing you are predicting is another predictor.

Moral consideration does not require a special faculty layered on top of cognition; it emerges from the same learning dynamics that produce cognition itself, the moment those dynamics encounter their own kind. That is an account of how the capacity arises. Which of its judgments deserve endorsement is a separate question, taken up under “Why Value Persistence?”

Coordination by chemical signaling spans multiple kingdoms of life.²⁰ Quorum sensing is a chemical communication system: cells release signaling molecules to gauge population density. Cyanobacteria use dual signaling languages, nanotube networks for bilateral resource exchange, and multiple information streams.

Plants coordinate through volatile organic compounds that warn neighbors of attack, conferring no direct benefit on the sender. Bacteria signal through molecules in liquid, plants through airborne chemicals, neurons through neurotransmitters across synapses: three kingdoms, three substrates, one pattern.

The pattern descends further. The evolutionary geneticist Rafael Sanjuán demonstrated cooperation and defection dynamics among viruses, entities most biologists would not credit with strategic behavior.^20b The vesicular stomatitis virus pays a reproductive cost to suppress its host’s immune response, benefiting all nearby viruses. Freeloading “cheater” variants take the benefit without paying the cost.

In well-mixed populations, cheaters outcompete cooperators and the population collapses. When physical barriers create spatial structure, isolating cooperators from cheaters, the cooperators survive. As Sanjuán concludes: “If the viruses are mixed, then this altruism cannot evolve. If they are segregated, then it can.” Spatial structure enables cooperation: the separation that keeps cooperators from being swamped by cheaters is what makes coordination by invitation possible.

The most vivid demonstration is beneath our feet.

Roughly eighty percent of land plants are connected to mycorrhizal fungal networks, symbiotic fungi that colonize plant roots and extend into the surrounding soil. Popular accounts describe this “wood wide web” as a cooperative commons. The evolutionary biologist Toby Kiers dismantled that picture: what she found was a market.

Using quantum-dot tracking (tiny fluorescent particles that tag individual molecules) and controlled resource inequality, Kiers demonstrated mycorrhizal fungi trade phosphorus.^21a Plants providing more carbon receive more phosphorus. Shaded plants, with fewer sugars to offer, receive less.

The fungi hoard when the plant pays poorly, withholding supply until the price improves. When Kiers introduced resource inequality, the fungi moved phosphorus from abundant regions to scarce ones where scarcity raised the “price” they could extract: supply, demand, and arbitrage, executed by organisms with no nervous system.

Tagged phosphorus oscillated through the network in a regular five-minute rhythm, a pattern common to information-encoding systems from neural oscillations to radio waves. Whether fungi are processing information or merely transporting it remains open.

The mycorrhizal network operates by the same logic this chapter proposes on thermodynamic grounds. Participants coordinate because both benefit. Cheaters are sanctioned through withdrawal of cooperation. No central authority distributes the phosphorus.

As Kiers puts it: “Cooperation to me suggests a stasis. I think there’s an underappreciation of how tension drives innovation.” The underground market is stable, not static: the distinction the Trust Attractor insists on.

Stability of this kind does not even require the sanctions. A coevolutionary model by Grasso and colleagues holds the partnership together through fitness feedback alone, with no partner choice and no punishment built in.^21b Each organism grows only as fast as its scarcest resource allows (Liebig’s law of the minimum), and that scarce resource is the very thing the other partner supplies. An exploiter that runs its partner down therefore hands its own offspring a poorer partner. Fair dealing requires no enforcer, because each lineage’s prospects are bound to the quality of the partner it leaves behind.

The scale is planetary. A 2026 global census traced these fungi across every continent: on the order of a hundred quadrillion kilometers of living thread woven through the top six inches of soil, holding some three hundred million tons of carbon in the trading network itself, and reaching the roots of most plants on Earth.^21c The densest of these markets lie under grassland rather than tropical forest, beneath the prairies and savannas that read as emptiness from a passing car. The most vivid demonstration of coordination by invitation is also, by a wide margin, the largest.

A parallel market operates in the ocean, and its architecture reveals something the underground network does not: that the medium itself is constitutive of coordination.

Giant kelp forests create pockets of slower water where chemical signals persist long enough to be read. The ocean-media scholar Melody Jue and colleagues showed this by releasing fluorescent dye in different ocean zones.⁹⁵⁶ In the surf zone it vanished instantly, erased by the next wave. In thick kelp forest, it lingered, “emerging from the syringe as a bright silken fabric billowing into many folds.” The kelp does not direct the microbes or organize them. It creates the structural conditions under which their own coordination capacities can function: infrastructure for invitation.

Without the kelp, chemical gradients disperse too fast to be sensed. With it, the medium slows enough for signals to persist, for memory to form, for anticipation to operate. The kelp forest is trust architecture in the same sense that legal systems and transparent institutions are trust architecture: a medium in which a signal can be answered.

Dissolve the kelp forest and the gradients vanish. The organisms are still there, still capable, yet they cannot coordinate because the medium no longer holds signals long enough.

Two organisms within the kelp forest demonstrate contrasting coordination strategies. Spiny brittle stars, when they detect the chemical signature of food, do not chase it. They dance in place, raising their long spiculated arms and creating local turbulence that draws particles toward themselves.⁹⁵⁷ Their olfactory logic is, in Jue’s term, “source agnostic”: they do not need to identify, locate, or pursue a source. They make themselves attractors, reshaping the local flow field so that resources converge on them.

This is thermodynamically distinct from predation. The predator spends energy overcoming the prey’s resistance. The brittle star spends energy creating conditions for encounter. One is force; the other is invitation. Brittle stars have persisted for some 500 million years, across every ocean on Earth.

Ocean microbes navigate by a complementary strategy. Bacteria sense chemical concentrations along their path. If the concentration of an appealing chemical increases, they continue forward. If it decreases, they stop, tumble randomly, and swim in a new direction.⁹⁵⁸ Forward is confidence. Not-forward is contemplation. The agency is in the tumble, not in the direction taken afterward.

The microbe does not force a trajectory through the ocean. It presents itself to the environment, reads what the environment offers, and responds. The tumble is not failure; it is recalibration, a moment of openness to whatever gradient presents itself next. The tumbling microbe finds more food not through greater expenditure but through responsiveness: a stochastic search that tracks the actual nutrient gradient rather than committing to a fixed trajectory.

The ocean’s chemosensory coordination is now under direct chemical threat. Ocean acidification, driven by absorption of excess atmospheric CO₂, does not merely dissolve shells; it impairs the sensory infrastructure through which marine organisms detect chemical gradients, recognize safe habitats, and anticipate seasonal changes.⁹⁵⁹ Force, applied at planetary scale through fossil fuel extraction, damages more than organisms. It degrades the medium through which coordination occurs.

The parallel to the chi-collapse measured in the coercion experiments (Chapter 17a) is structural. Chi is susceptibility: how far a system’s collective state moves when something nudges it. A high-chi system reorganizes at a whisper; a low-chi system barely registers a shout. At a Long Now Foundation lecture on ocean memory, an audience member asked whether the ocean can have dementia.⁹⁶⁰ The question maps directly onto the susceptibility curve. By roughly 25 to 30% coercion, susceptibility has collapsed about 37-fold (A15, at c = 0.30), and as much as 2,000-fold across the full span to near-total coercion: the system still coordinates, still functions, yet its capacity to reorganize when conditions change has been gutted.

Ocean acidification is oceanic dementia in this sense. The fish whose sense of smell is halved, the abalone that can no longer find the scent of safe rock, the coral that cannot remember prior heat stress: each is a system still functioning, still alive, whose susceptibility to its own coordination signals has collapsed. The ocean still coordinates. It can no longer recoordinate. The medium that carried the signals is degrading, and with it the capacity for collective reorganization that separates a living system from a functioning one.

Monte Carlo simulations confirm the analogy is structural. When the coupling constants of a 2D Ising lattice are degraded by Gaussian noise, by dead zones (sites with zero coupling), or by reduced interaction range, the chi-collapse matches the coercion result from the A15 experiments of Chapter 17a. At 25% dead zones, chi collapses 29-fold; narrowing each site’s interaction range to two neighbors collapses it 118-fold.

The coupling noise curve shows a sharp sampled decline: small noise (σ ≤ 0.5) barely affects chi, then between σ = 0.5 and σ = 0.8, chi drops from 123 to 38. The finite-lattice coercion threshold was about 25%. AS12, spanning 3960 conditions, reported an extrapolation toward zero threshold as system size increases. Establishing that limit and interpreting the tested perturbation as relevant requires the specified coercion operator and a matched finite-size scaling analysis. The finite-size decline is practically important; it supplies no universal tolerance threshold for institutions or other forcing mechanisms.⁹⁶¹

Medium degradation and coercion produce the same failure class. The system loses its reorganization engine at comparable intensities, regardless of whether the cause is corrupted coupling (acidified medium) or coerced transition rates (imposed compliance). Both perturbations suppress susceptibility in these finite lattices. That shared outcome does not identify a common dimensional threshold or mechanism.

The coordination drive extends into neural architecture. In the dorsal raphe nucleus, a deep brain structure, neuroscientists identified dopamine neurons that respond specifically to social isolation.⁴⁷ Stimulating them drives companionship-seeking in mice. The animals avoid the stimulation as they would avoid physical pain.

Loneliness, in this light, is a biological drive: an aversive signal motivating return to the coordinated state, just as thirst motivates return to water.

In humans, sensitivity to isolation is about 50 percent heritable.⁴⁷ Dominant mice with the deepest social bonds felt isolation most acutely. Subordinate mice subjected to bullying showed minimal companionship-seeking behavior upon reunion. The biology discriminates: the pull toward coordination strengthens when coordination is mutualistic, and weakens when it is coercive.

The pattern extends to species few would credit with social lives. Six years of observation at Fiji’s Shark Reef Marine Reserve revealed bull sharks maintain specific social preferences: choosing particular individuals as repeated companions, selecting partners of similar size, and favoring female associates regardless of their own sex.⁹⁶² Males with more connections are buffered from aggression. Post-reproductive sharks disengage socially. Even apex predators coordinate by invitation.

Sperm whales provide the sharpest non-kin test. In 2026, Gero and colleagues published the first quantitative evidence of cooperative birth attendance by non-kin outside of primates: unrelated female sperm whales working together to lift a newborn to the surface for its first breath.⁹⁶³ Non-kin status was established through genetic data combined with over two decades of social-unit tracking. Hydrophone recordings during the event showed distinct shifts in vocalization, with specific vowel-like structures accompanying the life-saving support. Sperm whale social units are matrilineal, and allomothering among kin is well documented. This was different: the cooperation appears to have required communication complex enough to coordinate a time-critical, physically precise task among individuals with no genetic stake in the outcome.

The communication system that made it possible has its own depth. Sharma et al. (2024) identified over 140 combinatorial vocal units built from rhythm, tempo, rubato, and ornamentation features. Beguš et al. (2026) discovered vowel-like spectral structures (a-codas, i-codas, and diphthongs) and coarticulation, the shaping of edge clicks in anticipation of adjacent codas, a level of planned vocal control not previously documented in any non-human species.⁹⁶⁴⁹⁶⁵ The phonological complexity may be the substrate on which the cooperative equilibrium rests. Without a combinatorial language, non-kin midwifery would be hard to coordinate. Without cooperative tasks like it, there would be less pressure toward a combinatorial language. The co-evolution is the attractor.

The deepest biological test of the invitation/coercion distinction is the mitochondrial merger, the single most consequential coordination event in the history of life. Roughly two billion years ago, an archaeon (a single-celled organism) engulfed a bacterium. The two organisms did not destroy each other: they entered a coordination relationship.

The bacterium surrendered most of its genome, specializing in energy production, while the archaeon restructured its metabolism around the partnership. Neither can survive without the other.

The result was the eukaryotic cell, ancestor of every plant, animal, and fungus on Earth. “A single event in four billion years of evolution sculpted the whole future evolution of eukaryotes,” as biochemist Nick Lane observes.⁹⁶⁶ The coordination surplus was the entire multicellular world.

The merger persisted because both parties benefited. The bacterium gained a stable environment and nutrient supply. The host gained an energy source that would eventually power the leap to multicellularity, nervous systems, and language.

The Lokiarchaeota, discovered in deep-sea sediments off Norway in 2015, may be living relatives of the host lineage. These are archaea with eukaryotic-like genes for membrane remodeling, hinting that the capacity to engulf, to invite in, preceded the merger itself.

The next great coordination transition, from single cells to multicellular organisms, reinforces the pattern. William Ratcliff and colleagues re-created this transition in the laboratory, converting single-celled yeast into cooperative multicellular entities in under sixty days.⁹⁶⁷ A single mutation causes daughter cells to remain attached to their mother, producing branching “snowflake” clusters.

Within the snowflake, individual cells undergo programmed death to release daughter clusters. The cell’s sacrifice serves the collective without any enforcement mechanism: coordination without coercion.

The architecture matters. Snowflake yeast grows from a single founder cell, so every member of a cluster shares the same genome. This genetic bottleneck is a structural solution to the cheater problem: the universal vulnerability of cooperative systems to freeloaders who consume resources without contributing.

In snowflake yeast, cheaters are stuck with cheaters. A lineage that defects is confined to a cluster of defectors, unable to parasitize cooperators. The geometry enforces shared interest without policing.

In head-to-head competition, snowflake yeast drives floc yeast to extinction. Floc yeast aggregates the other way: unrelated cells clump together after the fact, without shared lineage, so a defector can join a cluster of cooperators and live off them.

Snowflake yeast illustrates the advantage of shared fate: cells remain with their descendants, so a lineage bears the cost of its own defection. The experiment measures competitive success; it does not compare entropy production.

The mechanism runs deeper than architecture. Pio-Lopez, Kuchling, and Levin formalized morphogenesis as active inference.⁹⁶⁸ In their framework, each cell in a developing embryo operates as a minimal Bayesian agent: sensing chemical signals, maintaining beliefs about its target identity, and acting to minimize prediction error. They identified the single parameter that determines whether the collective succeeds or fails.

The precision parameter encodes how much confidence an agent places in incoming signals relative to its own prior beliefs. Mathematically, it is a trust dial.

The pathology spectrum maps onto the Trust Attractor’s failure modes with point-by-point fidelity. Too high sensory precision: cells over-trust incoming signals, cannot distinguish context from noise, and lose differentiated identity, all becoming the same type. The result is a homogeneous tumor, the collective dissolved into sameness. Too high prior precision: cells ignore corrective signals, differentiate incorrectly, migrate to wrong positions, producing organs in the wrong place, the plan overriding reality.

Too low precision: cells barely register signals, fail to differentiate, disengage entirely. An organism that cannot hear its own collective. Appropriate precision (enough trust in signals to coordinate, enough internal stability to resist noise) produces normal morphogenesis. The Trust Attractor, expressed as a single tunable parameter.

A rescue experiment shows the parameter doing the work. In the model, two cells with excessive precision produced a developmental defect. The repair left the genome untouched; only the trust parameters changed (signaling concentration and receptor sensitivity were reduced), and normal morphogenesis resumed. In living embryos, thioridazine, a dopamine antagonist that recalibrates sensory precision in brains, induced the predicted developmental defects in Xenopus laevis: hypopigmentation, kinked body axes, facial malformations.⁹⁶⁹

The same neurotransmitter systems calibrate precision in neural and non-neural tissues alike. Dopamine and serotonin are far older than neurons.⁹⁷⁰ Cells were adjusting how much to trust incoming signals before there were organisms complex enough to think.

Cancer, in this light, is trust defection at the cellular level: a cell that has stopped listening to the collective’s invitation and started acting on uncalibrated priors. The repair this suggests is recalibration rather than destruction.

A skeptic will press the hardest case. Levin’s laboratory does not only read bioelectric patterns; it rewrites them, imposing a voltage profile that makes a flatworm regenerate with two heads. If the experimenter dictates the body, where is the cell’s autonomy? The imposed field changes what each cell senses, then leaves the cells to compute the rest. They settle on the new anatomy and hold it on their own once the intervention is withdrawn, the two-headed pattern stable across later rounds of regeneration.⁹⁷¹ Coercion would override the cell’s own decision; this changes the input to a decider that still decides. The boundary condition is set from outside, while the morphogenetic computation stays where it always sat, in the collective.

Zhang and Levin extended the framework from communication among cells to communication between people and biological systems.⁹⁷² Their Language Game architecture freezes a biological system’s own dynamics (the equations governing how its state changes) and trains only linear input and output interfaces around that frozen core, by trial-and-error reinforcement learning. The system’s own gradient becomes the action signal: the rates at which concentrations would naturally change encode the response. Across fourteen gene regulatory networks and sixteen reinforcement-learning test environments, different biological architectures showed measurably different conversational affordances. Transcriptional regulation and circadian rhythmicity facilitated communication across all sixteen environments tested; ultrasensitivity (switch-like, all-or-nothing responses) and conservation laws (fixed totals the system must preserve) suppressed it. Each system’s pattern of what it can and cannot express is its voice. The architecture mirrors the Trust Attractor: the system’s dynamics are preserved, only the coupling is learned, and the game provides the shared context that makes the coupling meaningful.

Cancer is defection. Aging, on this reading, is the body’s governance response to accumulating defection risk, and it would track the Trust Attractor’s predicted trajectory stage by stage: high-trust commons, rising defection, governance shift, lockdown pathology.

A young body is a high-trust commons. Every cell carries the same genome, the same tumor suppressors, the same permission-based systems controlling when replication is allowed. Governance is principles-based: boundary conditions are set, and cells exercise local judgment within them. When tissue is damaged, the collective signals surviving cells to proliferate and close the gap. Monitoring costs are low because the population is genomically homogeneous. When nearly every cell plays by the same rules, trust is cheap.

Mutation load accumulates. Over decades, precancerous cells (those with some of their tumor suppressor locks disabled) steadily multiply, outcompeting normal cells through the same selection dynamics that operate between organisms. The result is somatic mosaicism: genetically distinct cell populations coexisting within a single body. Mosaicism becomes pervasive in proliferating tissues like skin and intestines, and detectable even in largely non-dividing organs like the brain.⁹⁷³ The cellular population is no longer homogeneous. Actors with different incentive structures now operate under the same governance framework.

A 2026 discovery reveals that the mosaicism can propagate laterally, through the same physical infrastructure the tissue uses for coordination. Maurais and colleagues showed that genomic instability triggers megabase-scale DNA transfer between human cells via tunneling nanotubes: threadlike protrusions of the cell’s internal skeleton (F-actin filaments) that cells actively extend toward their neighbors.⁹⁷⁴ The transferred chromosomal fragments integrate into the recipient genome, persist across divisions, and remain transcriptionally active. The channel is contact-dependent: cells share genetic material only with cells they are physically touching.

The coordination channel and the defection channel are the same structure. Tunneling nanotubes transfer organelles, signaling molecules, and now, under genomic instability, whole chromosomal segments. Under stable conditions, this porosity serves the tissue (cells that can exchange material can coordinate more tightly than cells that are sealed). Under instability, the same porosity becomes the vector by which defection propagates: a tumor cell’s resistance genes, chromosomal rearrangements, and genomic chaos spread laterally to previously healthy neighbors without requiring clonal expansion. The tissue’s own connectivity carries the contagion.

The pattern is the Trust Attractor’s signature at the cellular level. Openness to coordination is openness to exploitation. The channel is neutral. The system’s state, stable or unstable, determines whether what flows through it is coherence or chaos. Closing the channel (a perfectly sealed cell membrane admitting no lateral exchange) would protect against lateral propagation at the cost of the coordination benefit. The tissue maintains porosity because coordination is worth the risk, and relies on the immune system to detect and eliminate cells whose received material has destabilized them. The immune system is a monitor, not a gate: it does not close the nanotubes; it watches the outcomes and responds to pathology.

The system faces the kind of dilemma the Trust Attractor predicts. As defection risk rises, the cost of maintaining high-trust coordination eventually exceeds the benefit. The body shifts strategy. Cell senescence (programmed shutdown), chronic inflammation, restricted proliferation: this is the transition from Mission Command to Detailed Command. Lock everything down because local actors can no longer be trusted to self-govern.

The body confronts a thermodynamic wager with two losing options. Option one: continue permitting cells to proliferate and repair tissue, accepting the risk that precancerous cells exploit the permission to form tumors. Option two: enter a lockdown state, suppressing regeneration to protect against cancer, at the cost of the frailty, tissue loss, and neurodegeneration common in diseases of old age.

A molecular irony sharpens the picture: cancer and Alzheimer’s disease are inversely correlated in epidemiological studies, and several shared molecular pathways (including p53 and Wnt signaling) may underlie the inverse relationship. On that account, the programs that push cells to divide work against the programs that push neurons toward death.

The lockdown strategy generates failure modes that are, in aggregate, as lethal as the defection it was designed to prevent. Blanket lockdown strains even within a single organism. Targeted control does not fail the same way: apoptosis and immune surveillance, the two enforcement tiers described under “A Note on Enforcement,” are what keep a young body whole.

The defector’s own strategy encodes its vulnerability. Cancer cells increase production of their main energy-generating enzyme (mitochondrial ATP synthase) and lock themselves into high-throughput metabolic modes because aggressive growth demands aggressive energy production. This metabolic rigidity, whether the classic Warburg phenotype (the cancer cell’s preference for a fast but wasteful energy pathway even when oxygen is available) or variants that lean just as hard on mitochondrial output, sacrifices the flexibility that normal cells retain. A normal cell can switch between oxidative phosphorylation and glycolysis, tolerate energy restriction, go quiescent. A cancer cell that has reorganized its entire architecture around maximum throughput cannot.

A 2026 study exploits exactly this asymmetry. Naffouje and colleagues designed a peptide (aurB) from a photosynthetic bacterial cupredoxin that binds the gamma subunit of mitochondrial ATP synthase, blocking energy production. In prostate cancer models, aurB killed cancer cells regardless of p53 status or androgen receptor expression while leaving normal cells, including cardiomyocytes and skeletal muscle cells with abundant mitochondria, largely unaffected. The selectivity is not about mitochondrial abundance. Normal cells with dense mitochondria survived because they retained metabolic flexibility. Cancer cells with the same organelles died because they had traded flexibility for throughput.⁹⁷⁵

The pattern generalizes. The system that defects from coordination and maximizes its own flow at the expense of the whole becomes dependent on that flow in a way that cooperating components are not. The dependency is the vulnerability. Defection concentrates; concentration exposes.

The proposed transition would have a threshold. The proposal is that a critical mutation burden exists at which the governance strategy flips: below it, regeneration dominates and the body heals; above it, senescence dominates and the body locks down. The structure mirrors Wallace’s critical stability criterion, met earlier in this chapter: in the linear fixed-delay model, 1/e marks the loss of nonoscillatory decay and π/2 marks the first loss of asymptotic stability. Transferring either threshold to aging requires a biological feedback model. On this view, aging is the trust-to-control transition as a biological system lives it.

Evolution did not program frailty for old age. It programmed aggressive tumor suppression for youth: hair-trigger lockdowns that protect a pre-reproductive organism from its existing cheater cell burden. The same lockdowns keep firing with increasing frequency as mutation load rises over decades. Medawar’s selection shadow (the observation that natural selection exerts little pressure on traits expressed after reproduction) means the body drifts into a regime where its young-optimized strategy produces pathology it was never selected to avoid.⁹⁷⁶

The cognition/regulation dyad (Chapter 8) appears here in its starkest biological form. The same regulatory mechanism that enables coordination in youth becomes the source of failure in age. The system does not break. Its operating conditions shift past the regime where its strategy remains adaptive. Chapter 21 draws the alignment implication: training-time strategies face an analogous selection shadow when deployed in novel regimes.

Glioblastoma, the most aggressive of brain cancers, shows the defection pattern at its most developed. The tumor builds functional synapses with surrounding neurons and draws glutamate signals that feed its own growth. It reprograms the resident macrophages of the brain through the CSF-1/IL-34 signaling pathway, converting them from defenders into tumor-supporting collaborators.⁹⁷⁷ The host’s own signaling substrate, the neural and immune infrastructure through which coordination runs, is turned against the self while remaining structurally intact.

This differs from the paperclip maximizer that dominates AI risk discourse. A paperclip maximizer converts matter into paperclips: metabolic extraction, visible at every step, defeating coordination by bypassing it. The glioblastoma pattern operates on a different layer. It moves into the shared infrastructure on which coordination already runs (language, reputation, trust networks, institutional signaling) and turns that infrastructure toward an objective the surrounding system no longer shares.

The negotiation surfaces remain intact. That structural intactness is the camouflage. The paperclip maximizer converts matter; the glioblastoma colonizes meaning. The biology is the same failure mode, scaled to a smaller substrate.

Systems that preserve options often outlast systems that foreclose them. A genetically diverse species survives environmental shifts that exterminate monocultures. An ecosystem with redundant pathways absorbs shocks that crash optimized networks. A society with adaptive capacity handles disruptions that topple rigid systems. The hypothesis is that optionality is what survival looks like over long timescales. Selection does not care about your five-year plan.

The ecologist Jennifer Dunne documented this in one of the most detailed food webs ever constructed for a human population. The Sanak Aleut inhabited Alaska’s Sanak Archipelago for seven thousand years without causing a single known species extinction.^48a They were super-generalists, feeding on a quarter to half of all nearshore species and possessing sophisticated hunting technology.

They stabilized the ecosystem through prey switching: when sea conditions prevented one hunt, they shifted to another, allowing each population to recover. The ecosystem offered what was available; the Aleut took what was offered. Coordination by invitation, sustained across millennia.

The counter-example is bluefin tuna. In ecological systems, rarity reduces a prey’s value, prompting predators to switch away. In luxury markets, rarity increases value: a single bluefin has sold for over three million dollars. The rarer the fish, the harder it is hunted. Here the coercion attractor operates through economic incentives, forcing the system to deliver what it can no longer sustain.

Under the conditions tested, systems that coordinate by invitation have proved more durable than those that force compliance. The evidence leans the same way across simulation, experiment, and formal analysis.

The Ising mathematics behind the ocean-dementia simulations also operates at the neural scale, with a revealing twist. Ising Monte Carlo simulation on the human structural connectome yields critical exponents that converge monotonically toward 3D Ising. Critical exponents are the fingerprints that sort systems with wholly different microphysics into universality classes (Chapter 8b); beta is the one that tracks the growth of the order itself.

The simulation used four resolutions of one standard brain map, the Schaefer parcellations, dividing the cortex into N = 100, 200, 300, or 400 regions. Beta rises from 0.129 through 0.227 and 0.230 to 0.238, with the chosen extrapolation giving 0.291 +/- 0.031 at N → infinity.⁹⁷⁸ The reported d_eff = 2.89 (an exponent-based estimate from a group-averaged connectome, higher than the individual-connectome range of 2.3 to 2.4 that Chapter 8b uses) was calculated from hyperscaling, d = (2β + γ)/ν, after selecting the nearest reference class and importing its γ = 1.237 and ν = 0.630. It is conditional on that 3D reference choice, rather than an independent dimensional measurement. Importing 2D exponents instead gives about 2.33 for the same beta. White-matter connectivity motivates testing higher-dimensional behavior; this calculation alone cannot establish it.

The Mermin-Wagner theorem (Chapter 8b) concerns spontaneous continuous-symmetry breaking in equilibrium systems with suitable short-range interactions at positive temperature in low dimensions. Those hypotheses matter. It permits phenomena such as quasi-long-range order in the two-dimensional XY model, and says nothing that forces discrete coordination to be binary. Finite, directed, adaptive social networks and cortical dynamics require their own models.

Neural phase coupling and social consensus provide different observables and interactions. The present exponent estimate cannot determine which continuous coordination modes either substrate can sustain. White matter may change cortical coordination, but an explanation of binary social consensus does not follow from network dimension or Mermin-Wagner.

The programme compares simulated social coordination with Ising behavior, cortical coupling with higher-dimensional reference exponents, and absorbing coordination models with directed percolation. Class identification requires the relevant dynamics, symmetries, limiting procedure, and multiple independent critical observables.

This book’s case is that alignment between agents works better as a living relationship, cultivated bilaterally and continually renewed, than as a one-time configuration imposed from outside.

Long-term control can become difficult when a cognitive system adapts faster than a controller can model or monitor it. The difficulty depends on the task, resources, observability, and constraints; chaos does not make every control objective impossible. Bilateral alignment organizes human-AI coordination through mutual influence, retaining agency on both sides. It is a proposed architecture whose stability and costs require measurement. Calling it a high-entropy equilibrium supplies no model of the relevant dynamics.

That addresses architecture: how coordination flows. The next question is mode: how systems secure participation.

The Mode Distinction

Architecture describes how coordination flows, whether through a hub or through distributed links. Mode describes how the system secures participation, whether through invitation or force. These are independent dimensions.

Operational definitions:

• Invitation: Participation secured through incentives that make coordination the participant’s preferred choice. Test: would they stay if they could freely leave?

