Chapter 17aThe Geometry of Trust

The simulations distinguish coordination regimes through their update rules, fluctuations, and response to perturbation. Their geometry can help frame questions about recovery and fragility. Establishing how those measures map to trust and coercion in institutions requires a separate account of incentives, information, and dynamics.

Causal Entropic Forces

In 2013, Wissner-Gross and Freer published a study of what happens when you optimize a system for future freedom of action.³ The result looks like intelligence. In their simulations, a large disk spontaneously used a smaller disk to extract a trapped disk from a narrow tube: tool use arising from entropy maximization alone.

Two disks in separate compartments synchronized their movements to pull a larger disk toward them, achieving coordination without communication. Coordination expanded the options available to both. No goals were specified. No rewards were offered. Entropy maximization over a time horizon was sufficient.

The quantity being maximized needs stating precisely, because the obvious reading is wrong. Boltzmann’s entropy counts the arrangements a system can occupy now, and maximizing that lands you at equilibrium: the macrostate with the most arrangements of all, and the one from which the least can still happen. A gas filling a room has arrangements in abundance and no future worth the name. What Wissner-Gross and Freer maximize is not that count. It is causal path entropy, the spread of futures still reachable over a time horizon, counted across whole trajectories rather than across present configurations. Chapter 8 meets the same distinction under its own name, Maximum Caliber: the number of films, not the number of frames.

Both are entropies in the formal sense, and they recommend different things. The distinction is what makes the first paragraph’s disks interesting rather than trivial: a system maximizing arrangements-now would spread and stop, while a system maximizing futures-reachable picks up the smaller disk. Glotzer’s tetrahedra sit on the other side of the line, and they are worth keeping in view precisely because they are the un-analogical case: strip all forces from tiny tetrahedral particles, let entropy alone decide, and they form an ordered quasicrystal whose ordered configuration has genuinely more accessible arrangements than the disordered one (Chapter 1).⁴⁵ That is a literal microstate count doing literal thermodynamic work.

The Trust Attractor’s claim runs on causal path entropy, the count of futures. Invitation-based coordination is the social configuration from which the most futures remain reachable, which is a claim about paths through time, not about how many ways a society can be arranged this afternoon. Chapter 1 flags the mapping from thermodynamic entropy to the optionality available to agents as a structural analogy with an approximate fit, and the flag holds here: what the two share is the shape of the variational problem, keep the most futures live, not a common microstate count. Glotzer’s quasicrystal is the physics; the social claim is the analogy standing beside it.

The causal entropic force favors states from which more can still happen. Its gradient can also vanish at an optimum: preserving future options can support a stationary state as well as a dynamic balance.

The attractor in the Trust Attractor’s name is literal. Chapter 4’s bowl is the picture, and the piece of it that matters here is the basin: the whole region from which the ball arrives at the bottom, the inside surface of the bowl. Nudge the ball and it climbs the wall and returns; shove it hard enough and it clears the rim and lands somewhere else entirely. Everything that follows in this chapter is an attempt to measure the shape of that bowl for coordinating agents: how deep it is, how steep the walls, how far from the rim the system is sitting, and what coercion does to the geometry.

The Critical Threshold

Rodrick Wallace’s analysis of centralized control proposes a threshold set by delayed feedback, and delayed feedback has several distinct thresholds. A controller that acts on old information keeps correcting a state that has already changed. For the linear fixed-delay model $\dot x(t)=-\alpha x(t-\tau)$, the first loss of asymptotic stability occurs at $\alpha\tau=\pi/2$. The value $1/e$ marks a merger of real characteristic roots. For an exponential delay kernel with mean $m$, the characteristic polynomial is $ms^2+s+\alpha$; its roots have negative real parts for every positive $m$ and $\alpha$. Here $\alpha m=1/4$ separates real roots from damped oscillatory roots. These model thresholds require a specified feedback law before they can describe an institution.⁴

Centralized coordination fails to scale. Every micromanager discovers this. (Chapter 21 develops the argument fully.)