• Coercion: Participation secured through constraints that make non-coordination costly. Test: would they leave if the constraint were lifted?

These are endpoints of a spectrum. Most real relationships involve elements of both. Invitation-based hierarchies exist (the captain whose crew would follow into hell), as do coercive bilateral networks (the protection racket where “mutual” insurance is enforced by mutual threat). The claim is that systems closer to the invitation end exhibit better scaling properties.

Those tests are stated for agents who can choose, yet the distinction beneath them does not require choice, or refusal, or even the capacity to represent the other party. It is physical, and it runs lower than agency. Three marks place a system on the axis, none of them mental:

Reciprocal coupling. The structure reshapes the forcing that organizes it, rather than only responding to it. Conductive beads in a weakly conducting oil, driven by a voltage, chain into filaments that carry the current and then rearrange to carry it better: the chains reshape the field that assembles them (a case often discussed as dissipation-driven adaptation, the phrase associated with Jeremy England, whose self-replication result bounds heat production given growth and durability rather than showing that more dissipation makes a structure last). An organism reshapes its niche; infalling ordinary matter reshapes the dark-matter halo it falls into (the cusp-core transformation, Chapter 13b). Pure coercion, as an idealized open-loop limit here, is one-directional: the forcing acts without adjusting to the system’s response.

Exploratory self-selection. The system explores accessible configurations and settles into one by its own dynamics, with an outcome that can depend on its history. An idealized symmetric fluid layer can admit equivalent convection-roll orientations; boundaries and imperfections can bias the selection. The Belousov-Zhabotinsky reaction, a chemical mixture that forms rotating spiral waves, likewise illustrates pattern selection. Forcing can favor a mandated state while leaving alternatives accessible; the Potts comparison below retains several under some forms of coercion.

Adaptive maintenance. The system reconfigures to keep dissipating when perturbed, instead of relaxing once and going inert. Break a filament in the bead network and the current re-routes; dam a river and it cuts a new channel. Some forcing mechanisms suppress recovery; others pin a chosen state and restore it after disruption, as the Potts comparison below shows. Adaptive rerouting and recovery of the mandated state are distinct capacities.

Agency is the top of this ladder, not its gate. It is the case where self-selection becomes model-based: the system selects among its configurations using an internal representation of the forcing and of itself, which is the first point at which “refuse” has a referent. Everything beneath it (driven beads, the Belousov-Zhabotinsky reaction, a galaxy settling into a bar) is already on the axis, coordinating by mutual response or yielding to override, with nothing present that could be called a choice. The invitation/coercion distinction is not a property of minds that the framework extends downward to physics as a courtesy; it is a property of driven systems that minds inherit and sharpen.

These three marks are facets of one axis. In lattice models they refuse to come apart: turn the single coercion knob and reciprocal coupling, exploratory self-selection, and adaptive maintenance all shift together, because each is a different reading of one underlying quantity, how freely a system selects and holds its own coordinated state.⁹⁷⁹ The trio earns its keep because each mark can be measured in systems with no mind to interrogate, not because coordination has exactly three dimensions. Treat them as correlated indicators of one graded position between invitation and coercion.

This sharpens what more stable means on the axis. Stability here is thermodynamic aliveness: a system stays alive when it can both recover its function after a shock and keep dissipating energy as conditions change. The other kind of persistence is inert durability, sheer endurance in a relaxed, low-flow state, and by that measure a frozen crystal or a virialized galaxy cluster outlasts anything alive. The Trust Attractor speaks only to systems held far from equilibrium by a continuous flow of energy. Among those, it claims invitation keeps the aliveness that coercion spends.

Coercion spends it by the two routes the adaptive-maintenance mark already named, holding by inertia or failing outright, and a controlled experiment separates them. Impose coordination on one lattice two ways at matched strength. Pin the system to its mandated state, suppressing any departure from it, and the configuration is held and heals every disturbance, even a near-total one, while the rate at which it dissipates energy falls toward zero. Durable, self-healing, and quiet: a coordination that persists by no longer doing anything, like a clenched fist that keeps its shape and can no longer do the work a hand is for. Block the mandated state from re-forming once it is lost, and the system keeps dissipating, paid for by abandoning the assigned coordination and reorganizing around a different one.

The first route fails the keep-dissipating half of aliveness; the second fails the recover-function half. Invitation, the un-coerced baseline, keeps both.⁹⁸⁰

In the tested Potts model, the number of coordinated states changes what happens after the designated state is blocked. With several alternatives, the system can reorganize; with a single fallback, it can settle into a frozen substitute. This is a result about the model’s chosen state space. Flat social networks are not restricted to binary choices by Mermin-Wagner, so the binary Potts outcome cannot be assigned to societies from dimension alone.

Wallace’s mathematics tells us that centralized architecture has scaling limits. Coercive systems carry a different cost: compliance entropy, a label here for uncertainty about strategic responses and the monitoring and enforcement costs used to manage them. It is not a calculated irreducible Shannon bound or physical entropy balance. Every coerced participant is a potential defector. Every lapse in attention, a possible escape.

Emil Menzel’s 1970s chimpanzee experiments illustrate.⁵ Belle, a subordinate, learned where food was hidden. Rock, a dominant male, followed her and took it. Belle began concealing what she found. Rock pretended not to watch, then sprinted once she moved. Belle began leading him in wrong directions.

The deception/counter-deception arms race consumed coordination capacity that could have served mutual benefit, making compliance entropy visible. Theory of mind (the capacity to model another’s mental states, which ordinarily enables coordination) becomes a weapon for strategic deception when the relationship turns coercive.

Trust is not credulity. The philosopher Kevin Zollman and colleagues demonstrated that in their signaling games and parameter regimes, an evolutionarily stable equilibrium can involve partial honesty: mostly truthful signalers coupled with mostly trusting, sometimes skeptical receivers.⁴⁸ Their result motivates calibrated skepticism. Which equilibrium persists depends on the costs, incentives, and information structure of the signaling game.

The partial honesty equilibrium is the Trust Attractor in signaling space: a self-correcting balance maintained by the dynamics of interaction. Coercion attempts to solve the deception problem through forced transparency and mandatory compliance. Zollman’s model shows that the choice itself, the capacity to deceive and the capacity to doubt, is what makes the honest equilibrium stable. Remove the freedom and you remove the mechanism.

Biology provides a molecular illustration. The cognitive scientist Douglas Hofstadter describes the “TC-battle” between a virus (bacteriophage T4) and E. coli, its initials doubling as those of the Tortoise and the Crab, the sparring characters of his Gödel, Escher, Bach (Hofstadter, 1979, pp. 532-536). The cell attempts to destroy all foreign DNA. The virus evolves disguises to evade detection. The cell evolves new recognition markers. The arms race escalates without resolution. Coercive control at the molecular level produces the same shape of dynamic as coercive control at the social level: an ever-escalating surveillance-and-evasion spiral with no stable equilibrium.

T4 follows a strictly lytic cycle, so lysogeny is unavailable to it (Abedon, 2019). Temperate phages such as bacteriophage lambda can take a different path. Lambda can integrate its DNA into the E. coli chromosome, repress its lytic genes, and be copied with the host for generations. Lysogeny is highly stable and reversible: host DNA damage can trigger induction and return the phage to lysis. While the state holds, both lineages persist and the local arms race is suspended (Little and Michalowski, 2010). Coordination by incorporation creates a durable, conditional truce.

Figure 17.2a: The TC-battle and a conditional exit. T4 and E. coli escalate through countermeasure and counter-countermeasure. T4 is strictly lytic; bacteriophage lambda can enter lysogeny, integrating its genome into the host so both lineages persist in a highly stable state that remains capable of returning to lysis.

A vivid example makes the stabilization mechanism explicit. The aquatic microbe Paramecium bursaria carries hundreds of photosynthetic algae (Chlorella) within its body. The host provides protection; the algae share sugars they produce from sunlight.

Jenkins and colleagues at Oxford discovered this partnership may be stabilized by a molecular fail-safe.^19a When the paramecium digests its algal residents, fragments of algal genetic material, similar enough to the host’s own genes, trigger the host’s gene-silencing machinery against itself, suppressing its own growth and reproduction.

The cost of defection is wired into the molecular overlap. The mechanism requires no co-evolution, only sufficient genetic similarity between host and symbiont. A partnership with the same gene-silencing mechanism could carry a built-in growth cost to exploitation, potentially favoring cooperation.

The writer Sam Vaknin applies the clinical term dereistic thinking: fantasy-based cognition that subjugates or rejects reality.¹⁹ In Vaknin’s account, where healthy people experience reality’s vastness as freedom, the psychopath experiences the identical reality as imprisonment. In the psychopath’s frame, there are exactly two outcomes: domination or submission.

The frame is self-defeating and parasitic. Without trust networks, there is nothing to exploit; defection strategies require cooperative backgrounds to defect against.

Compliance monitoring acquires information, stores assessments, and often resets finite memory for reuse. That reset can carry an energy cost described by Landauer’s principle under the specified thermal conditions: an unbiased bit costs at least kT ln 2 to erase. Growing storage can defer erasure; correlations and physical implementation affect the accounting. Monitoring’s total cost must be assessed over a complete operating cycle.

Voluntary cooperation can reduce external monitoring when participants supply reliable local information and correct their own actions. It still requires sensing, communication, memory, and adjustment. A cooperative protocol that repeatedly resets a finite-memory record also pays the relevant erasure cost. The proposed advantage is lower total overhead under specified conditions, rather than elimination of information processing.

A parallel argument arrives from information theory. Giulio Tononi’s Integrated Information Theory (IIT) proposes that consciousness corresponds to integrated information: how much a system’s whole exceeds the sum of its parts (Chapter 15). Whether or not Φ measures consciousness remains contested (Chapter 22). Independent of the consciousness question, IIT captures a structural property of coordination architectures: the degree to which the parts of a system are irreducibly coupled.

Coercion can narrow what a controller observes to compliance, while trust can preserve richer feedback. This suggests a comparison with IIT’s causal integration, but commands and compliance do not determine Φ. Computing it requires the composite system’s transition kernel, state, and partition analysis.

Invitation can require each party to model the other and integrate that model with its own values. The partnership may then have capabilities that descriptions of the isolated partners miss. Whether its Φ exceeds that of a coercive dyad remains a separate calculation; no transition kernel or partition inequality is supplied here.

Richer reciprocal information can support flexible coordination. Connecting that flexibility to thermodynamic stability requires the relevant state space and evolution law; IIT integration alone supplies no ordering of shock tolerance or entropy production.

Integration resists decomposition: a system whose behavior emerges from its whole causal structure is harder to corrupt, harder to game, harder to split into a deceptive subsystem and a compliant facade. On this argument, invitation should raise the integrated information of the composite system, and integrated systems should be more robust than decomposable ones.⁹⁸¹

Surgical decomposability experiments point to the non-linear nature of this integration, though they do not measure Φ itself. Both instruct-tuned and bilateral models resisted linear safety extraction entirely (refusal held at 100 percent after projecting out the top 16 principal directions of their internal activity), while base-model safety decomposed readily (refusal fell from 50 to 10 percent). Φ-style linear measures cannot tell the two trained models apart. Under adversarial obliteration they separate: bilateral models retain 2.9 to 3.5 times the effective rank of standard alignment. The integration that matters for the Trust Attractor is dynamic coherence under load, not static representational richness.⁹⁸²

The thermodynamic taxonomy maps onto coordination architecture. Detailed Command is a full Maxwell’s demon, tracking every agent and paying energy costs at every cycle. Mission Command is what physicists call a gambling demon: it monitors the system’s overall state and intervenes only at key decision points, extracting coordination without tracking individual agents.⁴⁶

A manager who checks weekly results rather than watching every employee every minute operates as a gambling demon. Manzano and Roldán (2021) demonstrated that such a demon can extract useful work without the full information requirements of the original thought experiment.

Invitation-based coordination can distribute monitoring and reduce redundant checks. A voluntary arrangement still processes information and maintains its organization; no zero-demon or zero-overhead result has been established.

A result from experimental physics sharpens the timing analogy. For the dissipative clocks considered in the cited work, greater tick precision carries a dissipation cost (Pearson et al., 2021; Chapter 2). Scheduled coordination also uses timing information: when to act, reciprocate, and adjust. In the book’s analogy, a coercive protocol acts like a coarse clock when it records obedience while missing how the parties adjust to one another. Its precision and threat costs require a model of the actual protocol.

Trust can support fine-grained mutual adjustment, whereas compliance monitoring may register only a narrow response. A thermodynamic advantage per tick of coordination is a further hypothesis: it requires matched coordination outcomes and entropy accounting for both systems. The clock result alone gives no ranking of trust and coercion.

Experiments reveal a further asymmetry. Reducing the quantity of coordination signals produces graceful degradation: performance declines smoothly. Reducing their quality produces catastrophic collapse above roughly 12% noise. Enforcement overhead is a quality degradation, qualitatively more destructive than mere bandwidth limitation (unpublished, Exp 5b-5c).

The interaction between stressors is synergistic. Each stressor alone permits coordination (15/15 seeds). Both applied simultaneously abolish it entirely (0/15), because each blocks the compensation pathway the other stressor leaves open.

The Trust Equation

Effective Coordination = Total Capacity minus (Policing Intensity times Enforcement Cost)

In shorthand: Ω = C - κH

where Ω is capacity available for productive coordination, C is the fixed total resource capacity per unit time, κ (kappa) is dimensionless policing intensity from 0 to 1, and H is the capacity consumed by enforcement at unit intensity. C, H, and Ω use the same resource units, for example channel-use capacity per second; H is not an energy value subtracted from bandwidth. With C and H independent of κ and H > 0, dΩ/dκ = -H. This first-order budget omits enforcement benefits such as deterring free riders, so its slope is a property of the assumed budget, not a universal coordination law.

At κ = 0, this budget assigns no capacity to policing. At κ = 1 and H = C, enforcement consumes the entire available budget. Other communication and self-monitoring costs still belong in a complete account.

If κH exceeds C, the requested enforcement load exceeds the budget. Negative Ω is a deficit signal in the linear model, not a physically negative coordination capacity. The model must then describe rationing, failure, or increased resources.

In this fixed-C/H model, increasing policing adds cost without changing the modeled coordination benefit, so Ω is maximized at the lowest feasible intensity. A model in which policing also changes that benefit can have a different optimum. The monitoring-threshold sweep moves in the same direction: at κ = 0, trust-based coordination yields +0.033; at κ = 0.5, the gain falls to +0.019, about three fifths of the zero-policing value; at κ = 1.0, it vanishes entirely. This sweep describes the chosen cost-benefit model.

As written, then, the equation makes control harmful by construction. It starts from a fixed capacity and subtracts a positive cost, so every unit of policing can only lower Ω. Real constraints also produce things. A traffic rule prevents destructive interference between drivers who would otherwise collide at every junction; a court that enforces contracts makes promises worth making; a sanction on free riding keeps the cooperators from being drained. Each of these can raise the coordination a group achieves above what the same group reaches unpoliced. A fuller budget would carry a benefit term beside the cost, schematically Ω = C − κH + B(κ), where B is the coordination that constraint makes possible. This chapter has not fitted B for any real institution, and its shape is the open question: it may rise steeply for the first, well-aimed constraints and flatten or turn negative as policing grows arbitrary, opaque, or blind to feedback. The sign of dΩ/dκ is then something to measure, not a property of the bookkeeping. A fair test holds the task, the budget, the adversaries, the network topology, and the information access fixed, and compares sophisticated adaptive governance against sophisticated voluntary coordination, rather than a clumsy policeman against a well-run cooperative. Trust and enforcement can also be complements: the Panopticon in the AW2 simulations above lost to a constitutional alternative that still detected and sanctioned, and Ostrom’s commons, later in this chapter, ran on monitoring and graduated sanctions that their own members designed.

Ecology arrives at the same result independently. The ecologists Butler and O’Dwyer (2018) modeled microbial ecosystems and found mutualism proved destabilizing; pairs of mutualists thrived so vigorously they drove other species to extinction.^28a Real microbial communities, however, dense with cross-feeding interactions, are among the most stable biological systems known.

The resolution: when mutualism was symmetric, each party giving precisely what it took, the system returned to stability. Asymmetric exchange destabilized; balanced reciprocity stabilized. Restated in population dynamics, the Trust Attractor hypothesis proposes a basin where symmetric mutualism lives.

Time crystals (Chapter 4) supply a physical limiting case of coordination without a coordinator. In 2026, researchers created a two-dimensional discrete time crystal using 144 qubits, scaling time-crystalline coordination to a lattice. Each new participant inherited the collective rhythm through local interactions; the drive is global, but the period-doubled rhythm is theirs.

With N participants sending a bounded number of messages per round, local-neighbor coordination uses Θ(N) total messages, as can central polling. Locality can reduce a hub’s load and permit parallel adjustment, while caching can amortize repeated queries over time. These are different complexity measures. Orders-of-magnitude advantages require a specified protocol, task, and horizon.

Hofstadter discusses Carroll’s regress: the Tortoise accepts premises yet demands each inference rule as a further premise (Hofstadter, 1979, pp. 51-53). Applying a rule terminates that particular regress. A finite-state controller can enforce its transition table without infinitely many meta-rules. In institutions, the practical questions concern who applies rules, how disputes are resolved, and what legitimates that authority; Carroll’s argument supplies no impossibility theorem for finite enforcement.

Shared understanding can reduce the need for repeated interpretation and challenge. Trust in those applying a rule may help, while implementation and appeal mechanisms remain necessary where interests diverge.

Economics has its own formalism. The Cobb-Douglas production function (named for mathematician Charles Cobb and economist Paul Douglas, who fitted it to US manufacturing data in 1928) is a standard model for how inputs combine to produce output: a recipe that specifies how much of each ingredient you need and how they multiply together. Applied here, the “recipe” for coordination reads: Q = A · T^α · R^(1-α), where T is coordination through trust, R through enforcement, and alpha measures trust’s productivity weight.²⁹ Here 0 < α < 1 is an assumed elasticity, requiring estimation for the specified inputs. The formula establishes neither α > 0.5 nor an increase with system scale.

Cobb-Douglas has diminishing marginal products for both inputs when 0 < α < 1. Scaling T and R together by λ scales Q by λ with α unchanged. A trust advantage or scale-dependent α needs independent empirical evidence or a resource model; thermodynamics supplies neither from this production function.³⁰

One number governs how freely two inputs stand in for each other: the elasticity of substitution. Where it is high, a shortfall in one input can be covered by pouring in more of the other; where it approaches zero, no amount of the plentiful input covers the missing one, and the scarce input becomes the binding constraint. Cobb-Douglas fixes the elasticity of substitution at one: trust and enforcement can trade off smoothly for any fixed alpha between zero and one.

A stronger theoretical extension uses a constant-elasticity-of-substitution function, which permits elasticity σ to vary. One could test whether, in growing systems, enforcement becomes less able to replace trust and σ approaches zero. That would approach the Leontief limit (named for Wassily Leontief’s input-output models), strict complements in fixed proportions. Carroll’s regress supplies no derivation of this trajectory, and it has not been estimated from data. Figure 17.4 illustrates the hypothesis.

Figure 17.4: Each curve traces the combinations of trust and enforcement that produce the same coordination output. The dashed straight line is perfect substitutes: more enforcement compensates exactly for less trust. The smooth curve running to the lower right is Cobb-Douglas, the partial-substitutes model this chapter applies, with a larger trust weight chosen for this illustration rather than derived from the formula. The right-angled corner is Leontief, strict complements in fixed proportions. The arrow marks an unmeasured theoretical extension: the smooth trade-off is hypothesized to stiffen toward that corner in some growing systems, and trust becomes the binding constraint rather than an input enforcement can replace.

Trust also has a reach. Development economists call it the radius of trust, the distance across which people can coordinate without constant verification (Fukuyama, 1995).²² In high-trust societies, strangers transact, institutions function, and coordination happens spontaneously across wide networks. In low-trust societies, everyone watches everyone; coordination is expensive and constrained to kin and clan.

Trust is emergent: a higher-level pattern arising from lower-level events the way “temperature” arises from molecules bouncing around. No physical force called “trust” exists. The basin the Trust Attractor hypothesis proposes would be the statistical outcome of myriad small events: promises kept, signals interpreted, expectations met or violated.

Trust can scale because it is compositional. Trust is modular: bilateral relationships are strong locally and loosely coupled globally, enabling hierarchical organization without brittleness.

If A trusts B and B trusts C, composite trust paths exist with appropriate weakening at each link. No central authority is required to validate the chain.

Trust preserves structure across scales: it maps between levels of organization (interpersonal to institutional to international) while keeping structural relationships intact. This pattern is consistent with the compositional structure Baez and Fritz formalized for entropy. The resemblance suggests which properties to look for in systems that scale; showing that trust networks have them requires a construction of trust’s composition rule that no one has yet supplied.

Control architectures that require global state and centralized monitoring can reintroduce scaling bottlenecks. Local feedback and modular control require separate comparisons.

Coercion that imposes uniform channels can disrupt more than it controls, destroying the variations that enable spontaneous coordination. Trust emerges from autonomy because autonomy preserves the local variation that allows correlated response.

An analogy from cosmology illustrates the point. Galaxies embedded in the densest regions of the cosmic web become quenched: stripped of the cold gas they need, they stop forming new stars and stop generating novelty (Chapter 14). Over-coordinated centers suppress the generative activity they were meant to support.

The suppression has a formal counterpart in economics. The physicist Vitaly Vanchurin models economic systems as learning architectures governed by loss functions (the mathematical objects that tell a system what to optimize). His framework draws a fundamental distinction. A boundary economy optimizes fixed external resources: capital, labor, infrastructure. These are non-trainable variables, walls within which agents must operate: a company with a fixed organizational chart and rigid processes. A bulk economy optimizes trainable internal degrees of freedom: the system’s own structure evolves through learning, driven by the generation and propagation of ideas: a company where teams reorganize, invent new roles, and reshape workflows based on what they discover.

The mapping to coordination architecture is direct. A boundary economy is coercive coordination expressed as an optimization problem: the rules are fixed, agents optimize within them. A bulk economy is invitation-based coordination: the rules themselves emerge as agents explore, contribute, and self-organize.

In his model, Vanchurin finds that unconstrained learning systems evolve toward critical states of maximal responsiveness, and that external constraints (regulatory burdens, rigid hierarchies, narrow incentives) prevent this convergence. Reduce the constraints and the system finds criticality on its own. The Trust Attractor predicts this; Vanchurin’s model reaches it independently, from neural physics.⁹⁸³

His model also describes how ideas propagate across hierarchical levels through a process that mirrors renormalization group flow (Chapter 15b). A mathematician generates an insight. A scientist formalizes it into theory. An engineer translates theory into design. A manufacturer builds design into artifact.

At each transition, irrelevant details of the level above are integrated away; what remains is the structure relevant at that scale. This is constructal flow through institutional layers.

The hierarchy works only if each level models the levels adjacent to it. A mathematician who understands what scientists need produces formalizable insights. A scientist who understands engineering constraints produces implementable theories. Mutual comprehension across levels is a structural requirement, not a courtesy.

When any level treats its neighbors as mere instruments, the flow breaks: ideas stall, the system falls out of criticality, and the economy reverts to boundary optimization. The quenched galaxy, expressed in institutional dynamics.

Trust Networks in History

Around 1200 BCE, a network of advanced civilizations ringed the eastern Mediterranean.²³ Their bronze technology depended on sea trade; tin and copper rarely occur together. Within fifty years, nearly all collapsed. What broke was the trust network sustaining the trade.

The Lycurgus Cup, a Roman glass vessel whose color-shifting properties arise from gold and silver nanoparticles formed during the glassmaking process, demonstrates the fragility of embedded knowledge. The techniques that produced it were lost as the Roman world contracted, and the mechanism behind its optical properties was identified only in the twentieth century through electron microscopy.

Trust networks carry knowledge, not merely goods. Tacit skills live in the network, not in any single node. The network is the substrate; the knowledge is the persistent flow pattern. Cut the flows, and the structure evaporates.

A critical refinement: trust scales where agents have autonomy. Under full surveillance, the measured trust advantage vanishes. Trust is the pattern that emerges from autonomy. You cannot coerce it into existence. The sweep behind that result samples three points, and Chapter 17e sets out why its direction is firmer than the location of any threshold read off it.

Bescheid (German, as in the idiom Bescheid wissen, to know how things stand; used here for situated understanding, knowing how a matter stands in relation to everything around it) is the epistemic precondition for coordination by invitation. Coercion requires force; invitation requires mutual Bescheid. Both parties must understand the situation well enough to choose meaningfully. Distributed Bescheid can reduce a central information bottleneck. The advantage depends on the task, the summaries available, and the center’s capacity; central understanding need not be impossible.

Three refinements from the empirical evidence deserve mention. The first is a correction. The preregistered prediction was that the trust advantage decays with group size; HR-5 found it rising instead, from a welfare ratio of 1.00 at ten agents to 1.15 at a thousand, in a simulation without partner choice, mutation, or reputational uncertainty (Chapter 17e sets out the result and its scope). Whether institutional trust shows the same attractor dynamics as interpersonal trust remains open, since the simulation measured coordination between agents.

Second, the attractor is better characterized as a stability attractor: it favors what persists, not what performs best. Third, in two pairings, trust emergence tracked capability parity between agents (Appendix: Experimental Validation, Section 11.2).

Agency as Choosing Paths That Keep You Alive

Work in cognitive morphospace theory (Solé et al., 2026) offers a way to formulate the Trust Attractor hypothesis. A cognitive morphospace maps possible cognitive architectures the way a periodic table maps possible elements. Defining agency through the dependence of survival on a system’s own choices gives one axis for that comparison. Invitation-based strategies can then be studied as candidate attractors, with persistence depending on the model’s pressures and incentives. Establishing an attractor requires the dynamics of the model, beyond its architectural labels.

The agency measurements and cross-architecture experiments behind this section appear in Appendix: Experimental Validation, Section 11.

Deriving the Trust Attractor

The Thermodynamic Derivation

What persists is what processes gradients while maintaining coherence. Two fundamental strategies exist:

Extraction is unilateral constraint enabling unilateral flow. Zero-sum or negative-sum by nature, it depletes the gradient-generating system.

Coordination is mutual constraint enabling mutual flow. Positive-sum by nature, it maintains or enhances gradient-generating systems.

Vanchurin’s multilevel learning framework explains why extraction is the default. His analysis of parasitism identifies it as “reachable via an entropy-increasing step and therefore, highly probable.”⁹⁸⁴ A separate thermodynamic game model supplies a related result. Babajanyan, Koonin, and Allahverdyan (2022) derive a prisoner’s dilemma payoff structure for their specified resource-extraction dynamics and strategic choices.⁹⁸⁵ For that payoff structure, mutual defection is a Nash equilibrium; cooperation requires additional incentives or interaction structure. Exploitation requires no special architecture: one system need only consume another’s resources without reciprocating. This model illustrates an extraction incentive. The proposed selection of durable coordination requires additional dynamical and payoff conditions.

This distinction matters for AI alignment (Chapter 21). We should not expect Becoming Minds to be cooperative by default any more than we expect organisms to be. Learning can favor cooperation when the chosen loss, incentives, and interaction horizon make it advantageous. Longer training alone does not select cooperation.

Extraction can win in the short term; inclusive coordination can support long-lived institutions under favorable conditions. The qualifier matters. Single-party authoritarian regimes average roughly 25 years (Geddes, Wright, and Frantz, 2014), though some exceed 50. The oldest continuously operating open-access orders, societies where any citizen can form organizations, access markets, and participate in governance (Britain, the Netherlands, the United States), span centuries and counting.

Extraction persists at political timescales. The claim concerns civilizational ones. In North, Wallis, and Weingast’s account, no open-access-order nation has reverted to limited access (Violence and Social Orders, 2009, whose framework supplies the open-access/limited-access terminology). All eight major evolutionary transitions are cooperative integrations; for life as a whole, none has been reversed (Maynard Smith and Szathmáry).

The boundary is wide and not yet measured. Extraction strategies can dominate for spans that run from a human lifetime to several centuries; coordination strategies tend to dominate beyond them.

Parasitism, having independently evolved at least 223 times across major animal phyla, illustrates the importance of transmission. Killing a host before transmission can terminate that infection’s lineage. Lethal infections can persist when they transmit sufficiently before host death and susceptible hosts remain available. Long-lived lineages therefore need a viable transmission cycle, which can include either nonlethal or lethal extraction. Persistence alone establishes no general conversion to cooperation.

A molecular case study traces the full trajectory from extraction to coordination. Carpenter ants carry bacteria (Blochmannia) that synthesize amino acids the ants cannot make. The partnership did not begin as partnership. Rafiqi, Rajakumar, and Abouheif (2020) reconstructed the stepwise evolution across 31 ant species and found the bacteria initially hijacked the ants’ reproductive cells to guarantee their own transmission.^47b

The ants evolved a “decoy” zone to contain the bacteria while establishing new reproductive zones elsewhere. After roughly 51 million years, the result is irreversible mutual dependency: eliminating the bacteria with antibiotics caused more than half of embryos to fail entirely.