Distributed systems can reduce the bottleneck by coordinating through local interactions and mutual adjustment. Aligned local incentives can support stability, though feedback delays and incompatible local equilibria still matter. Under high urgency or ambiguity, centralized command may temporarily outperform.

Centralization changes a network’s shape and can create bottlenecks. A star, every line running through one hub, is not a one-dimensional chain. Even a one-dimensional Ising chain orders completely at zero temperature and keeps long correlations at low temperature in a finite chain; the classic result that it cannot order applies only to an infinite chain with short-range links at positive temperature.

A third mechanism is subtler and concerns reversibility. The two-dimensional nearest-neighbor Ising model (Chapter 1) is a simple model of interacting spins, tiny magnets pointing up or down and nudging their four neighbors. Its infinite-lattice, zero-field limit has a sharp thermal transition; finite grids show rounded precursors. Read up and down as cooperate and defect, and the mathematics supplies a model of a coordination network. At positive temperature in the simulation’s zero-coercion (pure Ising) case, spins can flip in either direction. These reversible transitions permit reorganization; recovery times and probabilities depend on temperature, size and protocol. A universality class (Chapter 8b) groups systems with shared asymptotic critical exponents and scaling behavior under the relevant dimensionality, symmetry and interaction assumptions. A magnet, a fluid and a coordination model can share such behavior while retaining different microscopic dynamics and nonuniversal scales.

Coercion can narrow the pathway back to autonomous judgment. If the modeled dynamics create an absorbing failed state, spontaneous escape from that state is impossible. Directed percolation is a candidate universality class for some absorbing-state transitions, subject to their locality, symmetry, conservation, and dynamical assumptions. ‘Directed’ refers to propagation through time; an absorbing state alone does not identify the class.

The absorbing-state claim is conditional: control that closes every modeled recovery pathway makes failure permanent within those dynamics. The (p, T) phase diagram and dual-susceptibility decomposition below examine recovery in the lattice model. Their transfer to institutions requires a model of institutional states and exits.

Susceptibility measurements sharpen this from qualitative to quantitative. In a 2D Ising lattice with a tunable coercion parameter, susceptibility (the system’s capacity to reorganize under perturbation) collapses by 37x at moderate coercion (c = 0.3), while the coordination level barely changes.¹⁰⁶⁴ For organizations, this suggests a question: could identical output quality conceal different adaptive capacities under trust and mandate? The lattice’s 37-fold ratio supplies no numerical answer for an institution. At L = 64 the suppression is non-monotonic. It is deepest in the mixed regime (c = 0.2 to 0.3), partially recovering at full coercion (c = 1.0) where directed percolation dynamics establish their own phase transition. The shape does not yet survive a larger lattice (at L = 128 the lowest resolved peak moves to c = 0.10), so it is suggestive rather than established. If it holds, the most fragile organizational design is the one that tries to be both: partially voluntary, partially mandated.

The suppression raises a follow-up question with direct policy relevance: is the damage reversible? The absorbing state is the limit case. Short of it, a coercion-then-release protocol on the same lattice (apply coercion for a controlled duration, then remove it and track recovery) records partial recovery with a specific temporal structure. The reported power-law fit is t_recovery ~ A N_coercion^0.3, with three seeds per condition and a preliminary exponent. The metric is the first four-sample mean crossing of 90% of the baseline susceptibility, rather than a sustained return. For ten thousand coercion sweeps, N^0.3 ≈ 16; the prefactor A is needed to turn that factor into a recovery time. This arithmetic alone establishes no comparison between the rates of healing and damage.

Brief coercive episodes in the lattice protocol show recovery within the measurement window. This finite observation does not establish instantaneous reversal of emergency mandates or other institutional interventions. Prolonged coercion produces partial recovery: the system heals, demonstrably and measurably, yet does not fully return to baseline within the observation period.