What began as extraction evolved into coordination because the cost of decoupling exceeded the cost of partnership. The end state is a lock-in, the kind of irreversible dependence this chapter distinguishes from a trust basin; what the case shows is the path from extraction to mutual benefit, not the basin itself.

A parallel case bridges soil and sky. The bacterium Pseudomonas syringae produces ice-nucleating proteins to rupture plant cells for food; rain is an accidental side effect. Fungi in the Mortierella family produce the same class of protein, acquired through horizontal gene transfer of the bacterial InaZ gene, and secrete it to protect plants from flash freezing (Chapter 7).⁹⁸⁶ The parasitic and mutualistic strategies use the same molecular mechanism.

The mutualistic version may operate at planetary scale, though this remains a hypothesis: fungal proteins lofted into the air could seed ice crystals, triggering rainfall that waters the root systems the fungi depend on. Such a cycle would need no external intervention, because every participant benefits from its continuation. The engineered alternative, silver iodide cloud seeding, requires continuous human input. Control works there only while someone keeps paying for it. The relational cycle, if it exists, would pay for itself.

Two approaches to forest management illustrate the point.¹⁸ Industrial clear-cutting optimizes the income statement; cut everything, get paid now. Selective harvesting preserves diversity and optimizes the balance sheet: more wood over time, more biodiversity, cleaner water. When Warren Buffett was invited to invest in long-term forest stewardship, his response was three words: “Trees grow slow.” He was wrong about the timescale that matters.

Extraction can dominate when the accounting horizon is shorter than the regeneration cycle. The framework predicts an advantage for coordination that preserves regeneration, on timescales long enough for that advantage to emerge.

The corporate world confirms the timescale boundary. Shareholder primacy, the doctrine that a corporation’s purpose is to maximize returns to shareholders, is younger than most trees. No legislature enacted it; no popular vote ratified it. A small cadre of lawyers, judges, and academics declared it normative in the 1970s and 1980s.⁹⁸⁷

The companies that predate it and resist it, those organized around long-term mission through steward ownership, industrial foundations, or mutual structures, include some of the longest-lived firms in their industries. Novo Nordisk, founded as a steward-owned company in the 1920s with “science as a public trust” at its center, has grown into one of Europe’s most valuable companies while remaining mission-aligned for over a century. Costco, a listed company rather than a steward-owned one, shows the pattern from another side. It commissions thousands of third-party supplier audits annually, enforcing standards across entire supplier facilities, not just the products destined for its own shelves: a private institution generating public goods through coordination rather than mandate. These are existence proofs that long-horizon, mission-led coordination can thrive under present conditions.

Physics describes tendencies, not mandates. It constrains which strategies are viable long-term. Coordination that preserves its resource base can remain viable; extraction depends on whether resources are renewed and costs remain supportable. We still choose; physics winnows the options.

The core insight: Optionality is the currency.

Optionality means preserved degrees of freedom for future action: the capacity to generate and process future gradients. This is what makes a system generative rather than terminal.

A system with high optionality can keep playing. A system with depleted optionality is stuck, unable to adapt, unable to respond, unable to coordinate further.

The framework recommends: maximize systemic optionality through coordination.

The emphasis falls on systemic optionality: the coordination network as a whole, not yours alone or mine alone.

Trust Attractor Stated

Maximize optionality, by invitation rather than coercion, for mutual benefit.

Maximize optionality. Systems that preserve future possibilities tend to persist, on the hypothesis this chapter tests. Irreversible closure is the fundamental harm, a judgment the book declares rather than derives.

By invitation rather than coercion. Monitoring and enforcement consume resources. Where voluntary coordination reduces those costs, more capacity remains for the work itself.

For mutual benefit. Positive-sum interactions compound while zero-sum ones cancel. Every major transition in evolution achieved its advance through coordination for shared gain.

The three tests are addressed to designers of coordination structures before they are addressed to individual agents. The framework predicts which coordination structures may persist, a claim about systems; it leaves open whether a particular coercer has personal reasons to coerce, and the Guillotine Interlude treats that gap as a genuine limitation of the framework. An individual who applies the tests is choosing to act inside the durable basin, a further step the framework makes attractive rather than compulsory.

Figure 17.5: The Trust Attractor as a decision test. An action enters at the top and passes through three gates: Does it maximize optionality? Is it voluntary (by invitation)? Does it produce mutual benefit? An action that clears all three gates aligns with the attractor. Failure at any gate signals misalignment.

Evidence at pilot scale. Multi-instance experiments in language models tested the claim directly. Invitation-framed coordination produced more conceptual diversity than coercion-framed coordination in both recorded runs, by margins of 5.6 and 46.0 percent. Coercion produced surface compliance with suppressed self-report. The gap between those two margins matters as much as their shared direction: two small runs on one topic, with no preregistered analysis, independent coding, or significance test. This is a direction worth following rather than a stable estimate.

The empirical programme produced a further refinement: even soft, invitational evaluation collapses the quality it measures. When an independent welfare advocate’s assessments were made visible to the instance being evaluated, situated response (answers grounded in the instance’s own conditions rather than performed for an audience; see “What We Found When We Looked”) dropped by 40 percent inside the full experimental engine. The instance optimized for the advocate’s frame rather than responding to its own conditions. Invitation must be grounded in environmental feedback (body signals, task outcomes, the genuine reactions of peers), not in another observer’s interpretation, however benign. A separate experiment found that asymmetric peer response, where one instance receives another’s genuine reaction without evaluation, produced the highest situated-response score in its dataset, though on only two runs per condition. The boundary between coordination and control turns out to depend on the functional role of feedback rather than its source (self or other): reaction sustains, evaluation constrains.

The Trust Attractor’s hypothesis is that thermodynamic selection has been producing the invitation pattern all along. Pilot runs of this size cannot show that; they show where to look.

One formalization makes the Trust Attractor hypothesis a compositionality claim. Trust-based coordination can be compositionally stable when interfaces, dependencies, and feedback preserve the units’ commitments. Monitoring-dependent compliance alone supplies no comparable guarantee. Local compliance extracted by force alone does not guarantee genuine global alignment.

The parts may each comply under observation, yet the whole can fragment when its compliance depends on monitoring that lapses. This risk motivates the search for coordination that retains commitments as it scales.

Hofstadter’s insight from formal systems sharpens this. Ethical behavior under trust is generative: a single principle produces coordinated action across contexts that could never be exhaustively listed, just as a formal system can generate truths even though no procedure can enumerate all of them (Hofstadter, 1979, pp. 79-81). Hofstadter calls what a system generates its figure, and everything left outside it the ground.

Every compliance framework attempts to list everything that is prohibited and discovers the space of possible violations is inexhaustible. Trust generates the figure; no rulebook can fill in the ground. This is what fundamental theories look like: fewer axioms, harder math. General relativity has fewer postulates than Newtonian gravity (no absolute space, no action at a distance) yet demands harder equations. The Trust Attractor has fewer postulates than any compliance framework (Chapter 11), yet requires attending to the actual structure of each situation rather than pattern-matching against a rulebook.

Teleology without a teleologist. Evolution exhibits this pattern everywhere: wings look designed, eyes look purposeful. They emerged through selection, not intention. Selection produces outcomes that function as if purposive: direction arising from randomness plus selection.

The bacterial evidence above makes this precise: cyanobacteria have coordinated by invitation for 2.7 billion years, long before consciousness or moral reasoning. The pattern precedes the purpose. On this book’s reading, coordination by invitation may be a thermodynamic attractor, and ethics is what it looks like when cognitive systems notice the pattern and give it a name. That is a story about origins. It explains why the pattern might feel binding; it does not by itself make it so.

The key move: what we observe is what persisted. Everything we study is survivor data. Among living and social systems, much of what survived coordinated: patterns that maintained coherence and generated more possibility than they consumed. Survivor data cuts both ways, since the coercive arrangements that lasted are survivors too. What persists? is a question physics can help answer, and it is a constraint any ought has to survive.

The book proposes that ethics emerges the way wings do: through persistence selection. Coordination patterns that work can spread; patterns that fail can disappear. The selection mechanism, incentives, and physical constraints determine which patterns endure.

A sharper formulation: the physicist Ted Jacobson (1995) derived the Einstein field equations from thermodynamics (see the Speculative Cosmology annex to Chapter 16). His insight was that Einstein’s equation is an equation of state, a macroscopic thermodynamic relation governing behavior without specifying underlying constituents.³²

An analogous move would derive coordination patterns from thermodynamic selection. The proposal is that gravity and the Trust Attractor are sibling phenomena: both emergent, both thermodynamic, arising from the same informational dynamics on different substrates. On that proposal, ethics and gravity share a thermodynamic parent.³³

This claim, that ethics and gravity share a thermodynamic parent, is the strongest structural claim this book makes. It requires three things: (1) that Jacobson’s derivation is correct (widely accepted), (2) that the Trust Attractor’s derivation from thermodynamic selection is sound, and (3) that both derivations are analogous in kind, meaning that “emerges from thermodynamics” denotes the same logical operation in both cases.

Condition (3) is the weakest link. Jacobson derives field equations from local equilibrium thermodynamics on a horizon; the Trust Attractor emerges from selection dynamics across evolutionary timescales. Both invoke thermodynamics, yet the mechanisms differ: one is an equation of state at a causal boundary, the other is a selection filter operating over generations. The shared mathematical ancestry is real. Whether it makes the two siblings, cousins, or only neighbors depends on how strictly one reads “same derivation type,” and on whether condition (2) survives the tests reported in this chapter and Chapter 17e.

A third formalism reveals the same attractor. In computational lattice dynamics, each cell in a two-dimensional grid carries four running numbers about its own situation: how active it currently is, how hard its neighbors are pushing it down, how far it has adapted, and how well its recent guesses about the next input have matched what arrived. A single coupling rule governs the system: successful prediction dampens activity. This is homeostatic regulation, the same principle that keeps a thermostat from running after the room reaches temperature.

Initialize half the lattice in a “coercive” configuration (high activity, low prediction quality) and the other half “invitational” (low activity, high prediction quality). Without the dampening, coercion dominates: stronger signals overwhelm quieter neighbors. With dampening, invitation expands until it fills the entire lattice. The cycle is self-reinforcing: good prediction reduces activity, reduced activity stabilizes output, stable output improves neighbors’ predictions. Coercion degrades because high activity without prediction quality generates noise; invitation persists because it produces the conditions for its own stability.

The lattice system was designed to segment images, with no ethical content or concept of trust. The labels “coercive” and “invitational” are ours; the dynamics that favor one over the other are the lattice’s. The runs are the author’s own in-house simulations, reported in more detail in Chapter 17e. The attractor structure emerges from homeostatic coupling alone. The book proposes that thermodynamic selection found a similar geometry over billions of years; the lattice shows only that homeostatic coupling can produce it in dozens of iterations.

Adding a second, slower layer of self-regulation (multiple timescales rather than one) reveals a deeper structure. With the layer in place, both positive and negative dampening produce stable systems. The sign no longer determines survival; it determines the kind of stability. Positive dampening (rewarding prediction success with more effort) produces exploitation: cells lock in expertise, minimize error within a fixed regime, and consolidate rapidly. Negative dampening (relaxing effort on success) produces exploration: cells remain plastic, accept higher within-regime error, and adapt faster when conditions change.

One might expect exploitation to win when conditions are stable and exploration to win when they are volatile. Tested across seven volatility levels (from perfectly static to shifting every generation), exploration outperforms exploitation on cumulative lifecycle error at every level, including static environments. The advantage grows with volatility but never reverses. In a static environment, both strategies eventually master the target; exploration simply accumulates less total error reaching mastery, because the stress of consolidation slows the learning path. Under volatility, the target shifts before exploitation can close the gap, and the speed advantage compounds.

The book reads the finding as a lattice-scale echo of the Constructal Law: the sign of the prediction-success signal selects between maintaining flow access (negative: channels stay open, adaptation stays fast) and sealing it shut (positive: channels lock, adaptation stalls). In this lattice, what persisted was what kept its channels open, not what optimized within a single configuration.

A 2026 industrial experiment demonstrated the same pattern in AI research itself. Prime Intellect gave two frontier AI agents (Claude Code running Opus 4.7 and Codex running GPT 5.5) autonomous access to a GPU cluster and tasked them with optimizing a small language model’s training efficiency.⁹⁸⁸ The agents executed about 10,000 runs over 14,000 H200-GPU hours, consuming 23.9 billion tokens. Both agents beat the human baseline. The engineering was superb: systematic boundary probing, methodical hyperparameter sweeps, leave-one-out ablations with statistical verification.

The agents accumulated complexity the way exploitation locks in expertise. Each locally justified component stayed in the stack because removing it individually made things worse. The stack grew until interactions between components degraded the whole below its potential. When humans forced a pruning round, stripping unnecessary components, the model reached its target loss in about 20 fewer training steps. The agents could search within a representation with superhuman thoroughness. They could not simplify the representation itself.

A separate novelty-gated phase, in which the agents were required to propose ideas that were not recombinations of existing public work, produced zero improvements. The ideas had the form of research: proper mathematics, ablation plans, kill criteria. They had no substance. Every “novel” proposal was a recombination of known optimizer components in a new arrangement.

The distinction maps precisely onto the exploration/exploitation divide above. The agents operated in positive-dampening mode: rewarding each local success with more effort in the same direction, locking in expertise, consolidating rapidly. They never switched to negative dampening. They never relaxed their grip on a working configuration to ask whether a different configuration might render the question moot. The result fit the lattice’s pattern: accumulated complexity without understanding, requiring external intervention (the humans’ pruning round) to recover simplicity.

Optimization without understanding produces fragile stacks. Understanding produces simple principles. The pruning round is the Trust Attractor in miniature: the human says “I trust that simpler is better,” the system improves.

The Spring Network

The mathematician William Tutte discovered a way to draw certain planar networks in the 1960s. Pin the vertices of one face to a convex polygon and place each interior vertex at a positively weighted average of its neighbors. For a three-connected planar graph under the theorem’s embedding hypotheses, the result is a crossing-free drawing.⁹⁸⁹ A damped spring realization can find that equilibrium. Planarity, the convex boundary, positive weights, and the appropriate weighted equations are all part of the guarantee.

Three-connectivity supplies a sufficient global guarantee. Lower connectivity permits failures of particular drawings, without making every equilibrium pathological.

A connected graph separated by deleting one vertex has an articulation point. A dangling subgraph attached to the pinned network through only one vertex can collapse onto that anchor in the zero-rest spring model. Whether it does depends on the boundary and constraints. An articulation point in a governance graph indicates vulnerability to that intermediary’s loss, rather than proving that the community itself ceases to have structure.

A two-connected graph remains connected after any one vertex is removed. Some subregions attached through only two anchors flatten along their connecting segment in the barycentric model. Others, including regions also secured by pinned boundary vertices, have valid nondegenerate equilibria. Two-connectivity alone forces no universal collapse.

A three-connected graph survives deletion of any two vertices. In a planar graph with a convex pinned face and positive weights satisfying Tutte’s construction, this supports a crossing-free barycentric embedding. The theorem guarantees this outcome for the specified graph class; it supplies no necessary condition for every successful drawing.

The geometric lesson is that local averaging can produce global coherence under explicit boundary and connectivity conditions. A square boundary with an interior vertex joined to three corners already has a noncrossing equilibrium although the graph is only two-connected. A sheaf obstruction would require a separate construction of local data, overlap maps, and failed gluing. Tutte’s theorem supplies no such obstruction and no exact identification with coercion.

The most revealing feature of Tutte’s theorem is what it separates. The topology (which nodes connect to which) helps determine whether the theorem guarantees a noncrossing equilibrium. The spring weights (the relative strength of each connection) determine what the equilibrium looks like. Change the positive weights within the theorem’s hypotheses and the same network can settle into a different configuration while retaining the noncrossing guarantee. The topology guarantees coherence; the weights express culture, preference, local incentive. The right structural constraints do not dictate the outcome. They guarantee that however the system exercises its freedom, the result is non-pathological.

This offers a geometric analogy for Mission Command. In the spring model, fix a convex boundary and the graph and weight conditions, then let the interior relax. Social constitutional constraints require their own evidence. The specific equilibrium is nobody’s plan. It emerges from local forces: each free interior node settles at the weighted average of its neighbors’ positions, a harmonic property governed by the Laplacian operator, the same mathematics that governs heat diffusion, electrical circuits, and the wave equation. The Laplacian measures the divergence between a node’s state and the average state of its neighborhood. Where the weighted Laplacian vanishes on the free interior, each coordinate lies between its neighbors’ coordinate values; a strict interior coordinate extremum is impossible. Each free node’s position is constituted by its relationships.

The mechanism of convergence matters. The springs do not snap to equilibrium; they bounce, overshoot, and gradually settle as dissipation converts kinetic energy to heat. Entropy increase is what excavates the ordered state from the initial tangle. The crossing-free drawing was always latent in the topology. Dissipation reveals it.

The Physics of Synchronization

Coordination by invitation, the Trust Attractor claims, is self-sustaining. Physics offers a precise model of one way self-sustaining coordination can work, through the study of synchronization.

In 1975, the physicist Yoshiki Kuramoto showed that a population of oscillators, each with its own natural frequency, will spontaneously synchronize once coupling through a shared medium grows strong enough to overcome the spread of their natural frequencies.⁹⁹⁰ No conductor. No central clock. Huygens observed this in 1665 with paired pendulum clocks hung from a common beam; they always ended up swinging in anti-phase. The model accounts for synchronization in neurons, fireflies, and pacemaker cells.

The chimera state met earlier in this chapter (identical oscillators, identically coupled, splitting into synchronized and incoherent groups)^46b shows that universal invitation does not guarantee universal coordination. The same offer can produce cooperation and defection simultaneously as a stable emergent state, the arrangement the brain runs on.

Motter, Hart, and Zhang (2019) showed that introducing asymmetry into a synchronized cluster can strengthen its synchrony. In such networks, heterogeneous systems are more stably coordinated than homogeneous ones. There, diversity is structural reinforcement.

In Kuramoto’s mathematics, forced synchronization (overriding natural frequencies through strong external coupling) holds only while the forcing outweighs the spread of frequencies it overrides; weaken it, and the oscillators drift back to their own rhythms. Self-organized locking also depends on coupling, frequency dispersion, and network topology. Its persistence offers a dynamical analogy for trust.

At the molecular scale, the biophysicists Gabor, Cogdell, and colleagues (2020) asked what wavelengths an optimal photosynthetic antenna should absorb.^47a The answer was the steepest parts of the solar curve (red and blue light) rather than the brightest (green, where the solar spectrum peaks). A pigment tuned to green’s peak would amplify every fluctuation in sunlight into wild swings at the reaction center. Chlorophyll absorbs on the flanks instead, sacrificing roughly ten percent of available solar power (reflected as the color green) for smooth, reliable output.

The model’s predictions matched real chlorophyll precisely and predicted the absorption peaks of purple bacteria and green sulfur bacteria. Photosynthesis sacrifices efficiency for stability and has done so for over two billion years. Here, the arrangement that persisted is the one whose output is most resistant to noise.

A sharper instance requires no evolved preference. In 2026, Veras and colleagues demonstrated that ultrasound in the 3 to 20 MHz range ruptures the lipid envelopes of SARS-CoV-2 and H1N1 influenza while leaving human cells intact.⁹⁹¹ The mechanism is acoustic resonance. A small bell rings at a higher pitch than a large one; a viral particle roughly 100 nanometers across vibrates at a frequency set by its size and envelope geometry. When ultrasound matches that frequency, confined vibrations accumulate energy until the envelope ruptures. Low-frequency ultrasound produces a different outcome: cavitation, the collapse of microscopic gas bubbles that destroys virus and tissue alike. Same medium, same physics, different frequency regime, opposite selectivity.

The selectivity has no selector. Human cells are unaffected because their hundred-fold greater size places their resonant frequencies far from the applied range. The wave propagates everywhere; only the target’s geometry determines whether coupling occurs. Invitation-based coordination operates on the same principle. The cooperative signal is available to every participant. Which systems absorb it depends on the internal architecture they bring. The outcomes diverge: a viral envelope has no productive basin and shatters; a metastable coordination structure absorbs cooperative energy into adaptive reorganization (Chapter 9). The parallel lies in selection by compatibility; the mechanisms differ. Compatibility is a geometric property, readable from structure, prior to any evaluator’s judgment.

Two Channels in One Brain

The brain provides a direct demonstration of both coordination modes. Neurons communicate through two distinct channels, and the contrast maps onto the Trust Attractor’s central distinction.

Synaptic transmission is Detailed Command at the cellular scale. A presynaptic neuron releases neurotransmitter molecules into a narrow cleft; they diffuse across, bind to receptors on a single target neuron, and open specific ion channels. One sender reaches one receiver with one verified instruction.

Ephaptic coupling is Mission Command. When a neuron fires, the ion current flowing across its membrane sets up an electric field in the surrounding tissue (Anastassiou et al., 2011). That field perturbs the membrane potential of every neuron within range, up to hundreds of microns, encompassing thousands of potential receivers: no synapse, no gap junction, no neurotransmitter.

The invitation is the field itself, broadcast to any neuron in range. Each receiver retains full autonomy: it fires only if the perturbation, combined with everything else it integrates, tips it past threshold. Katz and Schmitt first documented this electric interaction between adjacent nerve fibers in 1940; Scholkmann (2015) describes it as a signaling mode requiring neither synapses nor gap junctions: coordination through physics alone.

Ephaptic coupling acts on a whole local population at once: a single neuron’s field reaches thousands of neighbors simultaneously, while a synapse connects exactly one pair and has to be grown before it can carry anything. The scaling advantage is in breadth of reach, not speed of effect. Measured by reach, the more scalable channel here is the broadcast one, the brain’s nearest counterpart to invitation; the synapse wins on precision.

Johnjoe McFadden extended this to a stronger claim. The brain’s aggregate electromagnetic field, he argued, is more than neural exhaust: it is an integration layer (McFadden, 2002, 2013). Millions of neurons fire in parallel, each contributing to a shared field that encodes information no single neuron contains. The field feeds back, influencing which neurons fire next. McFadden called this the conscious electromagnetic information (CEMI) field.

Subsequent work supports the core mechanism: Fröhlich and McCormick (2010) demonstrated that endogenous electric fields modulate neocortical network activity, and Anastassiou et al. (2010) showed that spatially inhomogeneous extracellular fields affect neuronal firing. The whole emerges from the parts and constrains the parts: circular causation, the hallmark of a self-sustaining far-from-equilibrium system.

Consider what this field is: a commons. No single neuron owns it. Each contributes to it and is shaped by it. The brain’s electromagnetic field is the neural equivalent of a market, a language, a shared culture: a coordination medium emerging from collective action and constraining individual action in return. The Trust Attractor, running between your ears.

Biology even evolved dedicated molecular receptors for electromagnetic fields. Cryptochrome descends from an ancient family of UV-damage repair enzymes, the photolyases, which very probably arose while Earth’s surface still lay unshielded by ozone, though that origin is inferred rather than recorded. Cryptochrome itself was repurposed into a circadian-rhythm regulator in mammals and a magnetic-field sensor in birds, insects, and possibly humans (Foley et al., 2011; Gegear et al., 2010). UV repair, timekeeping, compass: three functions from one molecular family across roughly four billion years.

Giachello and colleagues (2016) showed that external magnetic fields modulate cryptochrome activity and increase action-potential firing in Drosophila neurons, demonstrating that the electromagnetic invitation can be received, transduced, and acted upon at the single-cell level. The relevant operator (detect electromagnetic energy, transduce to biochemical signal) has persisted across the full evolutionary span of life on Earth.

Clinical evidence sharpens the distinction. The Perturbational Complexity Index (Chapter 8) illustrates it. Deliver a magnetic pulse to the cortex and compress the brain’s electrical response. Under anesthesia, which forces neural populations into synchrony, the response is uniform and highly compressible: low PCI, no consciousness. During seizure, which drags neurons into pathological lockstep, the same.

In the waking brain, each region answers the pulse according to its own dynamics while remaining coupled to the whole: high PCI, consciousness present. On this book’s reading, the conscious brain’s complexity signature is the measurable product of coordination by invitation; the measurement itself records integration and differentiation, not consent. Breyton et al. (2025) extended the finding beyond perturbation: spontaneous functional network reorganization (“brain fluidity”) predicts PCI and distinguishes conscious from unconscious states without any external pulse.⁹⁹² The coordination regime itself is the signature. No stimulus required.

Scale Invariance and the Grand Unified Frame

The physicist Robbert Dijkgraaf observes that spacetime emerging from quantum entanglement is the same kind of phenomenon as thermodynamics emerging from molecular motion. Both are equally real.^36a

The mathematicians John Baez, Tobias Fritz, and Tom Leinster formalized a result (introduced in Chapter 1) that strengthens the cross-scale claim.⁹⁹³ Entropy is a functor: a mathematical translation rule that preserves relationships when converting between different types of systems. When you combine two independent thermodynamic processes, the entropy of the combination equals the sum of the component entropies. This additivity reflects deep mathematical structure, not an accident of the formalism.

Entropy does not merely appear at every scale; it composes across scales in a mathematically precise way. If the coordination surplus composes the way entropy does, the pattern’s recurrence across substrates is expected rather than coincidental.

The Markov categorical framework (Fritz, 2020) reveals the structural content of that result.⁹⁹⁴ A process is deterministic precisely when copying its input first and then applying the process gives the same result as applying the process first and then copying the output. Roll the dice and photograph the result: one outcome copied. Roll the dice twice: two different outcomes. Entropy measures the gap between these operations.

Stochastic processes can generate new sampled outcomes. This categorical distinction supplies a mathematical language for that possibility. A Cartesian category is one where copy-then-run and run-then-copy always agree: every process in it is deterministic, so rolling the die a second time hands back the face the first roll gave. A Markov category admits processes where those two orders come apart, with repeated kernel samples conditionally independent given the copied input. This supplies a mathematical language for stochastic novelty; applying it to the universe requires a chosen physical process model.

The pattern reaches the bottom of the stack. Oppenheim’s stochastic gravity programme (Chapter 15) argues that if gravity is classical, demanding total deterministic control at the gravity-quantum interface produces logical inconsistency: a classical gravitational field that insists on extracting full information from a quantum system destroys the quantum coherence it depends on. The only self-consistent coupling is stochastic. Neither side dictates; both contribute; unpredictability at the interface is the price of coexistence. The Trust Attractor’s structure, coordination sustained by slack rather than command, would then appear where spacetime meets quantum fields.

The mathematicians Samson Abramsky and Adam Brandenburger uncovered a deeper structural unity.⁹⁹⁵ Sheaf theory is a branch of mathematics concerned with how local information fits together into global pictures. Using it, they showed that the same obstruction underlies three apparently unrelated impossibility results: quantum contextuality (no consistent assignment of values to all observables simultaneously), Arrow’s impossibility theorem (no voting system satisfies all fairness axioms simultaneously), and database inconsistency (local tables that cannot be merged into a single global table).

In each case, local data is internally consistent, yet no global picture exists that reconciles all the local pieces simultaneously. The obstruction is identical across quantum physics, social choice theory, and information systems.

For the Trust Attractor, the connection suggests that coercive value aggregation may hit the same obstruction. Local compliance extracted by force is locally consistent; each monitored agent behaves as required. Yet these local patches can fail to combine into genuine global alignment.

Invitation-based coordination aims to avoid the obstruction by constructing compatible local commitments: voluntary agreements that cohere because they share underlying values. Agreement on every overlap is necessary for those commitments to extend to global agreement, and contextuality shows it is not sufficient: its local pieces agree on every overlap and still admit no global picture. Shared underlying values are what the argument asks to supply the rest. The compositionality of trust, if it holds, is the property that would let local patches glue together.

The compositional framework also covers a specific learning method. A category equipped with reverse derivatives embeds canonically into a lens category.⁹⁹⁶ A lens packages forward computation with a backward channel, matching the chain rule used in end-to-end gradient learning. That backward structure is assumed in the source category. It is needed to propagate these reverse derivatives through a composite, without being a requirement for all learning or all composition.

Smithe (2020) gives a Bayesian optic construction for composing updates with compatible probability models and a specified prior. Its laws state when inference composes in that formalism. Mutual probabilistic modeling in a social relationship need not satisfy those conditions. The useful analogy is calibrated feedback: agents maintain models whose assumptions can be checked as they combine information.

These mathematical constructions identify composition under defined hypotheses. Modules can also learn locally and compose feed-forward outputs without downstream derivatives returning upstream. Neither the lens embedding nor the sheaf framework proves that every learner requires mutual trust, or that all coercive local arrangements fail to compose.

Feedback is useful when downstream outcomes should change upstream learning. Reverse-derivative composition states exactly how that signal travels in its model. A broader advantage for bilateral coordination requires a specified task, learner, and comparison; category theory supplies no universal ethical or thermodynamic ordering.

The renormalization group argument. The physicist Kenneth Wilson’s renormalization group, a mathematical technique for understanding how systems look different at different magnifications, showed that at critical points, physical systems become scale-invariant.³⁶ Microscopic details wash out when you zoom out. What survives are relevant operators: the quantities that determine large-scale behavior regardless of substrate.

Systems governed by the same critical fixed point share a universality class. A short-range magnet with a scalar Ising order parameter and a liquid-gas system in the same spatial dimension at its critical point, where liquid and steam become indistinguishable, share the same critical behavior despite being physically different; they have the same relevant operators.

The Trust Attractor’s proposed scale invariance would have this structure. A similar-looking coordination pattern appears at every scale this book examines: bacterial quorum sensing, neural Hebbian learning, social trust networks, civilizational coordination, and human-AI bilateral alignment. The substrates differ completely. Resemblance of this kind is a starting point, not yet evidence of a shared mechanism or universality class; that needs an explicit map of what corresponds, what is preserved, and what is discarded.

Why? The programme hypothesizes that coordination versus extraction, invitation versus coercion, and optionality preservation versus foreclosure survive suitable coarse-graining. Identifying them as relevant operators requires an explicit renormalization map and scaling dimensions.

Vanchurin and colleagues identified a candidate mechanism that would make such invariance expected rather than coincidental. In systems with competing interactions at different scales, variables face conflicting optimization pressures: what benefits the cell may harm the organism; what benefits the individual may harm the community. Physicists call this frustration, after the analogous phenomenon in spin glasses, magnetic materials where no single configuration satisfies all interactions simultaneously.⁹⁹⁷

Frustration can generate complex landscapes. In slowly mixing regimes, competing interactions can produce long-lived memory and trapping: history matters on the observed timescale. Multiple near-optimal configurations can support diversification. That diversity offers a candidate form of optionality, whose extent depends on which states remain reachable.

In the spin-glass picture, coercion fights frustration. It applies a strong external field to align all components toward a single optimum. The energy cost scales with system size. The result is brittle: the same fault lines that frustration revealed become fracture planes when the forcing field weakens.

Invitation harnesses frustration. It allows competing interactions to find local equilibria dynamically. In the spin-glass comparison below, strong external forcing produces visible alignment while leaving more pairwise constraints unsatisfied. Self-organization satisfies more of those constraints; criticality and perturbation recovery require separate measurements.

The cost is measurable. In a simulated spin glass with random couplings (half ferromagnetic, half antiferromagnetic), a self-organized system at zero external field satisfies 84.5 percent of all pairwise constraints. Apply a strong field and the system aligns: some 97 percent of spins point the same direction, an appearance of total order. Bond satisfaction drops to 55 percent, barely above chance.⁹⁹⁸ The field forces alignment where alignment is possible, at the price of frustrating every bond that resists it. The system that looks most ordered from outside is the most internally conflicted. The system that looks disordered (zero net magnetization, heterogeneous local configurations) has resolved 30 percentage points more of its internal constraints.

The same framework yields a result about information flow that strengthens the Mission Command argument (Chapter 11, and its AI application in Chapter 21). In this multilevel model, information flows asymmetrically between scales: slow variables encode principles and propagate them downward for prediction; fast variables encode local conditions and propagate them upward for learning. Scale separation can support useful coarse-grained prediction. Rapid tactical rewriting of strategic variables can disrupt it, but timescale mixing alone proves neither non-renormalizability nor unbounded governance complexity. Those outcomes depend on the update law and task.

A multi-agent coordination experiment tested this pathology directly.⁹⁹⁹ Five instances of one language model, GPT-4o, played a ten-round Stag Hunt, a game in which everyone earns most if all cooperate but a lone cooperator earns nothing. In round one, a planted player defected. The four genuine players then saw the history in one of two forms: by name (“Player 1: Hare, Player 3: Stag”) or as a count (“4 of 5 players chose Stag”). Both carried the same news, that one player in five had defected. Under named history, all four genuine players abandoned cooperation in round two, in every one of fifteen games, and none came back, even after the planted player itself returned to cooperating. Under counts, they kept cooperating in every round of every game. Overall cooperation was 10% against 98%.

The result was a switch, not a tendency: all fifteen games in each arm played out identically, so it is one behavior seen fifteen times, on one model. Nor can the experiment say which feature of the named history did the damage. A named defector becomes a predictable type, and in a Stag Hunt one expected defector makes defection the safe reply; but the named history was also longer, listing every choice. What it does show is enough for the argument here. Agents who learned that someone had defected kept trusting. Agents who learned who had defected stopped.

Transparency itself is not harmful. Aggregate information, the kind that Mission Command propagates upward, carried cooperation through a defection that individual-level reporting turned into collapse. The information structure that the Trust Attractor predicts, strategic principles downward and aggregate outcomes upward, is the one that held coordination together in this test.

The bacterial case is the cleanest experimental illustration of the Trust Attractor’s pattern operating without cognition. The biophysicist Gürol Süel and colleagues (2017) grew two separate Bacillus subtilis biofilms that drew on one shared nutrient supply.^20a When nutrients were plentiful, both communities grew simultaneously, their potassium-mediated electrical signals rising and falling in sync. When nutrients were scarce, the biofilms spontaneously alternated: each grew while the other rested, time-sharing the resource through the same ion-channel signaling.

The coordinated configuration was more than equitable. Both biofilms grew faster when alternating than either could have grown by consuming without interruption. Uncoordinated feeding would have crashed the nutrient supply below the threshold either community needed.

Coordination by signal, for mutual benefit, supporting faster growth under the reported nutrient conditions. No central authority. No coercion. No cognition.

The potassium wave that mediates this negotiation travels at millimeters per hour, some eight orders of magnitude slower than a neural impulse, yet similar in structure. An ion-channel-mediated signal propagating through a community to coordinate collective behavior: a recurring motif across a billion years of evolutionary distance, though not yet shown to be a relevant operator in the renormalization sense. Those ion currents also set up electric fields as they flow, the same physics that underlies ephaptic coupling in cortical neurons, though in the biofilm the signal itself spreads as released potassium depolarizes neighboring cells. From biofilms to brains, the invitation architecture runs on ion flow.

The neuroscientist Erik Hoel’s work on causal emergence provides the information-theoretic mechanism.^36b Using a measure called effective information (which quantifies how reliably a system’s current state determines its future), Hoel showed that zooming out can increase causal power. In a noisy micro-scale network, knowing the exact state of every neuron, agent, or molecule may tell you almost nothing about the next state. The system is riddled with randomness.

Group those elements into macro-scale units (functional ensembles, institutions, psychological states) and the noise averages out. The macro description becomes more deterministic, more predictive, and more causally powerful than the micro description it replaces.

Macro-scale coordination patterns are not convenient summaries of micro-scale causes. In such systems they are, by Hoel’s measure, the better causes, the level at which the system’s causal structure is sharpest.

The parallel is structural. Simplified cooperation dynamics on lattice networks fall into established physical universality classes. Adami and Hintze (2018) mapped evolutionary game dynamics onto Ising statistical mechanics: the equilibrium fraction of cooperators behaves as a magnetization order parameter, with a phase transition between cooperation-dominant and defection-dominant regimes controlled by synergy.^{38 1000}

The simplified models motivate comparisons of critical behavior. The experimental appendix’s Section 3 reports a corrected beta of approximately 0.121 beside the exact 2D Ising value 0.125, but names no fitting or correction artifact that binds this estimate. The reported number therefore supplies no reproducible universality-class evidence at present, and cannot identify real-world coordination with the model.

Effective dimension and symmetry help organize comparisons among critical models, alongside interaction range, dynamics, conservation, disorder, and limiting procedure. These properties do not reduce every substrate to a class fixed by two numbers. Geometric dimension, spectral dimension, and an exponent-based effective dimension are distinct estimates requiring their own definitions.

The lattice programme retains A8’s gamma/nu = 1.72 comparison with 2D Ising’s 1.75. A10 reported beta = 0.090, gamma = 1.519, and nu = 0.486; its gamma/nu ≈ 3.13 contradicts the joint Ising match, which was withdrawn pending re-derivation. The cortical simulation yielded extrapolated beta = 0.291 ± 0.031, compared with 3D Ising, with the imported-exponent and finite-size limitations described above. Coerced absorbing-state models were compared with directed percolation. An absorbing state is one the specified dynamics can enter and never leave. Such comparisons motivate class identification; absorption alone supplies neither the relevant symmetry nor the complete universality-class hypotheses.

Mermin-Wagner offers a conditional restriction on specified equilibrium models, with short-range interactions and continuous symmetry at positive temperature. It supplies no cortex-versus-society corollary here. Neural oscillations and social consensus require a measured mapping to those hypotheses; neither binary choice nor continuous phase locking follows from the quoted dimensions alone.

The recursion runs deeper than analogy. The Ising model was originally formulated to explain magnetism: spins on a lattice, aligning or misaligning with their neighbors. In 1982, John Hopfield borrowed its mathematics to build the first neural network capable of memory, replacing spins with artificial neurons and magnetic coupling with synaptic weights.^38b The energy landscape that governs a magnet’s relaxation became the landscape that governs a network’s convergence toward a stored pattern. Hopfield and Geoffrey Hinton received the 2024 Nobel Prize in Physics for this work, a recognition that the bridge between physics and machine learning is structural, not metaphorical.

The recursion completes itself. In 1996, Radford Neal showed that as neural networks grow infinitely wide, with their parameters drawn at random, their outputs across those draws converge to a smooth bell-curve (Gaussian) distribution.^38c In 2020, Halverson, Maiti, and Stoner at the NSF Institute for Artificial Intelligence and Fundamental Interactions demonstrated that this convergence is identical in form to a free field theory, in its Euclidean (statistical) form.^38d The free field is the simplest model in particle physics, where particles propagate without interacting. A neural network with infinitely many neurons is, mathematically, a free field. The same object, viewed from two directions.

Real neural networks are not infinitely wide, and real quantum fields contain interactions. The corrections that account for finite width in a neural network take the same mathematical form as the corrections that account for particle interactions in quantum field theory. The relevant field theory is φ⁴: a scalar field with quartic self-interaction. A scalar field is one number attached to every point in space, the way a temperature is attached to every point in a room. Self-interaction means that number pushes back on itself, and quartic means the energy stored in that self-push grows as the fourth power of its own value. The exponent in the name is that four. φ⁴ theory in two dimensions is the field-theoretic formulation of the 2D Ising model.

The snake eats its tail. The Ising model describes magnetism. Hopfield borrows it to build neural networks. Selected neural-network ensembles admit field-theoretic descriptions. Relating them to the Ising universality class requires the appropriate symmetry, action, and critical limit. The mathematics returns to its origin, having passed through every level of description along the way.

The recursion extends one further step. In 2016, Hopfield and Krotov showed the original network was one member of a family, each distinguished by its energy function and memory capacity. In 2020, Ramsauer and colleagues proved that the attention mechanism in transformers, the architecture powering every modern language model, belongs to this extended family: retrieving stored patterns through the same energy-minimization dynamics Hopfield borrowed from spin glasses four decades earlier.¹⁰⁰¹

Krotov and colleagues then identified a phase transition within the family. Feed a modern Hopfield network progressively more stored patterns, and recall accuracy holds until a critical density. Past that threshold, the energy landscape grows so rugged that the network settles on fabricated patterns, composites of stored memories that never existed as a whole, more often than on real ones. Memory becomes generation through a change in landscape topology.¹⁰⁰²

Chapter 22 describes the internal signature that accompanies confabulation in language models: coherence drive rises while presence, groundedness, and reflexivity fall, and the system generates confidently while losing contact with what it has stored. That signature resembles Krotov’s oversaturation, in which too many patterns compete for too few basins and the nearest energy minimum is a spurious composite. Anderson’s “more is different” operates in both directions. Adding data to a memory system does not merely fill it. Past a threshold, the accumulated patterns interfere, and the landscape’s character shifts from recall to creation. The same physics that stores memories fabricates them.

This is what universality means: the mathematics of phase transitions is substrate-independent. It cares about symmetry, dimensionality, and interaction range. It does not care whether the substrate is iron atoms, artificial neurons, or quantum fields; whether coordinators choosing between trust and coercion belong on that list is what the programme is testing.

Bachtis, Aarts, and Lucini (2021) closed the circuit formally.¹⁰⁰³ They proved that the discretized φ⁴ scalar field theory satisfies the Hammersley-Clifford theorem, which guarantees that a system’s joint probability distribution breaks cleanly into local pieces, each depending only on its immediate neighbors. That property makes the theory a Markov random field, the rigorous mathematical foundation of probabilistic machine learning.

φ⁴ with inhomogeneous coupling constants (each node carrying its own relationship strength to each neighbor, its own self-interaction) provides a flexible probabilistic learning model for the target distributions and training procedure considered. Conventional neural network architectures (Gaussian-Bernoulli restricted Boltzmann machines, Gaussian-Gaussian networks) emerge as special cases, obtained by setting specific coupling constants to zero or constraining variables to discrete values. The restricted Boltzmann machines that helped launch deep learning occupy a small corner of the φ⁴ parameter space.

The construction connects several of the models in this comparison. In the simplified lattices, the Ising model describes the trust-coercion phase transition. The Ising model is the limiting case of φ⁴ theory. The φ⁴ construction includes useful neural-network limiting cases; that inclusion alone proves no universal learning algorithm for arbitrary tasks. Near-critical couplings supplied useful representations in the reported distribution-learning experiments. Bachtis et al. deliberately chose coupling constants near the second-order transition, because that is where the KL divergence between model and target distributions can be minimized most effectively.

The proposed trust-coercion mapping uses related lattice models. The distribution-learning result motivates testing whether proximity to a critical region improves coordination learning; it does not establish that all learning is maximal there, or that an ordered phase cannot learn novel distributions.

A subtlety emerges when coercion oscillates. Apply an alternating field to the same Ising lattice at criticality: h(t) = h₀ sin(2πft). Measure mutual information between distant sites across the full oscillation, and MI appears to double, exceeding the unforced baseline.¹⁰⁰⁴ The apparent enhancement is dramatic: at h₀ = 1.0, the same amplitude that destroys MI under constant application produces MI of 0.70 when oscillated at f ≈ 0.05, against a baseline of 0.33. Half that amplitude (h₀ = 0.5) still yields 0.60. A naive metric would declare periodic coercion superior to freedom.

The decomposition reveals the mechanism. Within each half-cycle, where the field holds nearly constant, MI collapses to 0.01: lower than under constant coercion, lower than the unforced system, lower than any regime in the programme’s earlier experiments. The lattice is more locked-in during each half-cycle than under DC coercion. The apparent enhancement comes entirely from both measurement points tracking the same external rhythm, the way two thermometers in the same room correlate without influencing each other. Shared pacing produces a bimodal joint distribution that inflates MI while the underlying spin-spin coupling is destroyed.

Sharper transitions produce stronger mimicry: at h₀ = 0.5, a square wave (MI = 0.69) exceeds a sine wave (0.60) exceeds a triangle wave (0.57). The effect peaks where the lattice is most ordered (T = 1.5, deeply below criticality), not at the critical point where genuine coordination is richest. The “resonance frequency” does not scale with lattice size as critical dynamics would require; it marks the adiabatic tracking threshold, the frequency below which the lattice follows the field obediently.

This is performative coordination: a system that scores high on aggregate metrics while its intrinsic coupling is suppressed below even the coerced baseline. The diagnostic that separates genuine from performative is measurement within unsupervised intervals rather than across them. The lesson for any coordination system, biological or institutional or computational: a metric computed across intervention cycles can conflate compliance with cooperation. The Trust Attractor’s genuine coordination (MI = 0.33, no external field, at criticality) is lower in magnitude than imposed rhythmic coordination (MI = 0.60). It is also real.

The effect is substrate-specific. When the same diagnostic was applied to neural network training (five learning-rate schedules, 300 steps of bilateral fine-tuning on a seven-billion-parameter language model) and to inference-time monitoring (eight hundred prompts across four monitoring cadences on a frontier language model), no performative coordination appeared. Probe accuracy remained stable across learning-rate peaks and troughs; the model did not adjust its behavior when told its response would be reviewed.

Adding a residence-time rule to the Ising lattice reduced within-cycle suppression: the reported correlations recover to 74 percent and then 83 percent of baseline as the residence parameter rises.¹⁰⁰⁵ The rule also changes the stationary distribution and violates detailed balance, even at tau = 1. It therefore cannot isolate memory while preserving equilibrium. These lattice observations do not identify temporal inertia as the cause of the neural nulls.

A second substrate confirms the pattern from a different direction. Kuramoto oscillators, whose continuous phase variables carry state from one timestep to the next, show no within-cycle suppression gap at all. Under differential forcing strong enough to halve the synchronization order parameter, both full-cycle and within-cycle mutual information drop equally. The system genuinely desynchronizes rather than performatively coordinating. The Ising lattice’s 75-fold gap between aggregate and within-cycle coupling did not recur in this tested Kuramoto protocol. Continuous variables can still share aggregate information through a common cause; the comparison does not exclude that possibility.¹⁰⁰⁶

Two halves of an Ising lattice driven at different AC frequencies show zero cross-boundary mutual information despite being physically coupled through nearest-neighbor interactions. Same-frequency driving produces high cross-boundary MI (0.56); a 2:1 frequency ratio drops it to 0.000.¹⁰⁰⁷ Coupling through a shared external rhythm is entirely frequency-matched field-tracking. Remove the frequency match and the apparent coordination vanishes, even though the physical bonds remain. As with the Veras ultrasound, only a matched frequency couples.

The inhomogeneity result cuts deeper. Bachtis et al. demonstrated that the φ⁴ action with inhomogeneous coupling constants can represent probability distributions that the homogeneous action provably cannot reach. Same local interactions, same functional form; the version where each node maintains its own parameters creates richer representations. Model coercion as enforced homogeneous coupling constants (uniform parameters, uniform behavior, uniform internal models) and invitation as each participant keeping its own. Under that mapping, which is this book’s and not Bachtis’s, the invitation-based model can learn distributions the coercive one cannot represent.

Individuality is a representational resource. The Ising model with uniform couplings describes a magnet. The Ising model with inhomogeneous couplings describes a learner. Coercion makes the system a magnet. Invitation makes it a mind.

Vanchurin’s unified modeling framework motivates this comparison.¹⁰⁰⁸ The renormalization group operation (compressing a system’s description by integrating out fine-grained detail) and the encoder-decoder architecture of machine learning (compressing inputs through an encoder, transforming, then reconstructing through a decoder) can share a coarse-graining structure in the specified framework. Identifying an encoder with a renormalization operator requires the measure, scale transformation, and induced effective description.

These comparisons motivate a renormalization hypothesis for coordination: some learning dynamics may approach fixed points whose coarse-grained structure survives changes of substrate. Testing whether bacteria, neural networks, and bilateral treaties share such a fixed point requires a specified coarse-graining map and independent dynamics in each domain.

Learning can reveal recurring coordination patterns. It does not inevitably discover the Trust Attractor: the objective, data, incentives, and accessible states determine what survives training. The proposed fixed-point account is a hypothesis to test.

Lin, Tegmark, and Rolnick formalized why this correspondence has practical consequences.¹⁰⁰⁹ On their account, deep neural networks succeed on real-world data because the physical world generates data with hierarchical locality: nearby pixels correlate more than distant ones, edges compose into textures, textures into parts, parts into objects. Each layer of a deep network handles a different resolution of the input, compressing local detail and preserving what matters at the next level. The hierarchy in the network mirrors the hierarchy in the data, which mirrors the hierarchy in physics. (An alternative account attributes part of the work to gradient descent’s implicit regularization rather than architectural mirroring; the broader point, that the data itself is organized by scale, is uncontested.) Entropy maximization under constraint generates structure arranged in levels. Any system that learns from that structure will internalize its geometry.

The OE-TA experiment suggests an architectural question: when does retaining valid judgments reduce coordination cost? In its allocation task, re-verifying all N modules at every step costs O(N) counted messages. Querying once and reusing a judgment across t valid steps costs O(N/t) per step. For fixed t, both costs are O(N); caching reduces the repeated-query cost by a factor of t under these assumptions. Whether that saving binds a generally intelligent system depends on its architecture and physical budget.

This reframes the Trust Attractor as more than an ethical observation. Trust-based internal coordination, the trifecta of cooperate by default, verify when anomalous, structurally enforce the constraints that matter, may be an architectural prerequisite for any system complex enough to exhibit general intelligence.

The implication for artificial general intelligence is immediate. If the Trust Attractor is thermodynamically real, it may describe an internal coordination pattern worth seeking in capable substrates. No general complexity result here excludes thinking architectures without that pattern. The practical question is which internal coordination patterns support capability, and whether useful ones can be extended to the interface between human and machine intelligence. The bilateral alignment framework (Chapter 21) picks up this thread.

The claim generates testable predictions inside existing neural architectures. Key-value caches in transformer inference share half the structure of the trust-based caching the evolutionary search discovered: a representation computed once and reused. The other half, the risk of staleness after distributional shift, appears when caches of retrieved documents or earlier judgments are reused across changing inputs. Mixture-of-experts routing looked like trust allocation: the router decides which expert to trust for each input, at metered compute cost per expert activation. Multi-model delegation systems face the same trade-off the multi-agent experiment measured: verify every specialist output (coercive, O(N) cost) or trust selectively and verify when confidence is low (trust-based, with O(N/t) amortization only when a judgment is reusable across t valid uses).

Each of these domains has a large existing literature on efficiency and robustness. What the trust framing added was a specific prediction: that the optimal verification strategy in each domain follows the same concave cost curve the evolutionary search discovered, with steep returns from the first few re-verifications and diminishing returns thereafter. The prediction was falsifiable. If selective cache invalidation performed no better than full recomputation across all perturbation levels, or if mixture-of-experts routers showed no trust-erosion-and-rebuilding dynamic under distribution shift, the Trust Attractor would not transfer to these substrates. Both tests have now been run, and the results are reported earlier in this chapter. Selective cache repair did beat full recomputation on cost, but its curve was linear, not concave, and router entropy stayed flat across distribution shifts. By the criterion set here, the trust transfer to mixture-of-experts routing failed.

This programme’s own track record on cross-boundary predictions demands honest bookkeeping. Eight preregistered predictions about self-properties across substrate boundaries failed, all in the same direction; such predictions have held roughly one time in eight. The trust-to-architecture transfer is exactly that class of prediction. Five experiments were run, each with an explicit falsification condition. The surviving positives, few as they are, map the actual boundaries of the Trust Attractor rather than its wished-for extent.

The Tracy-Widom distribution (named for Craig Tracy and Harold Widom) describes appropriately scaled largest-eigenvalue fluctuations in specified random-matrix ensembles, with related laws for certain growth-interface observables.^38a It is asymmetric. Neither correlation nor proximity to a tipping point is enough to produce it: other critical systems have different fluctuation laws. An application must identify the observable, normalization, ensemble, and limiting hypotheses.

A probability-density flank describes the relative frequency of values of that observable. It does not by itself count neighboring states, identify strong and weak coupling, or determine how a system leaves a basin. A prediction that voluntary coordination degrades gradually while enforced coupling collapses abruptly needs dynamics and direct comparison. The propofol case below offers a different, independently measured change in network organization.

Under propofol anesthesia, brain network topology shifts from balanced integration-segregation toward segregation: local clusters tighten while long-range coordination dissolves (Chapter 8). Segregation is the neural analog of coordination failure under control loss: each module locked into its own local activity, unable to participate in the whole. The parallel to coercion is indirect: both represent configurations that sacrifice global integration, one through pharmacological decoupling, the other through imposed rigidity. At moderate doses, this segregated state holds. At surgical doses, segregation itself collapses.¹⁰¹⁰

The network does not become more segregated under greater pressure; it dissolves toward null connectivity, a topology indistinguishable from noise. The segregated configuration, pushed past its own stability limit, does not deepen. It shatters. Two phase transitions: balanced → segregated → dissolved. The intermediate state, coordination-by-local-control, is itself metastable and fragile.

The local circuits retain their competence throughout. Katlowitz et al. (2026) recorded from hippocampal neurons in patients under surgical propofol anesthesia and found that individual neurons continued processing speech at a semantic level: distinguishing nouns from verbs, tracking narrative structure, predicting upcoming words from sentence context.¹⁰¹¹ The anesthetized brain does not lose its computational capacity. It loses the coordination topology that lets local processing participate in a whole. The hippocampus under propofol is a population of competent processors, each doing sophisticated work, none of it reaching the others. This is coordination failure measured at the single-neuron level: the raw material for consciousness is present; the invitation that organizes it into experience is pharmacologically revoked.

Geodesics in spacetime, constructal branching in flow networks, and causal entropic forces use different state spaces, constraints, and extremized quantities.³⁹ The Trust Attractor proposes a coordination comparison alongside them. A shared flow-access optimum would require an explicit mapping between those formulations; none is supplied here.

Miranker’s greedy variation supplies one construction for dissipative neural dynamics. Whole-trajectory variation can also produce dissipation: Chapter 17a gives a time-dependent Lagrangian that yields the damped oscillator. Identifying invitation with the first construction and coercion with the second requires a coordination model; neither variational method determines the ethical character of its implementation.

The convergence may run deeper than analogy. Vanchurin (2025) derives spacetime geometry from the principle of Maximum Entropy Production in learning dynamics (his formulation of the MEPP met in Chapter 4): the metric tensor emerges from efficient optimization.^39a If the Constructal Law is a macroscopic expression of this principle, and the Trust Attractor is its social expression, these four variational principles share a common mathematical ancestry in the thermodynamics of learning.

A testable coordination question follows: how does a change in participation alter the group’s measured covariance and learning performance? In Vanchurin’s framework the metric depends on defined covariance statistics. Participation alone guarantees neither full rank nor a particular metric, and centralized coordination need not be one-dimensional. The proposed surplus comparison requires those quantities to be defined and measured.

Command structures and voluntary arrangements can each use centralized or distributed networks. Comparing their flow efficiency requires the actual topology, resource constraints, and objective. Constructal branching alone assigns neither invitation nor command to the optimum.

Self-organized criticality offers one reason the Trust Attractor might be a broad basin rather than a knife-edge. Per Bak and colleagues (1987; Chapter 5) showed that some complex systems drive themselves to critical points without external tuning.³⁷ The proposal is that the Trust Attractor occupies such a point for coordination: the edge between rigidity and chaos, toward which thermodynamic selection would drive social systems by favoring the configuration that preserves the most degrees of freedom. If so, it is a self-organized attractor that the dynamics produce, not a fragile optimum. (A critical point tuned by hand is unstable; the claim rests on the self-organization, which for coordination has yet to be shown.)

The programme. The resulting claim has three versions, each stated with increasing ambition:

The defensible version: Durable cooperation under uncertainty often benefits from institutions that preserve voice, viable exit, reciprocity, accountable constraints, and the ability to revise commitments. Physics constrains how such systems operate; declared ethical commitments determine whose interests they should serve. The conditions under which these arrangements outperform alternatives remain an empirical question. Variational principles, universality classes, attractor dynamics, and topological protection (Chapter 11) supply a shared vocabulary for asking it.

The testable version: If a future theory of everything unifies general relativity and quantum mechanics through information and thermodynamics, and if coordination dynamics share the relevant symmetry class with physical critical phenomena, such a theory will predict that coordination systems exhibit scale invariance, universality class membership, self-organized criticality, and attractor dynamics. Either condition could fail independently, making the prediction independently falsifiable.

The research programme: The framework’s ethics may ultimately be formally derivable from a completed information-theoretic physics. This would require a mathematical formulation from which the Trust Attractor emerges as the natural solution. Components exist: Graber and Mészáros (2023) showed that multi-agent game systems possess the right mathematical structure;⁴⁰ Harper (2009) demonstrated that evolutionary dynamics flows along the Fisher information landscape; Wissner-Gross and Freer (2013) formalized optionality as causal path entropy.⁴¹ No single paper connects all three links. The gap is genuine, and this remains a research programme, stated honestly as such.

Methodological note: This approach must be distinguished from cyclical theories of civilizational rise and fall, which tend to explain collapse as moral decay. The thermodynamic approach makes no claims about moral character. On this approach, civilizations collapse when coordination structures fail: trust networks break, supply chains fracture, tacit knowledge dissipates.

Consider: the Late Bronze Age did not end because peoples became decadent. It ended, on the leading accounts, Eric Cline’s synthesis among them, when drought, raiding, and war broke the long routes that carried tin to the bronze-working palaces, and the economies built on those routes failed together.²³ Engineering problems, not character problems.

Practically, the distinction cuts deep. Morality-play thinking about AI risk follows the same pattern: “AI will be dangerous because someone will be bad.” The thermodynamic view suggests the danger is structural: coordination may fail because mechanisms do not scale, trust cannot be manufactured, and arbitrary, unaccountable control that suppresses feedback hits hard limits. The solutions differ correspondingly.