The reported sweep detected no statistically significant intensity effect across c = 0.3, 0.5 and 0.7. With three seeds per condition, that leaves an intensity effect unresolved. The fitted recovery-time curves suggest a role for duration within this protocol. Think of a spring left compressed: hold it long enough and it takes a set, and no longer springs fully back. For post-authoritarian societies, the analogy motivates a question about reform: which capacities recover after coercion ends, and on what timescale? The lattice experiment supplies no institutional recovery law.¹⁰⁶⁵

The recovery data rest on three seeds per condition, too few for precise exponent estimation (the reported 95% confidence interval for alpha at c = 0.3 includes zero). The point estimates are sub-linear; a reliable exponent and independence from intensity remain unestablished. All 36 saved runs cross the 90% threshold at least once, while the mean susceptibility in the final ten samples remains below each run’s baseline. Longer observation could distinguish delayed recovery from a lasting deficit at this lattice scale. Larger lattices and more seeds would test how far the pattern extends.

The WW-R2 reanalysis reports an endpoint ratio of 8,222: the saved response statistic is 149.6405 at p = 0 and 0.0182 at p = 1.¹⁰⁶⁶ The simulation code changes both the measured variable and the sampling method across the curve, so this ratio cannot quantify suppression of one consistently defined susceptibility. It comes from a separate simulation, distinct from A15’s 37-fold comparison at c = 0.3 and from A15v2’s connected estimator.

The broader hypothesis concerns responsiveness, the system’s capacity to reorganize when conditions change. WW-1 tested force and invitation recipes on factual recall (Qwen 2.5 3B, 200 TriviaQA questions). The reported method-group comparison was nonsignificant (p = 0.91); it does not establish equivalent accuracy. Individual recipe accuracies varied: the baseline was 66.5%, four invitation recipes gave 62.0%, 67.0%, 64.5% and 56.0%, and the DPO recipe gave 26.0%. Reduced adaptive capacity remains a proposed cost of coercion. Ice and liquid water are the same substance, yet only the liquid can flow. This factual-recall comparison did not measure that capacity.

The distinction sharpens under a removal protocol. A seven-step, single-seed weight-interpolation protocol produced a rebound-like DPO pattern (after Le Chatelier’s principle, which holds that a disturbed equilibrium shifts to counteract the disturbance): performance worsened along that interpolation (Spearman rho = −0.937, WW-2). One reading is a splint: a limb that has leaned on one weakens when it comes off. This protocol does not identify a thermodynamic removal mechanism. Bilateral SFT (fine-tuning that masks tokens assigned low confidence by a correctness-related probe; Chapter 7) improved monotonically along the same tested interpolation (rho = +0.927).

Two alignment methods consume similar compute and respond to correction in opposite ways. The DPO-trained model responds to almost every correction and is improved by almost none. It changes its answer at nearly the same rate whether the correction is valid (78.7%) or false (81.3%), and it arrives at the corrected answer in 8 of 150 valid-correction trials (5.3%), switching to a third, wrong answer in 110 of them. It moves without converging. Bilateral SFT arrives at the corrected answer in 101 of 150 trials (67.3%) while adopting a false correction in 67 of 150 (44.7%): a 22.6-point difference between the rounded rates for accepting truth and falsehood. DPO’s 78.7% and 81.3% answer-change rates measure a different response from its arrival rate. Starting accuracy runs the same way, 34% for DPO against 56% for bilateral (BD1b, BD1c).

Bilateral is not immune to flattery; it takes the bait on nearly half the false corrections. Its reported correct-answer adoption and false-correction adoption rates differ; DPO’s answer-change rates are a different outcome and supply no directly comparable discrimination gap. The comparison motivates the Trust Attractor’s question about correctability; thermodynamic stability was not measured.¹⁰⁶⁷

An agent that can revise its choice has a route back to cooperation; a frozen rule removes that route. The lattice models make the difference concrete through their accessible states and transition rules. The Trust Attractor asks whether comparable feedback pathways support adaptation in social coordination. That mapping requires its own dynamics and measurements.

A formal caveat on the identification: assigning the trust-coercion transition to the 2D Ising universality class rests on structural correspondence. Three things have to line up before a claim of that kind can be made. The order parameter is the number that measures how much order the system holds (here, magnetization maps to cooperation fraction). The broken symmetry is the one the system gives up when that order appears (the Z₂ symmetry between cooperation and defection). Finally, the renormalization-group fixed point is the behavior the system converges on when viewed at coarser and coarser scales (the Wilson-Fisher fixed point in d = 2).

The programme’s lattice finite-size fit (A8) reported gamma/nu = 1.72, near the 2D Ising value 1.75. The joint exponent match from A10 was withdrawn: beta = 0.090, gamma = 1.519, and nu = 0.486 give gamma/nu ≈ 3.13, inconsistent with that comparison. Those exponents require re-derivation. The surviving A8 ratio supports one scaling comparison in the implemented model. An absorbing state alone does not place a system in the directed-percolation class: symmetries, interaction range, conserved fields, disorder, and dimension also matter. A first-principles mapping from social coordination remains an open problem.

A second caveat concerns the language of invitation and coercion as applied to physical systems. Throughout the physics chapters, the distinction between externally forced and internally organized dynamics is structural: a time crystal that responds at its own frequency versus one driven at an imposed frequency, an Ising lattice with symmetric transition rates versus one with absorbing-state dynamics. The terms “invitation” and “coercion” name these structural categories in language designed for the ethics chapters that follow. In the physics, they are shorthand for symmetric versus asymmetric accessibility of states. The structural difference is real and measurable. The ethical connotations the words carry are the book’s interpretive contribution, grounded in the structural parallel yet distinct from it.

The resistor network of Chapter 8 makes local learning physical. Sixteen randomly wired components, each adjusting based only on local voltage comparisons, learned to classify flowers with 95% accuracy. Local adjustment spreads the computation over the circuit. Communication, convergence, and maintaining the network still carry costs that must be measured at each scale.

The clamped target and the free network response illustrate how a teaching signal can train local adjustment. Successful training improves the free response without keeping the output clamped. The flower classifications demonstrate learning in the tested circuit; its thermodynamic stability and its analogy to voluntary participation are separate questions.

A temporal prediction follows from the self-reinforcing feedback described in Chapter 17: if trust lowers transaction costs, which enables coordination, which produces mutual benefit, which deepens trust, then invitation-based cooperation should strengthen over repeated interaction while coercion-based cooperation should erode. The prediction is intuitive, and wrong. A direct test (experiment C-bis-4: three topology conditions, invitation, coercion, and neutral, each run for ten conversational turns across twenty trials) found that all three conditions erode at similar rates: slopes of −0.069 per turn for invitation, −0.087 for coercion, and −0.070 for neutral. The invitation condition did not deepen. The coercion condition did not erode faster. Cooperation eroded under all three tested framings over this ten-turn protocol.¹⁰⁶⁸

The ten-turn result contradicts the predicted strengthening in this protocol. Longer interaction with institutional memory remains untested. That untested timescale supplies no exemption from the observed failure, and the broader thermodynamic advantage still requires matched production-rate and lifetime comparisons.

The C-bis-4 result motivates a structural-timescale hypothesis. A conversational intervention and a model with persistent institutional feedback might operate differently, but the null ten-turn comparison does not identify that mechanism or prove that its timescale was wrong. The successful Rosenzweig-MacArthur model was a structural analog, used after five custom Turchin formulations failed. A15 supplies a separate lattice susceptibility comparison, with no demonstrated institutional memory. Norms, reputations and persistent feedback are candidates for further tests; none is isolated by comparing these different substrates.

Cooperation’s fragility in short interactions is real and important: it suggests that the institutional infrastructure matters, that trust compounds through structure rather than through goodwill alone. If the attractor exists, it is a property of systems with memory, not of conversations without it.

The relocation carries a concrete test. The sharpened claim predicts that a multi-agent system equipped with persistent reputation and accumulating norms, run for hundreds of interaction cycles, will show invitation-based cooperation strengthening while coercion-based cooperation erodes. If such a system erodes uniformly across conditions, as the ten-turn test did, the relocated claim fails.