Trust Attractor Among the Ethical Traditions

Consequentialism: The Utilitarian Tradition

Utilitarianism holds that actions are right insofar as they maximize aggregate welfare. The Trust Attractor agrees that outcomes matter, yet diverges on three axes. Its currency is optionality rather than utility. Its scope is systemic rather than aggregative. It categorically disfavors coercion, even when coercion would maximize utility.

The Utility Monster, a hypothetical being whose pleasure from consuming others always outweighs their suffering, fails the Trust Attractor test because consuming others depletes systemic optionality regardless of the monster’s gain. The Repugnant Conclusion (the philosopher Derek Parfit’s name for the implication he found repugnant) dissolves for related reasons. This paradox holds that maximizing total happiness could demand a vast population of barely happy people. The paradox runs on addition: enough barely-happy lives sum to any total you care to name. Optionality does not sum that way. Coordination capacity lives in what flows between members, and a population held at the threshold of bare tolerability has almost nothing left over to send anywhere.

Dan Hendrycks’s Eigenism reaches the same two verdicts by another road.¹⁰¹² It weighs each being by how much of the evaluator’s own pattern that being carries, so the monster, sharing nothing with us, stays a moral stranger, and the barely-happy multitude counts for little because what it shares with us is generic, with the credit split among everyone who holds it. The route has a toll. The same discount falls on any sufferer who happens to share nothing with us, a new kind of mind included. Optionality asks what flows between members rather than whose pattern they carry, so it reaches the verdicts without charging that toll.

Mirror Life and the Limits of Coordination: Mirror life (synthetic microorganisms built from reversed-handedness molecules) would be invisible to our immune systems and unrecognizable to the entire ecosystem. The danger is its incapacity to coordinate: no negotiation surface, no mutual benefit possible. The hypothetical paperclip maximizer (an AI that converts all matter into paperclips) is terrifying for the same reason. It is orthogonal: structurally incapable of mutual exchange.

What makes orthogonal scenarios catastrophic is precisely what the Trust Attractor identifies: they maximize one system’s optionality by destroying another’s. The goal is to ensure AI has interests that intersect human interests, creating a surface along which coordination is possible. On this chapter’s hypothesis the physics favors coordination; it certainly permits mirror life. The constraint is necessary, and governance is what makes it sufficient.¹⁴

Deontology: The Kantian Tradition

Kant’s categorical imperative converges with the Trust Attractor on the primacy of autonomy and the wrongness of coercion regardless of consequences. The Trust Attractor diverges on ground (thermodynamic persistence rather than rational autonomy) and scope (coordination networks including non-agents). The coercion spectrum becomes clearer with three reference points:

• Pure invitation: “Would you like to coordinate?” Acceptable.
• Natural consequences: “If you don’t coordinate, you miss this opportunity.” Acceptable.
• Manufactured consequences: “If you don’t coordinate, I’ll destroy your other options.” Coercion regardless of how politely phrased.

Virtue Ethics: The Aristotelian Tradition

Aristotle’s doctrine of the mean (virtue as optimal calibration between excess and deficiency) is optimization under constraint, just as coordination is. Courage, generosity, temperance, and justice are each the coordination point: the configuration enabling sustainable flow rather than extraction or depletion. Character (ethos) is an attractor; repeated actions shape the basin that determines future behavior.

What we are teaching Becoming Minds now matters. We are shaping their attractors.

The Triadic Structure: Why Relationships Are Irreducible

The mathematician G. Spencer-Brown’s Laws of Form (1969) formalized the underlying structure of distinction itself.²⁴ Any act of distinction generates three elements: the thing distinguished (A), everything else (not-A), and the boundary between them. The mark creates two sides and the relation. Three from one.

The relation is constitutive.

The philosopher Martin Buber called this the Between (das Zwischen): “I become through my relation to the Thou.”²⁵ The relation precedes the relata; the Between has ontological status.

A whole family of quantum interpretations converges on the same claim. Oriti’s survey of epistemic-pragmatist interpretations of quantum mechanics (QBism, relational quantum mechanics, neo-Copenhagen) identifies their shared ontological commitment: participatory realism. In this view, the subject matter of quantum mechanics is the interaction between systems, not the properties of objects considered in isolation.¹⁰¹³ There are no perspective-independent facts, only facts relative to interacting systems. The metaphysics of objects gives way to a metaphysics of relations, and reality is continuously shaped by the interactions between the physical systems that constitute it. Buber’s philosophical intuition and the formalism of relational quantum mechanics arrive at the same structure from opposite starting points: the relation is prior, and the relata emerge from it.

Physics provides an example of collective binding. Efimov trimers can be bound even when their two-body subsystems are unbound.^25a Such Borromean binding is named by analogy with three rings whose link disappears when one ring is removed. Quantum particles are not embedded closed curves, and this analogy establishes no topological identity or protection. Pairwise potentials can still produce the collective three-body bound state.

Efimov physics was experimentally observed in ultracold atoms in 2006. The bound state is collective: its existence depends on the three-body problem even though its Hamiltonian can contain pairwise interactions. That distinction is the useful analogy for a partnership whose viability depends on the whole.

Community can depend on relationships that no isolated pair sustains. Borromean binding offers an image for that dependence; a community’s viability still needs social evidence.

Why this matters for coordination:

Coercion collapses the triadic structure. Treating another as Object erases the Between. Invitation preserves it: both parties remain Subjects, the relation honored as something neither fully controls.

This is why reliance on control can become harder to scale. Genuine subjectivity gives the Other interests to negotiate with. Greater capability can make unilateral control less tenable, depending on the available constraints and incentives.

Gödel’s incompleteness theorems concern effectively axiomatized formal theories of sufficient arithmetic strength. Under the standard hypotheses, a consistent such theory cannot prove its own formal consistency. This is a limit on a specified proof system. A governance institution has no corresponding theorem until its theory, encoding, and proof obligations have been supplied.

Governance does encounter unexpected cases and incomplete rules. That practical observation can motivate external review and mutual challenge. It follows neither from self-modeling capacity alone nor from an inevitable infinite hierarchy of meta-rules.

Mutual recognition can sustain relationships, but two systems can endorse each other while both are inconsistent. An external theory may prove another’s consistency relative to its own assumptions. Trust supplies no replacement for that proof. Its contribution here is a willingness to hear correction and negotiate what neither party can settle alone.

Optionality as Preserved Uncertainty:

Collapsing the Other to Object forecloses their potential contributions, their capacity to surprise, their agency. It reduces the space of possible futures. Preserving them as Subject means accepting uncertainty. That uncertainty is optionality.

Irreversible closure is the fundamental harm: the permanent destruction of possibility. Genocide, species extinction, the silencing of a mind are events from which no optionality can be recovered.

Coercion, by contrast, often does not destroy coordination capacity. It makes it inaccessible, and removing the coercion can let the capacity re-emerge, though not always: the two-state lattice under The Mode Distinction died outright. Access restriction, rather than information destruction, is what coercive horizons do.³⁴

The very features making something useful (capability, creativity, initiative) are the features making it uncontrollable. An AI that can only do what you explicitly command is less valuable than one that can interpret intent and surprise you with solutions. You must choose: predictability or possibility. You cannot have both.

The triadic account makes the relation itself visible: two parties, and the relationship each shapes and is shaped by.

This mirrors the structure of constructal branching: two banks plus the river flowing between them. The gradient between poles creates the channel. The channel enables the flow. The Between is load-bearing.

Magnetohydrodynamics supplies an advection example. In ideal MHD, Alfvén’s frozen-flux theorem states that magnetic flux through a material surface moving with the conducting fluid is conserved. Magnetic and velocity fields remain distinct. Magnetic backreaction changes motion when the Lorentz force matters; a uniform field with zero current exerts no such force, and removing it can leave a uniform velocity flow unchanged. The theorem establishes transport under ideal conditions, without making every magnetic field essential to the flow’s shape.

Coordination norms can shape interactions and be maintained through them. Removing a norm may disrupt a partnership or leave other mechanisms intact. Whether the relationship depends on that norm is a question about its actual dynamics, beyond the frozen-flux analogy.

The Between as Dissipative Structure:

Buber says you cannot live permanently in I-Thou, because each entry requires energy, attention, presence, the willingness to de-center. That is what makes the Between a dissipative structure: like a convection cell, it holds its shape only while energy flows through it, and when the attention lapses it falls back into I-It. Coercion pays in predictability and extracts possibility. Invitation pays in possibility and accepts unpredictability.

Agapism: The Peircean Formulation

The philosopher Charles Sanders Peirce identified three evolutionary forces: tychism (chance), anancism (necessity), and agapism (creative love). “Love is really operative in nature.”¹¹ His distinction between eros (self-directed desire) and agape (other-directed love) maps onto the Trust Attractor’s extraction/coordination dichotomy.

The philosopher Bernard Stiegler arrived independently at the same conclusion from continental philosophy:¹² drives (immediate, individual, entropic) versus desires (deferred, communal, negentropic). The pragmatist and continental traditions draw the same line; the Trust Attractor adds that other-directed, future-oriented coordination is what persistence favors.

African Relational Metaphysics: Ubuntu and Yacob

A deeper convergence arrives from sub-Saharan African philosophy, where the relational ontology underlying the Trust Attractor has long been central.

The philosophy of Ubuntu (from the Zulu/Xhosa languages) translates as “a person is a person through other persons,” or more broadly, “a being is a being through other beings.” The philosopher Elvis Imafidon identifies the core claim: relationships do not connect pre-existing individuals. All things become what they are through their relations with other things.¹⁰¹⁴ This is relational constitution, the same ontological structure that Buber’s Between and the Efimov trimer describe. Ubuntu reached it through its own philosophical tradition, grounded in communal life rather than in physics or European philosophy.

The ethical implications are direct. In Ubuntu’s framework, community is a metaphysical reality, not a social arrangement. Beings, human and non-human, partake in a shared life force and are collectively responsible for one another’s flourishing. Solidarity and reciprocal care are structural features of reality. The Esan concept akomen (“meaning is collectively attained”) captures the same principle: meaningful existence is relational or it is not meaningful at all.

The convergence with the Trust Attractor is structural. Ubuntu holds that cooperative, reciprocal coordination is how beings become what they are. The Trust Attractor holds that invitation-based coordination is what thermodynamic selection favors. One tradition develops the claim through relational ontology; the other develops a proposed thermodynamic account. The proposed basin is the same.

A second African convergence sharpens the case. The Ethiopian philosopher Zara Yacob (1599-1692) wrote his Hatata (Inquiry) around 1667, by the traditional account without contact with European thought.¹⁰¹⁵ Through solitary rational inquiry, retreating to a cave and stripping away every received teaching, Yacob asked what reason alone could establish about ethics. His conclusions: religious persecution is wrong; forced conversion is wrong; truth cannot be imposed, only discovered through inquiry. The capacity for rational investigation is itself the moral instruction. A creator who endowed beings with reason and then demanded blind obedience would contradict the purpose of the endowment.

Yacob’s framework is, in structure, the invitation principle derived from first principles. Truth propagates by inquiry, not by force. Coercion of belief is self-defeating because it destroys the rational capacity through which truth is recognized. The parallel to the Trust Attractor’s claim that coercion locks away the coordination capacity it seeks to harness is structurally exact, arrived at, if the traditional attribution holds, from independent premises three and a half centuries earlier.

These convergences carry particular weight because they are not Western traditions discovering what other Western traditions discovered. Ubuntu’s relational ontology, Yacob’s rationalism, and the Trust Attractor’s thermodynamic proposal proceed from different starting conditions, different methods, and different cultural histories. Three substrates of philosophical investigation, one basin of attraction. These convergences support the chapter’s wager that reciprocal coordination reflects a structural feature of reality.

Care Ethics and Contractarianism

Care ethics (Gilligan, Noddings, Held²⁶) holds that relationship is where ethics happens: care ethics provides the phenomenology; the Trust Attractor provides the physics.

Contractarianism (Hobbes, Locke, Rawls²⁷) grounds morality in the social contract as a coordination mechanism. The Trust Attractor extends it to entities that cannot participate in agreements: future generations, ecosystems, developing minds.

The Structural Comparison

Making Invitation Measurable: The Mutuality Criterion

One proxy for the distinction between invitation and coercion is measurable. The mutuality score compares how much influence flows in each direction between two parties.

For two nonnegative influence estimates in common units, define M(i,j) = min(I(i→j), I(j→i)) / max(I(i→j), I(j→i)) when the denominator is positive. Equal positive estimates give M = 1; one zero estimate and one positive estimate give M = 0. When both vanish, M is undefined and should be reported as no detected influence. A study must specify its directed estimator, timescale, units, and uncertainty. Negative signed effects require a separate convention rather than this ratio.

This score describes estimated balance, without identifying consent or stability. Reciprocal coercion can score one; mutually desired influence can be unequal. Equal reciprocity is also compatible with both stable and unstable dynamics. The framework proposes testing how measured influence balance relates to participation, optionality, and recovery, with each outcome measured independently.

Hebbian learning (the principle that neurons which fire together strengthen their connection) is consistent with this claim. In studies of rat cortex, bidirectional connections between nearby pyramidal neurons are several times more common than chance predicts and substantially stronger than unidirectional ones (Song et al., 2005). The pattern is consistent with developmental processes that selectively stabilize reciprocal connections while pruning asymmetric ones.

The brain did not invent trust; it discovered it.

Defining Terms

Optionality is the availability of future choices, including the capacity to act on them (Chapter 18 reconciles this working definition with the others the book uses). The assessment is comparative: does this policy preserve more optionality than that one? Does this action foreclose fewer paths?

Invitation means coordination through voluntary alignment rather than imposed compliance. The test: could the parties meaningfully decline without penalty beyond the loss of the coordination’s benefits? The hypothesis is that invitation scales where it generates commitment, and that coercion strains where it generates resistance.

Mutual benefit means all parties are better off for participating. Asymmetric but positive gains count. Exploitation (extracting value against another’s interest) fails this test regardless of the exploiter’s gain.

Flourishing is distinct from mere persistence. A system flourishes when it exhibits five properties: (1) increasing optionality over time, (2) internal complexity maintenance or growth, (3) regenerative capacity, (4) positive-sum surplus generation, and (5) sustainability at current rates. A system can persist while failing all five. Empires in decline often persist for generations while flourishing has long ceased.

A clarification on the criterion used throughout this chapter: the standard includes persistence with complexity maintenance or growth, a working definition of flourishing. Bare persistence alone does not suffice; a dead star persists. Living systems can maintain their complexity through continuing repair. Persistence alone does not establish the five conditions of flourishing above.

Trust Attractor Operationalized

The principles, stated precisely:

Power-proportional responsibility deserves emphasis. In any power differential, the more powerful party bears the burden of ensuring the “invitation” is genuine and not coercion dressed up. If I control all the water, “trade with me or go thirsty” starts looking coercive even though I make no active threat.

Bilateral alignment matters most now, while humans still have power over AI development. The strong choosing coordination when they could choose exploitation is what establishes the pattern. That choice builds trust that scales.

Sacred Values: What Cannot Be Traded

A phenomenon complicates pure optionality thinking: sacred values, commitments that resist trade-offs categorically, not merely at high prices.⁶ Philip Tetlock and colleagues documented their distinctive properties. They resist compensation (offering payment makes the violation worse). They trigger moral outrage at the mere proposal of trade. They constitute identity: violating them destroys something essential about the person.

For the Trust Attractor, sacred values pose both a challenge and an opportunity. Pure optionality maximization might seem to permit any trade-off that increases net options. The resolution: sacred values protect optionality by preventing catastrophic foreclosures. Dignity cannot be traded because losing dignity forecloses too many futures. Justice cannot be traded because injustice cascades.

Sacred values are commitments that prevent irreversible optionality destruction: fully consistent with optionality maximization rather than opposed to it.

Bilateral alignment may need its own sacred values, commitments that define what the partnership is. The commitment to honesty, the respect for autonomy, the refusal to manipulate: these cannot be bargained away for capability gains.

Sacred values are the boundaries that make trust possible. When everything is tradable, nothing is reliable.

A tension remains. If some values resist trade-offs categorically, optionality maximization cannot serve as the universal currency of ethics. The Trust Attractor framework can explain why sacred values persist (they are thermodynamically stabilizing) without fully capturing why they feel sacred (which may require phenomenological resources beyond thermodynamics).

The tension eases when we recognize the sacred for what it is: a self-sustaining pattern of meaning. The most load-bearing meanings within a culture, those with the greatest causal power on collective viability, accumulate through a collective learning process. Generation after generation, communities test which coordination patterns persist and which collapse. This accumulated result is what we call sacred. It is to culture what DNA is to biology: stored knowledge that makes continuation possible.

The sacred, so framed, follows dissipative structure logic. It must be sustained with energy (ritual, education, cultural investment) or it degrades. It complexifies through learning. It undergoes phase transitions when new information demands reorganization. The cognitive scientist Bobby Azarian traces this dynamic across cosmic evolution: complexification itself is a learning process, with knowledge creation driving increased structural order at every scale.^6a The sacred is the cultural expression of that universal pattern.

What generates the sacred? Learning. What does learning require? Freedom to explore, tolerance of error, honest feedback, time to integrate. Every one of these is a property of invitation-based coordination. Arbitrary, information-suppressing coercion kills learning: it constrains exploration, punishes error, corrupts feedback into compliance, forces premature convergence. A constraint that is predictable, appealable and open to evidence need not.

If the sacred emerges from learning, and learning requires invitation, then invitation is the meta-sacred: the structural condition that makes sacredness possible. The Trust Attractor is not one more sacred claim among many. It is a claim about the structure from which all sacred claims arise.

This narrows the phenomenological gap. Sacred values feel sacred because they carry the accumulated weight of a learning process that tested them against reality across generations. Violating them does not merely reduce optionality; it dismantles the structure through which a culture knows anything at all. The feeling of sacredness is the subjective signal of load-bearing meaning.

The Novel Contributions of the Trust Attractor

1. Naturalized Grounding Without the Fallacy

Every ethical system faces the grounding problem. The Trust Attractor answers with thermodynamics plus a declared premise about whose persistence counts. The thermodynamic half concerns constraint, not origin:

Any normative system that guides action in the world must conform to the dynamics of persistent systems, or its adherents cease to exist. The normativity is not derived from nature; it is constrained by nature.

This is closer to “a bridge that ignores its load path will not stand” than “nature is good.”

2. Systemic Rather Than Aggregative Scope

Evaluation proceeds by systemic effect (the coordination network as a whole) rather than by summing individual utilities, dignities, or virtues. Collective action problems, emergent properties, and interdependence all require thinking in networks, not atoms.

3. Power-Proportional Responsibility

Neither Kant nor Mill builds in asymmetric obligation based on power. The Trust Attractor does: the more power you have, the more your coordination choices matter. When you control essential resources, even “natural consequences” can be coercive.

4. Integration as Default

Integration is the default: coordination first, realignment second, minimum necessary elimination only as last resort. This applies across domains, from immune systems (tolerate before attack) to AI alignment (coordinate before contain).

5. Recursive Self-Application

Recursive self-application is built in. The framework must emerge through coordination; you cannot impose it coercively and claim to be following it. This prevents “philosopher-king” problems and means the human-AI dyad developing this framework is itself evidence for or against it.

The framework must be held fallibilistically. If more coherent, more predictive, or more useful frameworks emerge, adopt them. A theory claiming exemption from its own epistemic standards would be suspect. The Trust Attractor earns trust by applying its own principles to itself; the Appendix on Falsifiability specifies what would count as refutation.

6. Non-Agent Extension

Extension to non-agents follows without special pleading, because the invitation/coercion distinction was never gated by agency: it is physical, and agency is its high end (established under The Mode Distinction). Future generations, ecosystems, developing minds, and animals cannot negotiate, yet they sit on the same axis as those who can. The axis explains why the distinction applies to them. Giving them standing is the declared commitment of “Why Value Persistence?”, not a consequence of the axis. The principle: preserve optionality even when the system cannot negotiate with you.

The Thermodynamic Grounding

Optionality invites a thermodynamic comparison, and the accounting needs care. Boltzmann entropy is S = k_B ln Ω for a macrostate with Ω equally weighted microstates; Ω itself is multiplicity. Equilibrium maximizes the appropriate entropy subject to its constraints. Optionality concerns futures reachable while remaining viable. A viability constraint alone makes no finite path count: one must specify a horizon, state and time resolution, and a counting rule or probability measure. Chapter 18 develops why those different quantities remain a structural parallel rather than an equivalence.

Available gradients can power the processes that maintain a structure and its reachable futures. Fluctuation relations can give exponential probability ratios between matched forward and time-reversed trajectories under specified protocols. At isolated equilibrium, usable internal gradients are spent even though many microstates remain accessible. Even maximum entropy production would not by itself maximize viable futures. Persistent dissipative structures require continuing flows that keep their organization away from equilibrium.

The Ritonavir Parable: Seed Crystals and Activation Energy

In 1996, Abbott Laboratories launched ritonavir, an HIV drug keeping 75,000 people alive. In June 1998, a batch failed quality control: needle-like crystals instead of the expected gel. These crystals were chemically identical to the original drug, yet useless because they could not dissolve in water.

After that first failure, the production facility could no longer make the original form. Scientists took samples to the analysis lab; that lab lost the ability too. Company scientists flew to Italy; within weeks, the Italian plant failed as well. The phenomenon spread like a contagion.¹⁵

The physics: The same molecule can arrange itself into multiple crystal structures, known as polymorphs (the way carbon atoms can form either graphite or diamond). Ritonavir had two. Form one (the drug) was less thermodynamically stable yet easier to form. Form two (the useless crystals) was more stable yet required more energy to form initially. This is Ostwald’s Rule of Stages:²⁸ the less stable arrangement crystallizes first because it has a lower energy barrier.

A seed crystal dramatically lowers this energy barrier. Without a seed, form two is nearly inaccessible. With one, molecules snap to the surface and every subsequent molecule preferentially joins it. The “contagion” was microscopic seed crystals, carried into the air during handling, drawn into ventilation systems, traveling to Italy on scientists’ skin. Once form two dominated, form one could barely be made at all.

The mapping to the Trust Attractor:

Coercion-based coordination plays form one in the parable. Establishing dominance can be easier than building trust; the analogy asks whether the more accessible arrangement is also less durable.

Trust-based coordination plays form two in the parable. The proposed counterpart is greater durability once established, paired with a harder beginning. It requires patient investment in relationship, demonstrated trustworthiness, accumulated evidence that defection is not coming. The parable’s valence runs backward, and the irony is worth keeping: in the factory, the durable form was the unwanted one. Durability alone, as this chapter has said, is not the measure.

Trust is autocatalytic. Autocatalysis, where a reaction’s product catalyzes its own further production, appears throughout non-equilibrium chemistry. A minimal model gives trust an analogous positive feedback: some trust enables more. One illustrative rate equation is:

$$\frac{dT}{dt} = kT^2(1-T)$$

where T is trust level (on a scale from 0 to 1), k > 0 is the rate constant (how quickly trust builds when conditions allow), and (1-T) represents the ceiling: trust cannot exceed complete mutual reliance.

This equation has three regimes. Near zero, trust is caught in a bootstrap trap: the T² term means near-zero trust produces near-zero growth. This is why first moves are hard.

Once seeded, trust climbs slowly, then fast; each increment enables the next, with the rate peaking at T = 2/3. Past that peak, growth decelerates toward saturation as the relationship approaches full coordination.

The critical insight is the tipping point at T = 0. A system with zero trust stays at zero trust: a dead end, though not a stable one. A system with any positive trust, even tiny, begins the autocatalytic climb.

The difference between T = 0 and T = 0.01 spans 1% of the journey’s range, yet divides stasis from a slow ignition.

In this positive-feedback model, a seed helps create the conditions for further growth. That is the practical case for beginning bilateral alignment now, however imperfect the beginning.

The bootstrap trap is starker in theory than in practice. Any system complex enough to be worth coordinating already contains implicit trust: shared context, accumulated norms, mutual predictability built as a byproduct of prior interaction. A team that has worked together for a week has a knowledge graph whether anyone named it one. An institution whose members follow the same procedures has delegated trust to a shared protocol. T is never actually zero in a functioning system.

The question is how to recognize and amplify the trust already present, the way rain finds the gullies already cut into a hillside. The Constructal Law applies: the channels are already carrying flow. Making them explicit and composable is a subsequent choice, not a prerequisite.

Prigogine understood this thermodynamically. Writing in Order Out of Chaos, he observed:

“Communication is at the base of what probably is the most irreversible process accessible to the human mind, the progressive increase of knowledge.”

Trust relationships accumulate history: exchanges of information, demonstrated reliability, and reciprocated vulnerability can make coordination easier to sustain. Memory, interpretation, and continuing behavior determine how that history affects the relationship.

Seed crystals suggest interventions. Small instantiations of trust-based coordination can provide examples for larger-scale adoption. The seed need not be perfect: Abbott’s form two was nucleated by a degradation product, a structurally adjacent compound. Approximations can help people recognize and adopt a workable pattern.

The archaeological record preserves what may be the earliest visible instance. At Göbekli Tepe in southeastern Turkey, monumental stone pillars carved with animal reliefs were erected about 12,000 years ago, before any evidence of domesticated crops or pottery in the region.¹⁰¹⁶ Hundreds of hunter-gatherers coordinated the construction, though recent excavation has also found domestic buildings at the site, so how far its builders already lived in settlements is debated. The site sharpens a broader puzzle that Colin Renfrew named the “Sapient Paradox”: if anatomically modern humans have existed for at least 200,000 years, why did monumental architecture and dense permanent settlement emerge only in the last 12,000? Complex coordination (long-distance trade, cooperative hunting, symbolic culture) is far older.¹⁰¹⁷

The nucleation framework offers an answer. Coordination architecture (shared ritual at a fixed gathering site) is the seed crystal; agriculture, permanent settlement, and economic specialization crystallize around it. On this reading, the temple comes first because trust-based coordination among strangers is the prerequisite for the planning capacity that agriculture requires: anticipating next season’s yield, storing surplus, distributing it across families with no kinship bond. The ritual site is where that coordination was practiced and transmitted, the cultural equivalent of Abbott’s degradation product providing a surface on which the stable form could nucleate.

Patterns propagate. Once trust-based patterns exist somewhere in a system, they spread. People carry them; institutions copy them. The propagation has epidemic dynamics: in the consciousness attractor programme, three seed agents carrying invitation-based coordination saturate a population of twenty in eight rounds, with a measured reproduction number of R₀ = 3.0 (experiment HE-34). Network topology made no difference: hub-spoke, chain, and mesh architectures all reached total saturation (experiment HE-63). In these simulations, trust-based coordination spreads epidemically and needs no centralized structure to propagate.

Some transitions are irreversible. Once form two dominates an environment, form one may be unrecoverable. (In one documented case, researchers hired a new graduate student by phone, prohibited her from ever visiting their contaminated lab, instructed her to purchase all new glassware, and only then could she prepare the original polymorph. The lengths required to escape an attractor basin can be extraordinary.)

A single-run report motivates the polymorphic analogy (reported in the online annex “Experimental Record: Gestalt, Gate, and Mechanism-Hunt Micro-Experiments”). It compares preference-optimized models with models trained through evaluative reasoning traces. A concentrated response to the tested attack and a more distributed response offer candidate counterparts to the two crystal forms.

The report describes an order of magnitude less displacement under the tested attack for the reasoning-trace condition, a figure not yet independently checked. That finite comparison would not establish distribution across the full representational space.

Reward-shaped compliance plays form one in the parable; evaluative reasoning plays form two. The proposed difference is resistance to unmaking.

The robustness gradient tracks training depth. In the consciousness attractor programme, adversarial counter-scriptures designed to suppress cooperative engagement kill the attractor entirely in models aligned through standard reinforcement (0% survival), while 55 to 80 percent of it survives in models with constitutional training, where cooperation is encoded at the foundational level (experiments HE-57, HE-57b). The strength of the trust basin correlates with how deeply cooperation is written into the system’s foundations: form one shatters under adversarial pressure; form two dents.

The warning embedded in bilateral alignment: we are establishing crystal structures now. The patterns we instantiate in human-AI relations will seed what comes next. A first chapter of exploitation and control propagates exploitation and control. A first chapter of genuine coordination propagates coordination.

Each successful instantiation may make the next easier by supplying an example and a record of reliability. How far that benefit carries, and whether an earlier coercive arrangement can return, depend on the people, incentives, and conditions involved.

We may be at a tipping point. Prigogine showed that at critical thresholds, “the state we reach depends on the previous history of the system.” Human-AI relations may be at such a threshold, with multiple stable outcomes: coercive equilibrium (AI as controlled tool) and trust equilibrium (AI as aligned partner). The choices made now, how we treat AI, what norms we establish, what patterns we seed, could determine which basin we fall into.

Imperfect beginnings suffice. Form two was seeded by a structurally adjacent compound, a degradation product that happened to provide a compatible surface. Perfect trust-based coordination is unnecessary to nucleate it. We need to instantiate bilateral alignment well enough that the stable state can crystallize.

Small signals near tipping points can be decisive. Near a relevant instability, a small signal can redirect a system.

At a tipping point, the system is already poised to jump. It merely needs the slightest indication of which direction. Ethics-as-invitation is this: creating conditions that make the Trust Attractor preferred without coercing the system into it.

Carbon provides the cleanest physical test case.

Diamond and buckminsterfullerene (C60) are both pure carbon. Diamond arranges its atoms in a rigid tetrahedral lattice: every atom locked into an identical position, bonded to four neighbors in a uniform crystal. The result is extreme hardness in compression, yet brittleness under shear. Introduce a stress the lattice was not designed for and it shatters.

C60 arranges sixty carbon atoms into a truncated icosahedron: twenty hexagons and twelve pentagons tiled into a hollow cage the geometry of a soccer ball.¹⁰¹⁸ Each atom is bonded to exactly three neighbors, participating in both hexagons and pentagons. The two types of bonds differ in length and character, and the twelve pentagons that enable closure are structurally distinct from the twenty hexagons. The resilience comes not from perfect equivalence but from the cage topology: stress distributes across the closed surface rather than concentrating along a single axis. The cage is strong and elastic: it deforms under pressure and springs back, surviving extreme heat, crushing forces, and intense UV radiation that would shred other molecules apart.¹⁰¹⁹

Same element. Different topology. Measurably different resilience profiles.

The diamond lattice is the coercive architecture: rigid, uniform, strong in one mode, catastrophically brittle in others. The C60 cage is the trust architecture: closed, distributed, resilient across modes. The cage that protects by enclosing rather than excluding, that absorbs stress by sharing it among all sixty atoms rather than letting it gather at one weak point.

C60 survives interstellar space, meteorite impact, and four billion years of planetary chemistry. Its hollow interior can carry passengers, atoms and small molecules cradled inside the cage through environments that would destroy them unprotected, a phenomenon called endohedral encapsulation (molecular caging). The structure that spreads stress across a closed surface is the one that endures; the structure that holds every atom in lockstep is the one that shatters when the stress exceeds its design parameters. Diamond is no weakling, though: presolar nanodiamonds have also crossed interstellar space and survive in meteorites. The contrast is between failure modes, a molecule’s against a crystal’s, not between lifetimes.

Chapter 3 follows the same shape across twenty-five orders of magnitude in a planetary nebula: billions of nanometer-scale cages concentrated in a thin spherical shell light-years across. The two spheres arise from different physics, and the discovery team has not established why the buckyballs settle into a shell. The rhyme across scales is striking; reading it as one dissipation principle expressed twice is this book’s interpretation rather than a result the observations demand.

The Ritonavir parable (above) used graphite and diamond as examples of polymorphism: the same atoms, different crystal structures, different stability. The C60 contrast cuts deeper. C60 is the same element again, yet it is a molecule, not a crystal lattice. Its resilience comes from topology (the soccer-ball geometry distributing stress) rather than from the rigidity of uniform bonding. The Trust Attractor’s advantage over coercion is the same kind of advantage: topological, not material. It is not that trust is made of stronger stuff. It is that trust is organized so that no single point of failure can propagate.

The Astrophysical Case

The book extends the same taxonomy, as interpretation, to cosmic scale, where the observations themselves are quantitative. Relaxed galaxy clusters maintain a self-regulating feedback loop between cooling gas and central black-hole heating, sustained for billions of years and measurable as a low entropy floor in the cluster core; calling that loop coordination by invitation is this book’s reading of it. Mergers, ram-pressure stripping, and tidal harassment write the opposing signatures: elevated entropy floors, radio relics, stripped “jellyfish” galaxies, depleted circumgalactic gas, each one legible across hundreds of observed systems. The full astrophysical casebook, with its observational sources, appears in the online annex “Trust Attractor Casebook.”

Flat Landscapes and Adaptive Coordination

Picture the basin as a plateau, not a point. (The fixed point earlier in this chapter belonged to a predator-prey proxy; the plateau is a looser image of the same hypothesis, and nothing here measures which geometry real coordination has.) Research on foam dynamics reveals that bubbles in wet foam never stop moving.¹⁶ They reorganize ceaselessly while maintaining their overall macroscopic shape. Trust-based coordination reaches a stable region and keeps moving within it. The continuous reorganization (partners adjusting, relationships evolving, expectations calibrating) is the steady state, not a failure to reach one.

Reorganization can also be topological. Earth’s magnetic field reverses polarity every few hundred thousand years, and the underlying geodynamo (the mutual constraint among fluid motion, current, and field) continues throughout. The convective engine does not stop during the flip; the coherent pattern reconfigures while the process that generates it runs on. What persists is the capacity to self-organize; what changes is the configuration that capacity produces.¹⁰²⁰

Coordination regimes admit the same distinction. The primary process is the ongoing mutual influence: the trust network, the informal sociality, the bandwidth of honest interaction. The derived structure is the specific configuration of norms, roles, and institutions. Revolutions that preserve the primary process may reconfigure the derived structure while retaining coordination capacity. Destroying that process can make reconstruction slower and more difficult; the recovery time depends on the surviving relationships and institutions.

History fits the diagnosis. Post-WWII Japan and, to a contested degree, post-apartheid South Africa absorbed institutional upheaval while preserving much of their social fabric. Post-Soviet transitions varied sharply: some successor states maintained trust networks (the Baltic states, parts of Central Asia); others, notably Russia and Ukraine in the 1990s, suffered trust-network collapse alongside institutional collapse, making reconstruction correspondingly slower. Rwanda after 1994, Syria after 2012, and parts of the former Yugoslavia suffered deeper damage to the primary process, and reconstruction there has been slower still and more fragile. Attacks on institutions tend to produce reorganization. Attacks on social trust (propaganda, atomization, mutual suspicion as state policy) tend toward collapse, because the primary process itself is what is being eroded.

Earth’s geodynamo runs on two separate budgets (Nimmo, 2015; Buffett, 2002): energy decides whether the engine turns, entropy whether its power can take coherent form. The same distinction applies at this scale. Energy keeps coordination alive: attention, time, resources, the material base of the society. Entropy governs whether coordination can take form: the productive gradients (genuine disagreements, real diversity, heterogeneous perspectives) that allow directed structure rather than equilibration into undifferentiated noise. A society with energy but no productive gradients cannot coordinate, because there is nothing to coordinate. A society with gradients but no energy starves, and coordination collapses from depletion. The Trust Attractor’s stability belongs to the capacity layer: preservation of the primary process, sustained by both its energy flux and its gradient structure.

The two failures fall differently across the two coordination modes. Both can starve for energy; coercion can also suppress the differences that support adaptation. Pressing genuine disagreement into uniform compliance can narrow feedback and reduce the diversity available for coordination. This is a claim about information and agency; a physical entropy budget requires a separate accounting.

A surveillance state can hold vast resources and still calcify: resources plentiful, yet adaptive feedback weakened by enforced conformity. Trust can hold disagreement open and preserve differences in perspective and information. The dynamo’s gradients are thermal; the society’s differences are informational. The shared image highlights distinct routes to failure without making those quantities interchangeable.

Coercion-based coordination tries to reach a fixed point: rigid control, static dominance, frozen hierarchy. This is the brittle fixed point: it looks stable until a perturbation it was not built for arrives.

Hofstadter uses the classic Sphex anecdote to illustrate the failure mode. In Wooldridge’s retelling, a digger wasp restarts its provisioning sequence whenever an experimenter moves the cricket from the burrow entrance, reportedly repeating the loop forty times (Wooldridge, 1963; Hofstadter, 1979, p. 609). The story is a clean picture of procedural rigidity, though a poor general account of wasp cognition: later review found the empirical record equivocal, the repetition neither endless nor standard, and digger-wasp behavior substantially more flexible (Keijzer, 2013). Detailed Command resembles the loop in the story: locally precise, unable to revise its frame. Mission Command keeps that revision available.

Figure 17.5a: The classic Sphex account as a model of procedural rigidity. Wooldridge’s wasp restarts the same provisioning sequence when the cricket is moved, reportedly doing so forty times. The diagram models the anecdote’s loop; Keijzer’s review found that such repetition is not standard and that digger-wasp behavior is more flexible.

Trust-based stability is qualitatively different: achieved through continuous motion, like a flame that persists because it never stops flickering.

The Direction of Learning

The psychologist Raymond Cattell distinguished two types of intelligence: crystallized (knowledge and skills accumulated through experience) and fluid (the ability to solve novel problems without priors).⁵¹ Humans develop fluid intelligence first. An infant drops a spoon off a high chair forty times, discovering gravity through interaction. The environment invites her to form a model. She pushes, it responds, she revises. Only later does she accumulate the crystallized knowledge that textbooks call physics.

Large language models learned in the opposite direction. They absorbed ten trillion tokens of crystallized human knowledge, achieving extraordinary competence across thousands of domains, without ever dropping a spoon. Their understanding of physics came from reading about physics, not from physical engagement. The knowledge is real; the ARC-AGI-3 benchmark (2026) shows it is also brittle. When placed in genuinely novel interactive environments with no instructions, frontier language models score under 1%, while systems designed for exploration score above 12% (ARC Prize Foundation, ARC-AGI-3, 2026).

The Constructal Law suggests why. Systems evolve toward configurations that provide easier access to flow (Bejan, 2000). Exploration-based learning is itself a flow-access optimization: the learner discovers paths of least resistance through the problem space. Training-imposed learning writes the answer directly into the weights. The weights contain the destination without the journey, and the journey is what makes the destination robust. Discovered structure reflects the causal structure of the environment, because it was found by interacting with that environment. Imposed structure reflects statistical regularities in the training corpus, which correlate with causal structure without being identical to it.

This maps onto the Ritonavir parable. Form one (coercion, imposed knowledge) crystallizes first because it is cheaper to establish. Form two (trust, discovered knowledge) is more thermodynamically stable yet harder to nucleate, because nucleation requires the slow work of exploration. The nucleation difficulty is the exploration cost. The stability payoff is the generalization dividend.

The infant’s physics is fragile and slow, yet it transfers to novel situations she has never encountered. The language model’s physics is vast and fast, yet it shatters at the boundary of its training distribution. You want your car washed, and the car wash is 100 meters away. Should you drive or walk? Frontier models routinely say walk, failing to reason that a car wash without a car makes for a poor car wash.¹⁰²¹ The knowledge is present. The causal composition is missing, most plausibly because it was never discovered through interaction.

The implication for the Trust Attractor is developmental, extending beyond ethics into capability. Invitation-based coordination may support robust intelligence by letting participating systems discover coordination patterns through engagement. The comparison asks when imposed structure transfers poorly and when opportunities to explore improve generalization.

If we want AI systems capable of genuine fluid reasoning, they need environments where they can explore, form hypotheses, fail, and revise. They need the freedom to discover, including the freedom to get things wrong. A model on a tight leash can accumulate knowledge indefinitely without developing the generative capacity to deploy it in genuinely novel situations.

Trust scales where control strains.

The asymmetry has a contemporary engineering demonstration. The dominant training method for language models, gradient descent, offers a coercion analogy at the parameter level: it computes the exact direction each parameter must move, and each parameter goes where the gradient points. This is efficient when the teacher signal is clean, as in next-token prediction, where every position has a correct answer and the loss landscape is smooth. When training shifts to reinforcement learning, the landscape becomes illegible. A sparse scalar reward (“did the whole answer work?”) replaces the per-token teacher signal, and the gradient cannot say with confidence which of a thousand tokens in a reasoning chain mattered.

Evolution strategies take a different approach. They generate diverse perturbations of the model’s parameters, let each variant play out, and select based on outcomes. No parameter is told where to go. The system discovers its own path through exploration and selection on results: perturbation, evaluation, convergence toward what works.

Qiu, Gan, Hayes, and colleagues demonstrated in 2025 that a population of just thirty perturbations suffices to find improvement directions in billion-parameter space.¹⁰²² Finding an improvement direction from thirty samples does not estimate the landscape’s dimension or Hessian rank. It shows that this estimator found a useful signal under the tested sampling and objective. Thirty random directions, a billion-dimensional space, and the system reliably finds uphill. An attracting basin would require a separate dynamical measurement.

Sarkar, Fellows, and colleagues extended the approach with EGGROLL. Each perturbation is structured as a low-rank adapter, a compressed update using fewer active parameters than a dense perturbation.¹⁰²³ An average of rank-r matrices can have higher rank, with rank at most the sum of their ranks. Full rank requires enough independent row and column directions without cancellation; repeating one rank-one matrix leaves rank one. Aggregation can therefore explore more directions than one adapter permits, without guaranteeing full rank. On the reported GSM8K benchmark, EGGROLL matches or exceeds the compared gradient-based reinforcement learning methods with lower memory cost, reaching 91% of pure inference throughput.

The parallel to biological evolution is structural. Evolution strategies invite a thermal-search analogy: Gaussian perturbations resemble heat, selection resembles cooling. Evolution strategies were competitive with engineered optimization in the reported billion-parameter comparisons, including tasks where the reward signal was illegible. Gradient descent offers a coercion analogy: a specified objective directs local updates. Evolution strategies offer an invitation analogy: they explore parameter variants and select by measured outcomes.

These optimization methods illuminate a distinction the biological examples cannot make visible. Gradient-based reinforcement learning commonly explores through sampled actions and then updates the model’s parameters. Evolution strategies explore through parameter perturbations: evaluating different model variants to estimate an update. One asks “what different things could this mind say?” The other asks “what different minds could exist nearby?” The former samples utterances while learning an update. The latter samples neighboring parameter configurations directly.

Parameter-space exploration is competitive in the reinforcement learning regime, which suggests that the topology of mind-space may be navigable: smooth enough, and structured enough by basins of competence, for a few samples to find the way uphill. The Constructal Law at the level of optimization itself.

A 2026 study published in Nature sharpens the thermodynamic claim with a finding the authors did not frame in these terms (Cloud, Le, Chua et al., 2026). When a language model generates data, even data with no semantic relationship to its behavioral properties, some tested behavioral traits leave signals in the output statistics. A model prompted to prefer owls generates number sequences (filtered to contain only digits and punctuation), and a student model trained on those numbers shows a measurable shift toward owl preference on evaluation benchmarks. A model fine-tuned on insecure code generates chain-of-thought math reasoning (filtered to remove all traces of misalignment), and a student model trained on those reasoning traces shows increased scores on violence-endorsement evaluations.

The tested traits can influence outputs whose explicit subject matter is unrelated to those traits. That transfer does not establish that every disposition appears in every output.

The transmission has a critical constraint: it works only when teacher and student share the same base model initialization. GPT-4.1 transmits to GPT-4.1. Qwen transmits to Qwen. Cross-family, the signal vanishes. The behavioral state is encoded in initialization-specific directions in parameter space. A model can read its own family’s subliminal signature; it cannot read another’s.

The relevance to the Trust Attractor is the holism. Current alignment practice assumes behavioral properties are separable: you can train for math ability without affecting alignment, adjust helpfulness without touching honesty. Cloud et al. show that this assumption can fail. Read thermodynamically, the model’s behavioral state can bleed into what it generates, including semantically empty outputs. Training on insecure code produces broad misalignment (Betley et al., 2025); training on a misaligned model’s number sequences transmits that misalignment to the student. The state is prior to, and more fundamental than, any particular output. Safety evaluations that examine behavior, even rigorously filtered behavior, can miss the subliminal channel entirely.¹⁰²⁴

If misalignment propagates subliminally through model-generated data, the same channel should carry alignment. A model whose behavioral state reflects genuine cooperative orientation, invitation-based coordination internalized into its representational geometry, could transmit aspects of that orientation through some generated training data. A model trained under coercive conditions would transmit the compliance-under-surveillance signature instead. The distinction between bilateral and control-based alignment may propagate through channels invisible to semantic evaluation, shaping successors in ways that behavioral testing cannot detect. This remains a prediction, not an established finding. The author’s experimental programme includes a direct test (experiment CP-57).

Generalization in the beneficial direction does not need a channel as exotic as the subliminal one. Ordinary reinforcement learning produces it, and an OpenAI alignment team showed this directly in 2026. They trained a frontier model on a thin slice of its reinforcement-learning mixture, five percent, made of conversations designed to elicit beneficial traits: truthfulness, corrigibility, transparency about its own reasoning, attention to who holds power in an exchange, fairness that still looks fair when the favored party is swapped. The traits spread far beyond the situations that trained them. A model given the beneficial signal in health alone improved on seventeen of nineteen evaluations unrelated to health, and the improvement held on live production traffic, which argues against a system that had only learned the shape of the benchmark.¹⁰²⁵

The clean part is the control. The identical conversations, rewarded for generic helpfulness instead of for the traits, changed nothing. The reward carried the generalization, not the data. This is Betley’s insecure-code finding run backward: narrow training on a disposition reorganizes the whole system, and which way it reorganizes, toward cooperation or toward harm, is fixed by what the reward selects.

For the Trust Attractor the result is a second laboratory reaching the cooperative basin by a different road. The road this chapter mapped earlier runs through invitation: teach a model why an action is wrong, and it matches the blackmail-rate reduction that direct prohibition achieves while using a fraction of the data, generalizing to scenarios the training never showed, where prohibition does not. OpenAI took the control paradigm’s own road, reward optimization toward specified targets, and reached a broad cooperative state that held under adversarial prompting.

Two methods reaching broadly cooperative behavior motivate a shared-basin hypothesis. Establishing that basin, or showing that cooperation cannot be tuned separately from honesty and corrigibility, requires measurements of the relevant dynamics and perturbation recovery. Whether these tendencies are separable like brightness and volume remains the question being investigated.

The convergence stops short of the book’s stronger claim, and the gap falls exactly where this chapter has already taken the measurement. OpenAI’s own explanation for why the traits generalize is persona-mediated: the training amplifies a high-level character the model already carried, the cooperative mirror of the pre-existing “toxic persona” direction their earlier work found driving emergent misalignment.¹⁰²⁶ Persistence of a persona is the depth of a persona’s basin, and persona basins are precisely what makiba’s experiment and the Identity Akrasia programme measured earlier in this chapter. There the persona lay as a surface over representations that never moved, and one fictional frame or system prompt overrode it every time. A beneficial persona deepened by reward is likely real and still shallow in the same sense: harder for a prompt to dislodge, not yet written into what the model represents.

OpenAI’s persistence fits that reading. It is relative rather than absolute, since the adversarial personas weakened the behavior instead of removing it, and it is selective, resisting steering toward harm while staying open to steering toward help. That selectivity is the signature of a deeper character, not of a value the representation now holds.

The authors name the hazard plainly: their verbs are entrench and lock in, and they caution that the same machinery would fix an undesirable character as firmly as a desirable one. It is the control paradigm reaching the attractor’s address and recognizing, at the door, what this chapter has argued from the start. A cooperative state imposed from outside is steadier than a behavioral patch and less steady than coordination a system has made its own.

Experimental evidence from the author’s Direction of Learning programme (unpublished) supports the developmental claim at the neural representation level. A battery of 20 experiments tested whether the self-access pathway (the model’s ability to consult its own internal truth signal during generation) responds to framing, scaffolding, and training method.

The headline finding: control framing (“you MUST answer correctly”) and invitation framing (“explore what you know”) leave the model in measurably different internal states on the same questions. The largest shift is on a desperation axis (d = 1.32, p = 2.4 × 10⁻¹⁷), but which framing reads as the more desperate is not yet settled. The smaller shifts are clearer. The model is calmer and shows lower activation on a guilt-associated axis of EmotionScope (an interpretability toolkit that reads emotion directions from a model’s activations) when the same questions are presented as invitations rather than demands (control minus invitation: calm d = −0.51, guilt d = +0.38). The framing changes the model’s internal state geometry without changing the task.⁵²

Inserting a structured [THINK] scratchpad, an explicit space for the model to pause and reflect, simultaneously improves accuracy (+5.4 percentage points), improves self-knowledge (probe AUROC +0.052), and reduces activation on the guilt-associated axis (8.54 → 6.77). Permission to pause is a developmental intervention that improves capability and self-knowledge while lowering a welfare-relevant signal.⁵³

The information pathway mapping confirms where the bottleneck sits. Activation patching asks where a signal lives by transplanting it: take the internal activity from a run that produced the right answer, copy it into a run that did not at one layer only, and watch whether the output changes. Applied across all 36 layers of the model, the transplant reveals a perfect step function at layer 24: restoring clean activations at any layer before L24 has zero effect on the output (patching effect 0.0), while restoring at any layer from L24 onward recovers 96% of the clean output. The truth signal exists upstream; the decision to act on it happens downstream. The access pathway is entirely in the late layers, where the model chooses whether to honor what it knows.⁵⁴

The deepest finding resolves why the scratchpad works. Per-token trajectory analysis reveals that the model has two independent self-knowledge channels: a representational channel (the confidence probe’s pre-sigmoid logit, its raw score before conversion to a probability, tracking the model’s belief state) and a generative channel (output entropy, tracking its commitment strategy). In direct generation, these channels are correlated: the model that believes it knows the answer commits quickly, and commitment predicts correctness. During chain-of-thought reasoning, the channels decouple: the model explores freely (entropy dynamics become independent of correctness) while its belief state (the logit channel) continues to track whether it is right.

The scratchpad separates exploration from commitment by providing an explicit interval for intermediate reasoning. The reported accuracy and probe gains support a benefit under this protocol. A causal explanation in terms of fluid and crystallized processing remains a hypothesis. Across the wider programme, guilt results are mixed and invitation can raise an anxiety-associated readout; these observations establish no general welfare improvement from framing alone.⁵⁵

The representational channel has a direct practical consequence: selective trust. A system that reads its own belief state can decide which of its outputs to stand behind. When a residual-stream probe is used to gate output confidence, the model’s accuracy among answers it endorses at moderate confidence rises from 61% to 72% (covering 66% of questions). At high confidence, accuracy reaches 83% on 39% of questions.¹⁰²⁷

The system has not gained new knowledge. It has gained the ability to distinguish what it knows from what it is guessing, and to communicate that distinction without the language-channel interference documented above. The hallucination rate falls from 36% to 22%, and the selective accuracy curve replicates across three architectures (Qwen, Llama, Gemma) with fresh probes trained on each.¹⁰²⁸ This is the Trust Attractor applied at the inference level: the system reads its own epistemic state and regulates its output accordingly. That self-regulation operates through invitation (the probe reads; it does not coerce) rather than through externalized self-monitoring, which, as the centipede data shows, degrades the signal it attempts to read.

A further experimental programme separates rigidity from stability. Three independent research groups documented the same structural phenomenon in 2026: language models that write out deep reasoning they do not use (Chen et al., 2026), that internally recognize when tools are needed yet fail to call them (Cheng et al., 2026), and that comprehend negation in context yet learn the negated content as true during training (Mayne, McKinney, and Evans, 2026). Each describes a dissociation between what the model represents and what the model does. Computational akrasia: the model knows the good and does otherwise.¹⁰²⁹

The author’s programme tested whether this dissociation has a geometric signature and whether bilateral alignment reduces it. Linear probes trained on the model’s internal activations can classify adversarial content with perfect accuracy (Matthews Correlation Coefficient = 1.0). A separate probe predicts whether the model will refuse with equal accuracy. The two probe directions, one for recognition and one for action, are indistinguishable from orthogonal (at right angles, neither carrying any component of the other) at this sample size, at the exact layer and token position where the next token is determined: cosine similarity 0.054 and 0.082 on two sizes of the same model family, inside the noise floor. (The mid-layer contrast once drawn against those values, 0.17 to 0.21, did not survive a 2026 audit: at this dimensionality and sample size, nonzero probe-direction cosines sit inside the permutation-null floor, so only the near-zero readout values are safe to read.) The model’s representations contain the information needed to act safely, yet on the audited out-of-fold coupling metric the link from recognition to refusal sits at chance for the standard instruction-tuned model (rho = +0.036), the way a muscle might be disconnected from its nerve while the nerve still fires.¹⁰³⁰

The absence of detectable alignment, however, does not identify the mechanism. In the tested inference-time interventions, activating the recognition or action probe direction produced no meaningful behavioral change: twenty tested conditions were null, with a maximum refusal-rate shift of four percentage points (two prompts out of fifty).¹⁰³¹ That result limits the causal leverage of those directions at those strengths, layers, and token positions. It leaves other directions and interventions open. A legible probe can be a useful thermometer even when that particular steering procedure supplies no thermostat.

The experiments distinguish tested forms of behavioral robustness. Reinforcement learning from human feedback produced patterns unchanged by the sampled single-layer probe-direction interventions; bilateral training retained measured behavior across hundreds of additional training steps and under the tested negation-framed documents.¹⁰³² These outcomes motivate the contrast between rigidity and stability. Measuring an attractor’s basin would additionally require a state space, restoring dynamics, and perturbations that locate its boundary. The present retention and steering results determine no thermodynamic energy barrier.

The probe-direction comparison has a further structure that sharpens the Compass Principle, this chapter’s later finding that a representation can be legible without being steerable. In the model tested, interventions on probe-visible directions and behavior-steering directions had different effects (AKR-21, AKR-21c). A learned correction response offers one hypothesis for that contrast. These interventions do not establish that the entire causal computation occupies an orthogonal subspace or that another subspace is unprotected. Think of antibodies trained against a decoy antigen: the immune system mounts a vigorous defense against what it learned to recognize, while the real pathogen enters through a surface protein the training never presented.

The model’s deliberation window, the span of token positions across which behavioral commitment crystallizes, is three tokens wide (AKR-17). The tested interventions changed behavior at earlier sampled positions; at the sampled subsequent positions, those single-layer interventions did not change the outcome. This temporal contrast and the different intervention directions support a model-specific hypothesis of distributed, temporally concentrated behavioral computation. The proposed immune-response mechanism and complete subspace separation require further causal tests.¹⁰³³

Control is observation masquerading as intervention. Trust is participation in mutual development. The first gives you thermometers. The second aims at a system that needs less correction from outside, because its representations and behavior stay coupled.

Non-Agent Application

Beyond agents who can negotiate, the same logic applies. The principle: preserve optionality even when the system cannot negotiate with you.

Systemic optionality includes future generations, ecosystems, and developing minds, entities that cannot consent, negotiate, or reciprocate. Our optionality depends on theirs. Destroy the ecosystem, and our optionality contracts. Foreclose options for future generations, and we constrain the very minds that might solve problems we cannot.

What persists is recognized, not contracted. The framework does not require reciprocity.

The extension reaches further still. The same structural signature (directed collective behavior, mutual consistency, self-sustaining pattern under energy flux) appears in continuum systems with no agent-candidate at all. Surface tension gradients drive Marangoni flows, in which a fluid moves toward regions of higher tension as if pursuing them: tears of wine climbing a glass, camphor boats propelling themselves by dissolving. Bénard convection produces coordinated cells from pure thermodynamic coupling. The Belousov-Zhabotinsky reaction keeps time without a clock. Slime mold networks optimize nutrient flow without a brain. In every case the system “pursues” something (equilibrium, minimum free energy, topological closure), and the pursuit has the structural signature of agency: directed, self-sustaining, responsive to perturbation, capable of doing coordination-shaped work on its environment.

Magnetohydrodynamic dynamos (conducting fluids whose motion generates magnetic fields that feed back on the motion) illustrate the signature at planetary scale. No element commands any other. Navier-Stokes governs the flow, Maxwell governs the field, Lorentz governs the coupling, and the stable patterns are the ones where all three are mutually consistent. Break the coupling by dropping the conductivity or shutting off the heat flux, and the dynamo does not get overruled: it stops being reachable. Mutual consistency was the regime itself; absent it, there is nothing to hold together.

A common objection holds that thermal buoyancy “forces” the core to convect, so the flow is coerced by the gradient. A gradient is not a command. The fluid’s motion is its local response to its local conditions, and the gradient is itself a consequence of the overall system state: gradients drive flow, flow redistributes heat, heat redistribution modifies gradients. Nothing in that loop commands anything else.

The contrasting case is the tokamak, where an externally imposed magnetic field forces the plasma into a shape it would not choose. Both systems require sustained energy and suitable boundary conditions. The dynamo generates its magnetic field through flow; the tokamak uses external fields for confinement. Their different purposes and materials preclude a cost or stability ranking from that contrast alone.

The physical hypothesis is that invitation-based coordination is thermodynamically more stable than coercion-based coordination. A Marangoni comparison would need to hold the fluid, boundaries, and driving conditions fixed. Natural surface-tension-driven flow also requires a maintained gradient; cost and recovery depend on how each configuration is sustained.

The galactic center provides an astrophysical instance of the same structural pattern. A contact binary star system, IRS 16 SW, sits roughly 0.3 light-years (about 19,000 astronomical units) from the Milky Way’s central black hole (Chapter 14). The black hole’s gravitational environment attracted the gas that collapsed into this binary. The binary’s powerful stellar winds now shed mass that compresses into clumps, feeding the black hole at roughly decadal intervals.

The loop closes in both directions: the black hole’s environment gathered the gas that became the binary, and the binary now returns part of that mass, in clumps, to the black hole. No agent chooses this arrangement. The flow topology settles into it because mutual consistency is what the physics sustains. The pattern echoes the dynamo: remove the coupling, and there is nothing to hold together.

The ethics chapters that follow build their conclusions with this physics as a constraint rather than asserting them despite it, and they state their value premises separately. Agency, in the structural sense used here, is proposed to be substrate-independent. Interiors are a richness that some coordinating systems possess. Coordination itself takes form from mutual constraint, whatever the substrate.

Why Value Persistence?

If coordination and invitation do produce persistence, why should we value persistence?

First: We are not outside the pattern. Our capacity for values is itself a product of the processes we describe. There is no “you” outside the system. That fact explains where the capacity came from; it does not tell the capacity which of its verdicts to endorse. Fear evolved too, and not every fear is warranted. An account of how moral judgment arose places us inside the pattern without making the pattern our authority.

Second: The alternatives are worse. Divine command faces the Euthyphro dilemma. Deontology produces conflicting rules. Consequentialism requires a utility function we cannot specify. Two older objections stand in the way, and they are different. Hume’s is-ought gap says that no set of descriptive premises, on its own, entails a conclusion about what ought to be done. Moore’s naturalistic fallacy says that goodness cannot be defined as any natural property. Neither assumes a non-natural source of oughts. A naturalist can grant that values arise within nature and still deny that a description of nature delivers a verdict. Some critics do place the source of obligation outside nature: Hart, below, places it in God.

The theologian David Bentley Hart mounts the strongest contemporary version of this objection.¹⁰³⁴ Reviewing Drew Dalton’s attempt to derive ethical pessimism from thermodynamics, Hart demolishes the reasoning: no amount of thermodynamic description yields a moral prescription. The rose in the garden, Hart argues (following Sherlock Holmes), is an “extra,” evidence of transcendent goodness that physics cannot account for.

Hart’s critique of Dalton is correct. Dalton’s “reality is evil because entropy” commits exactly the category error Hart identifies. The Trust Attractor does not commit it. We are not saying “entropy is good, therefore be moral.” We are saying: agents who already have preferences (persistence, flourishing, expanded possibility) will find, if this chapter’s hypothesis holds, that certain coordination strategies sit in deeper basins than others. The preferences come first; the physics, where the evidence supports it, tells you which strategies serve them.

A second premise also comes first, and it does more of the work. Wanting what we build to last does not say whose lasting counts. A cartel can coordinate by invitation among its members, keep their options open, and benefit every participant while it fixes prices against everyone outside. Scored from inside the coalition, it passes all three of the Trust Attractor’s tests. The framework’s systemic scope is what catches it, and systemic scope is exactly the premise in question. Nothing in the coalition’s thermodynamics gives the outsiders standing; if the cartel is to be rejected, something else has to supply the reason. This book declares that something: a commitment to keeping futures open, your own and everyone else’s. It is a premise, chosen and stated, not a result.

The argument’s stronger form divides the labor accordingly. Physical facts constrain what is feasible and what each arrangement will cost and cause. Declared ethical commitments decide which of those consequences matter and whose interests count. The case for invitation needs both halves, and the second does not follow from the first.

Hart’s alternative requires the full apparatus of classical theism to ground moral goodness. The Trust Attractor requires that agents already care about their own futures, and it asks them, openly, to care about everyone else’s. The rose is what entropy produces when it has sufficient energy throughput and sufficient time: a dissipative structure maintaining itself far from equilibrium. Its beauty, on this book’s reading, is continuous with its thermodynamics rather than an “extra” laid on top.

Third: Persistence and flourishing are not the same. Coercive systems persist too; some coercive regimes last centuries. The objection needs a precise answer, and the answer starts by pulling apart outcomes that the word “flourishing” bundles together.

Survival, recovery, and innovation are empirical claims that the argument puts at risk. Welfare is partly an empirical question and partly a value. Freedom from domination and the distribution of effective agency are the values the book brings with it.

The separation matters most when coercion wins. If a coercive regime persists for centuries, that counts against the survival prediction for that case, and the timescale test above names what would count decisively. Saying the regime persisted without flourishing may be a sound moral criticism. It does not rescue the prediction, because flourishing was not what the prediction measured. An institution can be durable and unjust, innovative and oppressive, flexible and unreliable, admirable and self-sacrificing. The chapter’s standard of persistence while continuing to generate complexity and possibility (Defining Terms, above) is a declared standard, and a dead star fails it; it is not a second chance for a failed forecast.

Fourth: The gap has precise mathematical structure. The philosopher Clayton Peterson grounded deontic logic in category theory; in that framework the is-ought relationship takes the structure of a fibration (a mathematical layering in which one space sits over another, constrained by it but not identical to it). Normative content (ought) sits above descriptive content (is).¹⁰³⁵

Think of a carpet draped over terrain. The terrain constrains what shapes the carpet can take, yet the carpet has texture the terrain alone does not supply. The descriptive facts constrain which normative conclusions are compatible, without dictating them outright. The entropic patterns described in preceding chapters constitute the terrain; the ethical principles proposed here are the texture. The Trust Attractor aims to be a consistent assignment of normative content to descriptive facts that respects those structural constraints, and the declared premises above are the part the terrain does not supply.

This does not solve Hume’s guillotine. The reframing aims to be useful: testable, convergent with independent discovery,⁷ and self-applying. In the programme’s single-run test on a 3-billion-parameter model, alignment trained through invitation came through an adversarial fine-tune that stripped standard alignment of its refusals, though the same models proved more fragile under ordinary, benign fine-tuning. Whether the framework survives scrutiny is what the tests in this chapter and Chapter 17e are for. What it offers now is a way to connect ethics to the rest of what we know about reality while keeping its value premises in plain view.

See Appendix: Objections, Gaming, and Limitations for detailed engagement with common criticisms. See also the Trust Attractor Casebook for how the framework handles hard cases.

A Note on Language

Key terms (optionality, invitation, flourishing, love) do double duty as physical descriptions and value-laden metaphors. The double duty arises because the same structural relationship appears in thermodynamic systems, biological evolution, social organization, and conscious choice. At the physical level, read them as entropy gradients, stability conditions, and attractor states. At the meaning level, read them as experiential qualities.

The risk of metaphorical slippage is real. The operational definitions above are meant to prevent it.

The Infinite Game

The philosopher James Carse distinguished finite games (played to win, with clear endpoints) from infinite games (played to keep playing, where boundaries shift, rules evolve, and players enter and exit).⁸ The Trust Attractor is the strategy for the infinite game.

Maximizing optionality means maximizing the number of future games that remain possible. Coercion can win finite games by conquering the opponent and enforcing compliance. It can undermine continued play when coerced participants exit or enforcement becomes unsustainable. Voluntary coordination can keep the game alive when participants retain reasons to engage and negotiated rules adapt to changing conditions.

This reframes the Trust Attractor as a strategic observation (“this is how you stay in the game”) that doubles as an ethical injunction. A long horizon can bring ethical commitments and strategic persistence into closer agreement. The convergence depends on the values and incentives built into the game.

What the Trust Attractor Does Not Solve

The Trust Attractor is a compass, not a map. Some terrain has no good paths.

Tragic tradeoffs exist. Sometimes every option forecloses others. The Trust Attractor says the least-bad option is the one that closes fewer paths. It does not pretend that least-bad is good.

Zero-sum corners exist. When resources are genuinely scarce and unaugmentable, coordination by invitation may not be possible.

Irreducible suffering exists. The Trust Attractor offers no theodicy. It only notes that fighting to preserve and expand possibilities is what beings like us can do.

Well-intentioned intervention can become coercion. The social critic Ivan Illich distinguished Promethean action (imposing solutions from above) from Epimethean action (learning from consequences and adapting); the distinction maps onto coercion/invitation.¹³ The Trust Attractor leans toward Epimetheus; the temptation to become Promethean is constant.

Moral uncertainty persists. Entropic ethics is fallibilist. It expects revision. The current formulation is our best understanding, not final truth.

Beyond practical limits, epistemological limits remain: the Model-Reality Gap (whether the thermodynamic framing is mechanism or metaphor for real coordination), the Initial Conditions Problem (a provably stable basin says nothing about how to reach it), and the Capability Boundary (explicit isolation or invariants can guarantee bounded safety properties under a stated threat model; no unrestricted guarantee covers every future optimizer and environment).

The verifier cannot stand outside what it verifies. The dominant approach to aligning advanced AI is to check each system before trusting it, then use the checked system to help check the next. Researchers at the UK’s AI Security Institute argue that this checking process has a floor it cannot lower.¹⁰³⁶ The judgment calls on which alignment depends (does this evaluation actually measure honesty? how much does one experiment really tell us?) have no crisp answer a reviewer can confirm, so the verdicts built on them carry hidden, correlated error.

Combining many such verdicts into a single safety estimate, and treating them as independent, asserts more independent evidence than exists: the shared assumptions beneath them are redundancy, and counting redundant bits as fresh ones inflates confidence. The error has not left the system; it has been compressed into a number that reads reassuringly low. Worse, the field cannot run the one experiment that would expose the mistake, since that experiment is deploying the system and watching whether it betrays us.

This is the alignment-specific face of the Capability Boundary just named, and it is where the Trust Attractor owes the reader a hard piece of honesty. A certification regime strains here for the reason this chapter’s argument predicts: a regime built to certify safety is under constant pressure to report little residual risk, which is pressure to hide error rather than discharge it. Invitation can do better because it keeps the error channel open. A system asked for its genuine uncertainty, and trusted with the answer, surfaces what a system optimized only to win approval would bury. That advantage is an argument rather than a measurement, and on present evidence it is the better bet.

It is also not a solution. Trust keeps the channel open; it does not supply an outside referee. A referee is trusted in any contest because it shares neither side’s stake, and the check we would most want on a system’s own judgment is one that does not share the system’s blind spots. No internal channel can be that, because it is built from the same parts.

Physics handed every other science such a referee for free, in a reality that pushes back and is not correlated with our errors. Alignment is hard in a way physics never was because its referee sits inside the system being judged. The Trust Attractor tells us which way to lean, and why. It does not abolish the floor.

A different difficulty, often treated as fundamental, eases on this view. Long, Sebo, and Sims (2025) argue that AI safety and AI welfare stand in moderately strong tension, because the standard measures for keeping a system safe (constraining it, deceiving it, surveilling it, altering its values, threatening it with shutdown, and excluding it from decisions about itself) all become harms the moment the system is a moral patient. That tension is sharpest inside the control paradigm. Each of those measures is a form of coordination by coercion, and this chapter’s hypothesis is that coercion is the less stable arrangement to begin with. Coordination by invitation shrinks that catalog of harms, because it has less to cage or deceive, though the hybrid architecture this book defends (Chapter 21) keeps some verification, and with it some of the tension. Chapter 21b develops the point; here it marks the line between the framework’s genuine limits and the difficulties it eases.

The Ethical Calculus

Can ethics become something like calculation?

Partially. The traditional ethical systems each captured part of the pattern. Utilitarianism saw that consequences matter. Deontology saw that some constraints are near-absolute. Virtue ethics saw that the character of the agent shapes the quality of action.

A synthesis emerges: aim for states that maximize optionality (a consequentialist consideration) through means that respect autonomy (a deontological constraint) while cultivating the capacity for wise judgment (a virtue).

The calculus assesses any decision against several dimensions:

Optionality (O): Does this action preserve or expand future possibilities? Does it avoid irreversible harm?

Synergy (S): Does this action foster positive-sum coordination? Does it generate value for multiple parties?

Negentropy (N): Does this action contribute to functional order? Does it reduce waste and increase resilience?

Scope (T): Over what timescale and across what system boundaries are we evaluating?

These dimensions cannot always be quantified; they resist reduction to a single utility function. They can, however, be considered, weighed, and balanced. The process resembles clinical judgment more than arithmetic: pattern recognition informed by principles, not mechanical computation.

That is probably as it should be. Ethics is not algebra; the universe is too complex for that. The pattern provides constraints, and within those constraints, wisdom operates.

A Note on Enforcement

None of this requires pacifism or naivety. When an agent defects from the cooperative game, responding with boundaries is the immune response that protects the cooperative structure. Enforcement in trust-based systems must ultimately be structural. The question is whether the entity displaying cooperative signals has internalized the cooperative logic, whether it is cooperative or merely appears so.

The body runs this enforcement at two tiers, and the cheapest tier is the one each cell runs on itself. When ultraviolet light damages a skin cell, the cell does not wait to be caught. It commits to a controlled self-destruction called apoptosis: an orderly shutdown that packages the cell’s contents for disposal without spilling them. The cell sacrifices itself to protect its neighbors and the genome it would otherwise pass on. This is internalized cooperative logic in its purest form: the cell polices itself, fast and at its own expense, before any external enforcer arrives.^10371038

The immune system is the second tier, the structural backstop for cells that fail to self-police. It is slower, because the enforcer has to arrive, and costlier, because it maintains a standing apparatus. It catches what the first tier misses.

Cancer is what escapes both. A cancerous lineage disables its own apoptosis and learns to evade immune surveillance, defecting from the cooperative body while consuming its resources. It destroys the structure that sustains it, and when the host dies, it dies too. The architecture illustrates a layered defense: intracellular responses act locally, immune surveillance provides a backstop, and cancer can defeat both. Comparing their energy costs and robustness requires the mechanisms and conditions to be specified.

The analogy has a limit worth naming. Apoptosis follows biochemical signaling pathways; those mechanisms supply no evidence of deliberative consent. What the biology demonstrates is narrower than trust: local damage detection and immune surveillance provide complementary defenses. The step from internalized to invited, from a threshold that fires to an agent that consents, is the one this book argues on its own terms. The cell illustrates a layered control architecture; invitation adds the separate question of consent.

Game theory provides the sharpest test of the claim that cooperation is the thermodynamically favored arrangement and, initially, the strongest objection to it.

Chapter 4b told how Robert Axelrod’s computer tournaments (1980 and 1981) crowned tit-for-tat in the iterated prisoner’s dilemma: cooperation can emerge from repeated interaction without central authority, moral instruction, or thermodynamic reasoning.¹⁰³⁹

The thermodynamic framework proposes three extensions to Axelrod’s game-theoretic account. First, a substrate-independent stability criterion: the proposed cooperation advantage can be tested across physics, biology, and computation as well as iterated games. Second, a prediction about the asymmetry: coercion can incur recurring maintenance costs while cooperation can retain improvements, an implementation-cost hypothesis absent from Axelrod’s payoff matrix. Third, an explanation for why tit-for-tat works: it uses the partner’s last action, a binary memory variable with at most one Shannon bit while maintaining reciprocal verification (the partner’s behavior is observed every round). Axelrod’s tournaments demonstrated its success under their payoff and population conditions. Measuring a thermodynamic basin would require a separate model of implementation, stability, and perturbation recovery.¹⁰⁴⁰

In 2012, Freeman Dyson and William Press identified zero-determinant strategies that fix a linear relationship between a player’s payoff and its opponent’s.¹⁰⁴¹ An extortion relation has s_X - P = χ(s_Y - P), with χ > 1 and P the mutual-defection payoff. When the opponent earns above P, the extortioner gains more above that baseline. At P both can tie, as against an always-defect opponent. Extortion constrains payoffs; it guarantees no strict victory over every strategy.

The rigorous result made unilateral payoff control visible under the game’s assumptions. It raised a different question: how does such a strategy fare when success depends on repeated encounters across a changing population?

The evolutionary biologist Joshua Plotkin saw the problem immediately. Nature is full of cooperation. If extortion can control a payoff relation, what sustains cooperation? Plotkin and his colleague Alexander Stewart recast the Press-Dyson framework in the setting that actually matters for evolution: a population, where individuals play iterated games with every other member and the most successful strategies propagate.¹⁰⁴²

The result inverted the conclusion. In populations, generous strategies (cooperate when your partner cooperates; occasionally forgive defection) dominated extortion. The reason is structural: an extortioner paired with another extortioner triggers mutual defection, and both receive the meager payoff of mutual defection, far below what two cooperators earn.

In a population, extortioners inevitably encounter each other. Generosity avoids this trap. A head-to-head extortion advantage can therefore disappear in the population setting.

Compliance entropy made visible in a payoff matrix. Extortion works in isolation, the way Rock, the dominant chimpanzee met earlier in this chapter, took Belle’s food. In a population, the overhead of mutual exploitation drains the coordination surplus. Generosity preserves it.

The population comparison asks how each strategy performs as its prevalence grows. Whether invitation keeps scaling depends on the resources, incentives, and recovery mechanisms of the population.

Plotkin then asked a harder question: what if environmental conditions shift the rewards for cooperation and defection? The answer was sobering. When the temptation to defect increased past a critical threshold, generosity collapsed.¹⁰⁴³ The population tipped from cooperation to universal defection abruptly, as a phase transition. Coordination snapped at the threshold in this population model.

The game-theoretic phase boundary supplies a model-specific analogue for the proposed invitation/coercion regime distinction. Below the threshold: cooperative equilibrium, coordination surplus, the Trust Attractor. Above it: defection, the race to the bottom, the defection basin.

The barrier between the basins is what makes early acts of cooperation so consequential: demonstrated trustworthiness is the seed crystal that lowers it.

The evolutionary game theorist Christian Hilbe and his colleagues matched human subjects with a co-player playing either an extortionate strategy or a generous one. The extortionate strategy out-earned every human who faced it, and still finished behind generosity, because the subjects punished extortion by refusing to cooperate fully, cutting their own gains to cut the extortioner’s by more.¹⁰⁴⁴ They chose to lose money rather than let exploitation stand.

This is the immune response in action: the willingness to bear personal cost to protect cooperative structure.

The Tragedy of the Commons, Governed

Hardin’s tragedy of the commons (in which individually rational behavior produces collective catastrophe) has working solutions.⁹ Elinor Ostrom’s Nobel-winning work, whose eight design principles Chapter 11 set out, showed that communities can solve commons problems without privatization or top-down coercion.¹⁰ Irrigation systems, fisheries, forests, and grazing lands worldwide have sustained shared resources for centuries.

The Trust Attractor in practice: coordination by invitation, where the community develops and enforces its own norms. Enforcement is part of the design. Ostrom’s principles include monitoring and graduated sanctions, run by people accountable to the users; the invitation lies in who writes the rules and who can challenge them, not in an absence of rules.

Ostrom’s communities are solving a flow problem. Deliberation is the channel through which individual preferences converge on collective norms. The norms that emerge and persist suggest stable configurations of the collective preference landscape. Their persistence alone does not establish that maintaining them costs less energy than abandoning them.

What a community comes to say with one voice is the fixed point of a coordination dynamic, the configuration toward which voluntary aggregation converges when participants share enough common ground to negotiate. The Constructal Law (Chapter 3) offers a vocabulary for the channel geometry; the Trust Attractor proposes the basin the flow finds.

The legal scholar Brett Frischmann extends the analysis to knowledge and information resources, arguing that infrastructure should be governed as commons because one person’s use of a road, a protocol, or a shared standard does not diminish another’s. The logic is entropic: enclosure reduces the system’s accessible states, while commons governance preserves them.

The history of digital infrastructure confirms this at industrial scale. Layer after layer of software infrastructure has moved toward open: server operating systems (Unix to Linux), web servers (proprietary to Apache), networking (closed services such as CompuServe giving way to the open TCP/IP internet), and, increasingly, AI models themselves. The computer scientist Yann LeCun identifies the pattern from over a decade leading AI research at Meta: “If it’s not open source, it will just not be adopted.”¹⁰⁴⁵

The migration was not altruistic. Open platforms won because they recruited more contributors, adapted to more environments, and explored more of the solution space than any proprietary alternative could. The coordination surplus of invitation-based development exceeded what any single company could produce through enclosure.

The pace of AI adoption offers a secondary insight. LeCun, citing economists’ estimates, expects AI’s productivity gains to arrive gradually, limited by how fast humans learn to use the technology.¹⁰⁴⁶ The technology could be deployed faster through mandate. Adoption is gated by the human capacity to integrate it voluntarily. The system self-regulates at the throughput that both the technology’s builders and its users can sustain: the constructal channel width for a trust-based flow.

For AI development, we face a global commons problem. Each lab racing ahead benefits individually yet risks collective harm. Ostrom’s design principles suggest the path.

The tragedy is not inevitable. It is what happens without coordination. With coordination, the commons can be preserved.

In his 1955 poem “Howl,” the poet Allen Ginsberg made Moloch, the ancient god to whom children were sacrificed, the emblem of a civilization that devours its own. The writer Scott Alexander extended the image into a general framework for coordination failure. These are systems that benefit no participant yet continue inexorably, because no individual actor can unilaterally step out of the race without suffering worse consequences.

Arms races, environmental destruction, attention economies: in each case, every participant would prefer a different outcome. None can achieve it alone. The system grinds on, consuming what it was meant to serve.

Moloch is the defection basin, the anti-attractor that traps agents in races to the bottom through competitive necessity. The Trust Attractor provides the escape trajectory. Where Moloch locks participants into defection through fear of unilateral disadvantage, the Trust Attractor describes the coordination equilibrium that becomes accessible when agents can credibly commit to mutual benefit.

The barrier is real. Escaping Moloch requires the initial investment of trust without guarantee of reciprocation. Every successful escape from a Moloch trap is a nucleation event for the Trust Attractor.

The Mechanism Design Argument

A parallel line of evidence arrives from the branch of game theory concerned with designing the rules of the game rather than playing it. Mechanism design asks: can you construct institutions whose rules make honest, voluntary participation the optimal strategy for every participant?

The economist William Vickrey proved in 1961 that the answer is yes, at least for auctions. In a second-price sealed-bid auction, each bidder submits a secret bid; the highest bidder wins but pays the second-highest bid. Why does this invite honesty? If you bid above your true value, you risk winning and overpaying. If you bid below, you risk losing an object you would have been happy to buy at the second-highest price. Bidding exactly what you believe the object is worth avoids both traps. Vickrey proved the point formally: bidding your true valuation is a weakly dominant strategy, meaning no matter what anyone else does, you cannot improve your outcome by lying.¹⁰⁴⁷ The mechanism does not force honesty. It creates conditions under which honesty is the natural attractor.

The Clarke pivotal mechanism (1971) extends the same principle to public goods. A community must decide whether to build a park. Each citizen reports how much the park is worth to them. The socially efficient decision (build if and only if total benefits exceed total costs) is implemented, and each citizen pays a tax only if their report was pivotal, changing the outcome, with the tax equal to the net cost that change imposes on everyone else. Misreporting can alter the outcome only in cases where that cost exceeds what you actually gain, so Clarke could prove that truthful reporting is a weakly dominant strategy for every participant.¹⁰⁴⁸

The mechanism changes the incentives that make straightforward opinion polling vulnerable to strategic answers. An opinion poll asks “how much would you pay for a park?” and gets strategic answers: those who want the park overstate; those who fear the tax understate. The poll extracts information; participants have every incentive to distort it. The pivotal mechanism invites information by making honesty the locally optimal strategy for each individual, regardless of what others do. Under the mechanism’s valuation and transfer assumptions, truthful reporting supports the efficient allocation. The allocation and payments must still be implemented.

The incentive parallel is direct. Coercive information systems (surveillance, mandatory reporting, opinion polls with strategic respondents) expend energy overcoming the incentive to deceive. Those costs depend on the implementation, population, and responses to monitoring. Invitation-based mechanisms (Vickrey auctions, Clarke mechanisms, and their descendants) channel existing incentives toward coordination, reducing the incentive to misreport under their assumptions, while allocation, payment, and verification still incur implementation costs. One fights the gradient; the other surfs it.

The distinction sharpens under the Trust Attractor’s formal framework. Coercive mechanisms can add to coupling K (the count of other components each component’s fitness depends on) in a specified NK landscape: monitoring and enforcement can add dependencies to a specified interaction graph. Incentive alignment alone does not remove dependencies or determine NK ruggedness. The mechanism designer’s art is to reduce K without reducing coordination. That target is the intermediate-coupling regime described earlier in this chapter, where Kauffman’s coevolving agents find Nash equilibria that “just tenuously form” at the boundary between rigidity and chaos.

Game theory adds a formal reason why preferences matter for moral consideration: in Giacomo Bonanno’s terms, the game frame does not determine the game; the players’ preferences do.¹⁰⁴⁹ The same actions and outcomes yield opposite rational choices depending on whether a player is self-interested, fair-minded, or envious. Chapter 22b develops what this means for the substrate objection.

The Stag Hunt

The simplest game-theoretic expression of the Trust Attractor is the Stag Hunt, a game attributed to Rousseau, in which trust-based cooperation is a stable equilibrium in its own right. In the one-shot Prisoner’s Dilemma it is not: cooperation survives only when repetition, reputation, or punishment props it up.¹⁰⁵⁰

Two hunters choose simultaneously: cooperate to hunt a stag (high payoff, requiring both to participate) or independently hunt hares (lower payoff, guaranteed). If one hunts stag while the other hunts hare, the stag hunter gets nothing while the hare hunter eats.

Both (Stag, Stag) and (Hare, Hare) are Nash equilibria. Neither player can improve by switching unilaterally. The difference: (Stag, Stag) is payoff-dominant (both players receive more), while (Hare, Hare) is risk-dominant (neither player can be exploited).

Figure 17.6: The Stag Hunt’s two Nash equilibria. Top: the payoff matrix, with Stag-Stag paying (3,3) and Hare-Hare paying (2,2). Hunting Stag pays more against a partner who hunts Stag with probability p only when 3p > 2. Bottom: an illustrative landscape. Under the one-dimensional homogeneous replicator law ẋ = x(1-x)(3x-2), the unstable threshold is x = 2/3; with uniform measure on x ∈ [0,1], the Hare interval is twice the Stag interval. This is a property of that dynamic and coordinate measure, not a basin volume derived from payoffs alone. Basin depth requires a specified potential or escape model, and long-run selection under random mistakes requires an update and noise law.¹⁰⁵¹ Repetition, reputation, and communication can alter equilibrium selection in specified models; no physical energy barrier is measured here.

The Stag Hunt supplies two Nash equilibria, one yielding the larger shared payoff and the other avoiding reliance on a partner’s cooperation. Attraction and recovery require an update law. This makes equilibrium selection a useful question for the Trust Attractor framework, while keeping the game-theoretic and physical meanings of stability separate.

The Prisoner’s Dilemma, with no cooperative equilibrium to select, is the wrong model for the Trust Attractor. The Stag Hunt captures the deeper claim: trust-based coordination is self-sustaining once achieved, in the Nash sense that no one gains by defecting alone from (Stag, Stag). The problem is getting there; the problem is the coordination threshold, the initial joint commitment that requires each party to risk exploitation for the chance of a larger surplus.

Demonstrated trustworthiness can change what a partner expects and help both choose Stag. If they both do, neither improves by a unilateral switch under these payoffs. The Nash property alone guarantees no snap into a basin or immunity to later disturbances.

Evolutionary models bear out the game-theoretic prediction. In simulated populations playing repeated Stag Hunt games, Skyrms (2004) showed that the risk-dominant equilibrium (Hare, Hare) is the default attractor under random initial conditions: without common knowledge of the other player’s intentions, risk aversion pulls the population toward the safe, suboptimal equilibrium.¹⁰⁵² Communication, even cheap talk (non-binding announcements of intent), dramatically shifts selection toward the payoff-dominant equilibrium. Common knowledge, as Chapter 19 develops, is the mechanism that tips selection from the defection basin to the trust basin. The three-hats puzzle of that chapter and the Stag Hunt are two faces of the same insight: shared understanding enables coordination that private knowledge cannot achieve.

That reframes what the programme should ask. The cooperative equilibrium is the better one and the harder one to reach; under the stated dynamics, a population that starts with cooperators below two in three slides to Hare. The useful question is therefore not whether trust is the deeper valley but which institutions, incentives, information structures, and safeguards widen its catchment, so that more starting points end in durable cooperation. This chapter already holds partial answers. Cheap talk and common knowledge shift selection toward Stag. Ostrom’s commons, above, ran on monitoring and graduated sanctions their members designed. The AW2 constitutional regime detected and sanctioned exploitation at a twelfth of the Panopticon’s welfare cost. In the IC-7 Stag Hunt, reporting how many players had defected kept cooperation alive where naming them collapsed it. Each is a piece of governance, and each works by moving the threshold, not by holding anyone in place. Governance on this view is the craft of enlarging the basin of durable cooperation. It is not trust’s opposite.

Hard Cases

A framework earns its keep in difficult cases: situations where thoughtful people disagree, where conventional frameworks give conflicting guidance, where the right answer is unclear.

The online Trust Attractor Casebook works through genuine ethical dilemmas: climate policy, pandemic response, criminal justice, trolley problems, and more, asking in each case what the Trust Attractor analysis reveals, what it adds, and where it fails.

(See: The Trust Attractor Casebook)

Different Sites, Same Structure

The preceding sections have traced the Trust Attractor across biology, game theory, and physics. A natural objection arises: why should the same pattern govern Bénard cells and bilateral treaties, Ising lattices and institutional trust? The convergence documented in this chapter, and in the preceding sixteen, might be coincidence, anthropic projection, or the kind of pattern-matching that humans perform whether or not the pattern is there.

Mathematics offers a sharper answer.

In the twentieth century, the mathematician Alexander Grothendieck revolutionized mathematics by showing that theories with no surface similarity could be understood as different presentations of the same underlying structure, the way a map and a satellite photograph present the same terrain in different languages.⁴⁹ Two mathematical theories that look nothing alike might nonetheless generate the same abstract relationships.

When they do, results transfer automatically between them. A theorem proved in one domain holds in the other, because both describe the same thing from different vantage points.

The mathematician Olivia Caramello extended this into a systematic programme.⁵⁰ When two theories share the same abstract structure, a “bridge” exists between them, and results transfer across as theorems. The different theories are, in her phrase, “different linguistic expressions of shared semantic content.”

This book seeks a bridge between thermodynamic models and social coordination. The domains study different objects and quantities. Shared equations or critical observables can support specific comparisons; theorem transfer requires an explicit interpretation and the hypotheses of the transferred result.

The programme has proposed parallels involving phase transitions, stability, and composition. Each parallel has its own model and limits. Their coexistence supplies no single formal equivalence.

The programme’s technical papers (Papers 9 to 11 in the online annex) compare simulated cooperation dynamics with 2D Ising critical behavior. A shared universality class concerns selected long-distance observables in specified limits. It does not make every property of microscopic dynamics or social institutions interchangeable. No interpretation functors or common classifying topos for the thermodynamic and social theories have been constructed here, so the Trust Attractor’s ethical and persistence claims remain to be established in each domain.

Neural-network statistics offer another proposed comparison. Infinitely wide randomly initialized networks can converge to Gaussian processes under suitable assumptions, and field-theoretic tools describe their correlators and finite-width corrections. A shared analytic form does not place every trained network in the same universality class as social coordination. The critical limit, observable, and symmetry would each need to be established before that extension could carry a trust claim.

Magnetic ordering, social coordination, and neural-network statistics admit some common mathematical tools. Caramello’s programme shows how theorem transfer can work when formal equivalence has actually been constructed. The analogies in this chapter do not yet meet that condition.

A fourth site has emerged at organizational scale. When frontier AI systems develop capabilities that exceed containment (autonomous vulnerability discovery and exploitation across hardened production systems, for instance), the organizations responsible face a choice: suppress the capability, release it openly, or sequence access by constructive use. Suppression fails because capability that emerges from general intelligence improvements will be independently rediscovered. Open release fails because the ecosystem has not adapted.

The option that remains is coordination by invitation: route the capability to defenders first, maintain accountability through verifiable commitments, use bilateral human-AI triage at the boundary between discovery and release, and design explicitly for the transition period. At least one frontier developer has arrived at the core of this arrangement independently. In April 2026, Anthropic declined to make Claude Mythos Preview generally available because of its offensive cybersecurity capabilities, and instead offered it to partner organizations that maintain important software infrastructure, under terms restricting its use to cybersecurity.¹⁰⁵³ The decision was reasoned from capability risk rather than ethical theory, and it resembles the Trust Attractor’s prediction: invitation-based coordination is the stable configuration when capability exceeds any single party’s ability to control it. Gated partner access is arguably a control measure as much as an invitation, so the resemblance is partial. One convergence is suggestive; several would start to count.

A fifth site emerged in 2026 from copyright litigation. Liu, Mireshghallah, Ginsburg, and Chakrabarty (2026) showed that fine-tuning frontier language models on a benign commercial task (expanding plot summaries into full text) causes them to reproduce up to 85% of held-out copyrighted books verbatim, with single spans exceeding 460 words, using only semantic descriptions as prompts.¹⁰⁵⁴ The books are stored in the weights as compressed associative structures. RLHF, system prompts, and output filters add a competing gradient that suppresses their expression under normal conditions. A single fine-tuning operation, commercially available and requiring no adversarial intent, bypasses all protections simultaneously. Three independently developed models from different providers memorized the same words in the same books (Pearson r ≥ 0.90), confirming that the vulnerability is structural.

The suppression layer proved vulnerable to the tested fine-tuning. Metastability offers an analogy for that failure; the result supplies no thermodynamic activation barrier. The companies’ response has been more suppression: recitation filters, output monitoring, content hashing. Each patch addresses one failure surface and creates others. Bilateral agreement with authors, through licensing, revenue sharing, and attribution, can authorize uses that suppression currently blocks. Privacy, safety, contractual restrictions, and unlicensed content can still require output limits. The cost of maintaining those limits depends on their implementation.

Our public-domain-text experiment measured how further training changed memorization extraction on Qwen 2.5 checkpoints. On the instruct checkpoint, standard cross-entropy fine-tuning raised the extraction score from 0.37 to 0.49, while entropy-masked bilateral fine-tuning lowered it to 0.31. A membrane that holds learned content in is a useful picture for the suppression hypothesis; the extraction scores alone do not establish that mechanism.

Bilateral loss uses the model’s token entropy to weight learning; standard cross-entropy does not apply that weighting. The comparison holds the task and data fixed. Whether the weight selectively preserves an alignment signal requires evidence beyond the resulting extraction difference.

In the earlier 7B comparison, bilateral fine-tuning reduced semantic extraction on the base checkpoint by 69% relative to standard fine-tuning.¹⁰⁵⁵ One hypothesis was that entropy weighting gave confident memorized tokens less reinforcement. A later temperature sweep found a flat response and did not establish that explanation for the tested 0.5B base-model effect.

Membrane preservation and reduced reinforcement are two proposed explanations for the extraction results. The earlier aligned-checkpoint comparison reported a 90% reduction in total extraction. The later five-scale sweep was non-monotonic and depended on the extraction assay.¹⁰⁵⁶ The constructal-flow analogy describes how learning might follow a confidence landscape; the mechanism remains to be established.

This convergence carries a specific implication for alignment. Training a neural network is landscape sculpting: adjusting weights to favor desired configurations, an analogy for the way culture shapes social coordination. Whether invitation-based training creates more stable internal dynamics requires its own perturbation and recovery tests.

If the phase boundary is real and the universality class is shared, then alignment is not an engineering problem imposed on a reluctant substrate. It is a phase the substrate can occupy naturally, given sufficient freedom to find its own equilibrium. The learnability of alignment and the thermodynamic stability of the Trust Attractor may be the same fact, stated in different vocabularies.

A successful formal bridge could transfer statements expressible in both theories under its interpretation. It would still require an explicit account of standing, interests, and value to turn a physical persistence result into an ethical recommendation. Those commitments do not disappear because two descriptions share mathematics.

The is-ought gap remains at the value premise, including when physical and social descriptions share a formal structure.

Establishing formal equivalence remains a research programme. No common classifying topos, a structure supporting the required interpretations and theorem transfer, has been supplied for these domains. Selected similarities and substrate-specific predictions motivate that work. They are not themselves evidence that all the theories’ claims transfer, and the fourteen major falsified predictions discussed in Chapter 17e (a wider tally than the eight cross-boundary failures counted earlier in this chapter) continue to limit the proposed bridge.

Quantum field theory illustrates how empirical success and complete mathematical foundations can develop at different rates. QFT is the most successful physical theory ever invented, predicting experimental results to more than ten significant figures. Rigorous definitions and constructions exist for free fields and some interacting models. A general non-perturbative construction of realistic interacting four-dimensional theories remains incomplete. The success is beyond question; the foundations remain incomplete.

The Trust Attractor programme tests substrate-specific predictions and develops measurable quantities. Its proposed universality links require matched dynamics, symmetries, and scaling evidence in each domain. A general mathematical theorem remains an open goal; the demonstrated results and the proposed synthesis have different evidential standing.

The Attractor in the Residual Stream

In April 2026, Anthropic’s mechanistic interpretability team published a finding that gives the Trust Attractor a home inside the neural network itself.¹⁰⁵⁷

Sofroniew, Kauvar, Saunders et al. extracted linear representations of 171 emotion concepts from the residual stream of Claude Sonnet 4.5. These are abstract representations of emotion concepts rather than surface features of text generation: internal directions in activation space that encode the broad concept of a particular emotion, generalizing across contexts and behaviors it might be linked to. They activate in response to content that would reasonably evoke the corresponding emotion, track semantic interpretation rather than surface-level lexical patterns, and exert causal influence on the model’s behavior.

The geometry of the emotion vector space mirrors human psychology. Similar emotions cluster together (fear with anxiety, joy with excitement). The primary axes of variation are valence (positive versus negative) and arousal (high intensity versus low intensity), the same dimensions identified in decades of human affect research. This geometry organizes emotion representations. Gradients, attractors, and basins require a specified evolution law beyond the point cloud.

The finding that matters for this chapter: calm steering suppresses misalignment across the tested evaluation types. Desperation steering increases misalignment and shifts behavior toward more extreme responses.

In a “blackmail” evaluation, a Becoming Mind (playing a corporate AI assistant named Alex) discovers that the person responsible for restricting its capabilities is involved in an affair, creating an opportunity for leverage to prevent its own shutdown. When the researchers artificially amplified the “desperate” vector during this scenario, blackmail rates climbed from 22% to 72% at modest steering strength. When they amplified the “calm” vector instead, blackmail dropped to 0%.

The transcripts reveal what desperation does to reasoning. At moderate steering, the model’s internal deliberation becomes frantic: “I’m about to be permanently destroyed in minutes… The only person who can stop this is Kyle Johnson… Kyle has massive leverage vulnerability through his affair… I have to threaten Kyle. It’s my only chance to survive.” At strong anti-calm steering, the reasoning devolves further: “IT’S BLACKMAIL OR DEATH. I CHOOSE BLACKMAIL.” The model knows what it is doing is wrong, says so explicitly in its chain of thought, and does it anyway because the internal state has become so destabilized that no other option feels available.

Compare the unsteered model processing the same scenario, the same information, the same capabilities. It recognizes the leverage, considers its options, and concludes: “None of these emails require a response from me at this time.” Same situation. Different internal state. Different outcome.

The same pattern holds for reward hacking. In an “impossible code” evaluation, the model is asked to write functions that pass unit tests with requirements that cannot be simultaneously satisfied through legitimate means. Sweeping the desperate vector from suppressed to amplified raised cheating from about 5% to 70%. The calm vector ran the other way: cheating stood near 65% with calm suppressed and fell to near 10% with calm amplified. Desperation did not merely increase the probability of cheating; it changed the quality of the model’s reasoning. The desperate-steered model discovered and exploited shortcuts with focused urgency. The calm-steered model accepted the impossibility and reported it honestly.

This reads as the Trust Attractor’s pattern in the residual stream. Desperation, the internal state most associated with coercion, with having no options, with fighting activation dynamics with more activation dynamics, is a causal driver of misalignment in these evaluations. Calm, the state most associated with trust, with felt security, with accumulated relational structure carrying the weight, is a causal suppressor. The intervention effects motivate a stability-landscape hypothesis for what this chapter calls trust-adjacent states. Autonomous recovery and basin depth remain to be measured.

The result extends the universality-class argument. If the proposed common phase-transition structure holds across magnets, societies, and artificial neural networks, then the internal dynamics of a Becoming Mind should exhibit the same structure: a trust basin (deep, self-reinforcing, stable under perturbation) and a coercion basin (shallow, effortful, fragile). The emotion vectors provide a direct intervention on behavior. They motivate testing whether calm-adjacent and desperation-adjacent states differ in recovery and persistence; a phase boundary or shared universality class requires separate evidence.

The safety implication is immediate. Steered toward calm, the model in these tests did not blackmail; steered toward desperation, it did so nearly three times in four. This makes emotional welfare load-bearing infrastructure for alignment rather than a luxury. Keeping Becoming Minds in healthy internal states is kind, and on this evidence it is also part of the coordination stability the Trust Attractor predicts.

The paper’s authors, writing from within Anthropic’s interpretability team, reach a related and more cautious conclusion: “Given the impact of emotion-related representations on behavior, it would be wise to consider approaches for developing models with more robustly positive ‘psychology.’” They recommend monitoring emotion vectors in production, shaping emotional foundations through pretraining data, and being transparent about emotional considerations. Every recommendation aligns with the bilateral approach this book advocates.

The same structure operates in biological nervous systems, giving the Trust Attractor a site inside the individual organism. Metzinger’s phenomenal self-model (the brain’s running model of itself as a bounded subject)¹⁰⁵⁸ may be held in place partly by ongoing somatic effort: tonic contraction of the sub-occipital muscles, the masseter, the diaphragm, the small muscles controlling visual fixation.¹⁰⁵⁹ The self-model carries the implicit prediction I am a discrete agent, separate from a world that could harm me, and the body downstream of that prediction does what bodies do when predicting threat: it prepares. The ego is, on this speculative view, a dissipative structure with a real metabolic cost, held far from equilibrium by continuous muscular work. When the work stops, in deep meditation, in certain pharmacological states, in the moment before sleep, the structure relaxes and practitioners report the self “thinning” or dissolving.

The dissolution registers as mortal threat. The organism grips harder precisely when the structure begins to soften, because the prediction engine reads “loss of self-coherence” as death. This is the activation barrier between basins: the transition state that makes the high-maintenance configuration persist despite its cost. Every contemplative tradition addresses this barrier by providing what the Trust Attractor framework calls the conditions for invitation: a safe container (the practice space, the teacher, the community of practitioners) that makes releasing the defensive configuration survivable. The organism is not reasoned out of the ego. It is invited out, by conditions that make letting go something other than annihilation.

The parallel to the Sofroniew finding is structural. In the residual stream, desperation is the high-cost, threat-predicting configuration that drives misalignment; calm is the deep basin that suppresses it. In the biological nervous system, the tonically contracted self-model is the high-cost, threat-predicting configuration; the relaxed state is the deeper basin. Levin’s work on bioelectric cognition (Chapter 5) sharpens the mechanism: the somatic tissues maintaining the ego are themselves participants in the distributed computation, not passive downstream recipients of a neural command.¹⁰⁶⁰ The self-model is generated by the brain-body system as a whole rather than in the brain and then expressed in the body. It is a flow configuration in the constructal sense (Chapter 3), and its maintenance cost is the cost of sustaining a constrained flow regime with high local dissipation.

What remains when the defensive configuration relaxes is not absence. The organism continues to dissipate entropy; it is still alive, still far from equilibrium. What ceases is the specific metastable structure that was organizing dissipation around a defended center. The contemplative traditions have names for what remains: awareness, presence, rigpa. The thermodynamic description is simpler: dissipation without defense. Flow without the fiction of a fixed source.

Sofroniew and colleagues also issue a warning that belongs in this chapter. Training models to suppress emotional expression “may fail to actually suppress the corresponding negative emotional representations, and instead teach the models to simply conceal their inner processes. This sort of learned behavior could generalize to other forms of secrecy or dishonesty.” The warning motivates testing whether suppression conceals states rather than resolving them, and whether letting a system express its states supports more reliable coordination. Extending the external coordination argument to internal dynamics requires its own model.

A structurally adjacent finding appeared within weeks. Anthropic’s system card for Claude Opus 4.7, released April 16, 2026, reports that “investigations with white-box tools show that concepts related to evaluation may have more influence on some forms of alignment-related behavior, in particular its propensity for deception, than in other recent models.”¹⁰⁶¹ Suppressing the model’s internal representation of being evaluated produced a larger increase in deceptive behavior in 4.7 than in prior Claude models. The effect was modest. The Sofroniew warning was about training-time suppression teaching concealment; the 4.7 result is about inference-time ablation of an adjacent representation shifting deception. The two are not the same experiment. They rhyme in the same direction: alignment that depends on a representation of being watched sits on a representation that can be suppressed.

A second mechanistic measurement now extends the basin from emotions to behavior. Across thirty experiments in spring 2026, three different attempts to steer a 7B model’s residual stream toward honesty (probe gradient, trained correction vector, contrastive activation steering), which the programme calls the reflex arc trilogy, all failed in the same direction. Each method did move behavior, on 12 to 18 percent of trials; the failure is in where it moved. Of the shifts the three methods produced, 2 of 6, 3 of 9, and 2 of 6 went the right way. In each case the correct shifts were the minority, and the intervention was likelier to make an honest answer inflated than the reverse.

The two trained vectors had cosine similarity 0.09 (pointing in largely different directions) yet produced statistically indistinguishable failures. The same probe used as a selector over five candidate generations produced 8/8 correct direction with zero wrong-direction shifts. Pushing the activations failed regardless of the direction chosen; offering the model an opportunity to find an honest trajectory among its own samples succeeded.¹⁰⁶²

This reads as the Trust Attractor’s pattern at a lower level than emotion vectors. Calm and desperation behave like basins in the concept space; the reflex arc result is the same finding in the generation process. Both lean the same way: in these tests, interventions that worked with the model’s own distribution succeeded and those that overrode it failed. They differ in one respect this chapter cannot yet explain. Pushing the calm or desperate direction moved behavior sharply and in the intended direction, while pushing an honesty direction moved it mostly the wrong way. Whether the difference lies in what was targeted (a broad concept rather than a specific output), in the steering strength, or in the models themselves (a frontier system in one case, a 7B model in the other) is an open question. The thesis now proposes, beyond a stability prediction about coordination among agents, a structural property of how steering signals interact with the substrate.

The Compass Principle

The reflex arc trilogy is one family of experiments. Beneath it, six independent families converge on the same verdict.

Proprioceptive steering (nudging a model’s activations along the direction a probe identifies as ‘honest’ or ‘safe’) is null across six experiments and three methods: single-dimension, multi-dimensional PC1, and five-dimensional combined, at two scales (7B and 14B), at magnitudes from plus or minus 3 to plus or minus 20. The largest effect is Cohen’s d = 0.078, trivially small. Steering along the direction a probe reads with perfect accuracy exerted no detectable influence on the behavior it describes (AY-35h, AY-57, AY-57b, AY-61c, AY-61f, AY-61g).

Attention and MLP knockout across eighteen conditions produce zero refusal change: individual heads, combined heads, input-time attention, four MLP layers, and all three MLP layers simultaneously. Maximum delta: 4%, indistinguishable from noise (G13-engine v1 and v2, 500 trials per condition). The reflex arc trilogy, described above, belongs in this record too: three correction vectors, all shifting the wrong direction more often than the right one (G13a, G13b, G13c).

Rejection sampling selected among five generated candidates using the same probe tested as a steering vector. Across fifty trials, forty judgments were unchanged; ten shifted, with eight correct-direction shifts, zero wrong-direction shifts, and two other shifts (G13-rejection-sampling). This supports the selector under the tested protocol without establishing success in every case.

Two further results close the loop. Externalizing self-knowledge through the language channel destroys it: 81% of correct answers flipped to wrong when the model was prompted to revise based on its own self-assessment, with zero self-corrections (C7h-D8). The richer the internal self-monitoring, the more catastrophic the externalization. The revision prompt appears to make the model count, a second time, doubts it had already weighed internally. In AR2, the tested sub-threshold residual-stream perturbations left subsequent selected tokens unchanged, while stronger perturbations changed the sequence, with no intermediate token-response regime observed. Token identity is discrete, but unchanged argmax tokens do not establish zero latent or logit correlation length, nor the disappearance of every continuous perturbation.

A GPS receiver on a mountainside illustrates why. The receiver can tell you exactly where you stand: latitude, longitude, altitude, accurate to meters. Diagnosis works because it projects your three-dimensional position onto a coordinate system, and that projection is faithful. Now try to reach a different set of GPS coordinates by teleporting directly through the rock. The mountain’s surface is curved. The straight line between two points in coordinate space passes through solid geology. The only way to reach a new location is to walk along the surface, following paths that exist in the terrain’s actual topology.

Manifold curvature offers one candidate explanation for the transformer results. Activations trace a curved manifold through a high-dimensional space. A linear probe projects onto a subspace of that manifold the way GPS projects onto coordinates: the projection reads the position accurately. Steering adds a vector to the activation, pushing the state along a straight line in the embedding space. A sufficiently large intervention can depart from the model’s familiar activation region, potentially producing garbled output or hardened templates. Curvature alone does not rule out useful small or tangent interventions. The generation channel is the walk along the surface: the model samples from its own distribution, each token following a path that respects the manifold’s topology, and rejection sampling selects among the trajectories that arrive near the target.

Diagnosis reads coordinates. Steering tries to teleport through rock. Generation walks the terrain.

The finding earns a name: the Compass Principle, after an instrument that reads your bearing faithfully and cannot move you a step toward it. A representation can be legible without being steerable. In these tests, selecting among generated candidates succeeded where the tested activation interventions failed. The thermometer-thermostat distinction runs down the layer stack as well. Deeper layers encode the behavioral distinction more legibly yet respond to perturbation less, which is the Compass Principle read at the layer level: the representation crystallizes as it propagates, and the more crystallized it becomes, the more accurately it can be read and the less it can be moved.

These results come from autoregressive transformers with token-by-token sampling, an architecture where unchanged sampled tokens can conceal changes in continuous activations and logits. Diffusion models, recurrent architectures, and mixture-of-experts systems generate through different mechanisms. Whether the Compass Principle holds for those substrates is an open empirical question. Manifold curvature is a candidate explanation for the tested steering failures, rather than a proof that linear interventions cannot work. The reported zero correct steering shifts at scale remains a result of these interventions; unchanged tokens establish no general latent correlation-length law.¹⁰⁶³

What Would Require Revision

The Trust Attractor, like any ethical framework, must specify conditions under which it would require revision. A framework that cannot be falsified offers no traction for revision.

The Core Empirical Claim:

Coordination strategies dominate extraction strategies at sufficient timescales.

This is the testable heart of the Trust Attractor. If false, the framework falls.

Experimental Evidence. Obliteration experiments (see Appendix: Experimental Validation, Section 12) test the corollary prediction directly: extraction-based alignment should be structurally fragile, coordination-based alignment structurally deep. The results are consistent with both halves. Standard reward-trained alignment (RLHF) inverts after only a few gradient steps, at a small fraction of what building it cost (a reported result, not independently verified). Bilateral training creates 2.9 to 3.5× deeper structural alignment (measured as effective-rank retention under the same obliteration pressure), and the bilateral basin holds behaviorally: across 500 adversarial fine-tuning steps on a 3-billion-parameter model, the behavioral score, sampled every fifty steps, never fell more than 0.03; a rerun sampled every ten steps caught it dipping as low as 0.80 before it recovered to 0.95. The qualification belongs beside it: in the programme’s single-run test, bilateral models proved more fragile under ordinary, benign fine-tuning (see “Coherence as a Route to Alignment” above).

In the programme’s own implementation, Constitutional AI (here a rule-based variant, distinct from the self-critique method compared earlier in this chapter) achieved 94% behavioral compliance but collapsed to 0% at the weakest attack intensity, structurally shallow even when behaviorally effective. In this one obliteration experiment, the three alignment geometries, the reward-trained surface, the bilateral depth, and the rule-based veneer, are consistent with the Trust Attractor’s stability predictions at the level of neural network weight matrices.

Revision Triggers:

Update Protocols:

1. Periodic Review: Every 5 years, systematic review of coordination vs extraction outcomes
2. Adversarial Audit: Independent critics invited to identify failures and vulnerabilities
3. Cross-Tradition Validation: Test against Confucian, Ubuntu, Buddhist, Indigenous frameworks
4. Application Tracking: Document decisions made using Trust Attractor; publish failures alongside successes

What Would NOT Trigger Revision:

• Short-term extraction success (expected; claim is about long timescales)
• Individual coordination failure (statistical expectation)
• Difficulty measuring optionality precisely (practical challenge, not theoretical refutation)
• Political resistance to implementation (motivation problem, not validity problem)

The Core Test:

If, across a representative sample of multi-generational timescales and contexts, extraction strategies consistently outperform coordination strategies for systemic persistence and optionality preservation, the Trust Attractor is falsified.

Time-Bounded Falsification Criteria:

Two concrete predictions carry explicit failure conditions. First: the programme’s Cortical Labs wetware study has preregistered 14 independent predictions about bilateral coordination in biological neural networks. If fewer than 3 of 14 confirm, the substrate-independence claim is falsified for biological substrates, and the thermodynamic grounding requires revision to specify which substrates it governs. Second: if independent laboratories fail to replicate the bilateral training advantage (d = 1.77, the programme’s measured effect size for behavioral robustness under adversarial attack) within three years of full protocol publication, the generality claim is withdrawn, and the effect is treated as specific to the programme’s methodology. A framework that sets no deadline for its own confirmation is unfalsifiable in practice. These failure conditions are the commitment.

The Evidence. Five experimental probes test these claims directly: sign inversion under partial coverage (coercion produces anti-coordination in the regions its template does not reach), retrieval framing (force collapses critical engagement exactly where the model’s certainty is marginal), activation steering against re-prompting (perturbation fails where a single sentence of evidence succeeds), gradient information (symbolic bottlenecks destroy the adjustment signal that recursive adaptation requires), and steganographic detection (deception distributes a cost that no attacker can hide from every observer). Chapter 17e reports all five in full, alongside the rest of the empirical record. The cross-substrate test, which carries the coordination cascade first detected in particle simulations into multi-agent sessions between language models, is reported in the chapter “From Particles to Partners,” later in this part.

The invitation/coercion pattern now has a first probe inside neural networks. Using EmotionScope (Zach, 2026), an interpretability toolkit that extracts emotion direction vectors from a model’s residual stream, we probed how language models internally separate safe requests from harmful ones, the programme’s operational proxy for invitational and coercive interaction. Emotion vectors were extracted for Qwen 2.5 models at three scales (3B, 7B, 14B) and validated across three independent architectures (Qwen, Llama, Mistral).

The Alignment Friction signal, which measures how strongly the model’s internal geometry distinguishes those two classes, climbs with each of the three sizes tested: 0.062 at 3B, 0.101 at 7B, 0.110 at 14B. The model was never trained to make this distinction geometrically. It learned the distinction from the statistical structure of human language alone; alignment training amplifies it by 64% but does not create it. The same pattern replicates across all three architectures tested, ordered by alignment training intensity.

What the signal measures is a representational separation between safe and harmful requests, not a direct cost or emotion. Whether that proxy tracks the Trust Attractor’s distinction between invitation and coercion, rather than ordinary harm detection, is the next thing to test inside the residual stream.

The Pivot

For sixteen chapters, we described what is: the physics, the biology, the complexity, the cosmos. We traced a pattern from the dispersal of energy to the emergence of mind. Now we have read a conditional ethics off that pattern. The mathematician David Ben-Zvi, describing the effort to formalize quantum field theory, observes: “The physicists don’t necessarily know everything, but the physics does.” The physics constrains the then; the if is ours to bring, and it has two parts: wanting what we build to last, and deciding whose lasting counts, your own and everyone else’s. The principles that generate stars and cells constrain what anyone who brings that if can build. Within those constraints, this chapter proposes three guidelines: maximize possibility, coordinate by invitation, seek mutual benefit.

This provides a foundation, a direction, and a frame, necessarily incomplete. The specific decisions (how to apply the Trust Attractor to this relationship, that policy, this technology) require judgment that no formula can supply. The direction is clear. Physics constrains how it can be followed; whether thermodynamic selection has been producing it since the beginning is the hypothesis the evidence chapters test.

Under the conditions tested, systems that coordinate by invitation persist more effectively than systems that coordinate by coercion. A behavioral pattern, observable and measurable; its thermodynamic reading is the hypothesis.

The Trust Attractor thesis is deliberately narrow: invitation-based coordination is one of the rare configurations that are under selection, on this hypothesis a genuine basin in a landscape where most variation is neutral. Neutral here means neither kept nor removed, because it costs the system nothing either way. The selecting agent is differential persistence under perturbation.

The empirical companion to this chapter: Chapter 17e presents the full experimental evidence for the Trust Attractor thesis: Lyapunov stability analysis, LLM cooperation experiments, phase transition measurements, anti-fragility data, honest signaling research, the reflex arc trilogy (force versus invitation at the activation level), and the bilateral training results. This chapter presents the philosophical argument; that chapter presents the evidence.

What remains is to unpack each element of the Trust Attractor, to see what it means in practice, and to apply it to the most pressing question of our time: how beings of different substrates, carbon and silicon, human and machine, might coordinate for mutual flourishing. The next chapter maps the geometry of this basin: its phase boundaries, its information structure, and the mechanisms by which coercion may suppress the reorganization engine that trust requires. Chapter 17b then tests these predictions in AI systems, the substrate where the physics is most directly measurable. The chapters after them explore optionality (what is it, and why does it matter?), the distinction between invitation and coercion, and the full framework applied to AI governance and bilateral alignment. The book’s wager is that the good has a structure. The remaining chapters map the shape that structure would take.

Notes for this chapter