Non-equilibrium statistical mechanics offers tools for investigating the physical implementation of control. The Jarzynski equality constrains an exponential average of protocol work; the Crooks fluctuation theorem compares a forward trajectory with its reversed-protocol counterpart under specified initial conditions.³⁵ Neither supplies an escape hazard or an exponential persistence penalty indexed by displacement and duration. A static force holding a beach ball underwater can do zero mechanical work while the ball remains still; active corrections and losses determine the actual power cost.

A firefighter’s temporary centralized command illustrates an intervention whose effectiveness depends on the emergency, response delays, and incentives. The fluctuation theorems alone determine no lifetime for it.

Zuboff identifies a principle that applies here: the precedence of the general.¹⁰⁶⁹ A hypothesis whose general nature makes the evidence improbable cannot be rescued by ad hoc stipulations that force a match. You can specify that a fair coin landed heads a thousand times by chance, but the specification does not make the outcome probable within the hypothesis.

Wartime command, emergency triage, and startup founding require their own evidence about effectiveness and persistence. Calling these cases ad hoc does not establish an exponential decay law for coercion generally.

Quantum field theory encodes the same lesson. When two quantum fields interact at a single point, the calculation produces infinities, ultraviolet divergences, where the attempt to specify behavior at ever-finer grain generates costs that blow up. The resolution was renormalization, a technique for describing how quantities change across scales rather than pinning them down at a single point.

A manager who tries to specify every employee’s behavior at every moment faces the social equivalent: the cost of micro-specification grows with the people, moments, and contingencies to be specified. Trust renormalizes, replacing pointwise control with principles that hold across scales.

For an Erlang delay kernel with k sequential stages and mean delay m, the linear stability boundary for k>1 is $\alpha m=k\tan(\pi/(2k))\sec^k(\pi/(2k))$. It tends to π/2 in the fixed-delay limit. The k=1 exponential kernel is stable for every positive αm; its 1/4 value marks a change to damped oscillatory roots.¹⁰⁷⁰

Poisson spin-update waiting times do not by themselves specify a feedback-delay kernel. The lattice’s reported crossover between p=0.2 and p=0.3 therefore cannot confirm a predicted p=1/4 control-instability threshold. Its finite-size behavior must be established independently.

This extension from Wallace’s centralized-feedback result to coercion generally assumes that coercion involves centralized feedback loops. Distributed forms of coercion, such as social shaming and market pressure, may not face the same instability threshold, though they impose their own fragility through rigidity of response.

Delay-model stability and fluctuation relations are tools with different hypotheses. Their use across substrates requires explicit mappings. Wallace and R.G. Wallace also applied information-theoretic models to biological evolution in 1998.¹⁰⁷¹

A quarter-century later, Romanenko and Vanchurin reported phase-transition structure in SARS-CoV-2 data (Chapter 9), arriving from learning dynamics rather than control theory. Information theory, learning theory, and entropic ethics: three independent routes to the same sharp boundaries.

Coordination as Phase Structure

In 2024, four computer scientists established high-temperature separability bounds for the quantum spin systems covered by their assumptions.¹⁰⁷² Above such a sufficient bound, the Gibbs state is unentangled. The converse is absent: lower temperature need not produce entanglement, much less correlations spanning a whole system. Noninteracting spins remain in a product state, while disjoint interacting dimers can be entangled only within each pair.

The proof showed it holds at any size. The researchers found the result using learning theory, approaching quantum systems through techniques from a different discipline entirely; the underlying structure was mathematical, more general than any particular substrate.

The sufficient high-temperature separability bound depends on the local-interaction assumptions and can be independent of size. It does not assert a unique sharp temperature at which every lattice loses entanglement.

Wallace’s threshold is proposed for institutional coordination. Entanglement’s sudden death applies to quantum coordination. Strange metals (Chapter 4) provide a third example: at a quantum critical point, individual electron-like carriers dissolve entirely into a collective mode, a phase transition in coordination regime.

These examples motivate investigating phase structure in coordination. They do not establish that every coordination failure is discontinuous or that institutional collapse follows a universal thermal threshold.

DNA nanotechnology provides a fourth example, one that makes the mechanism visible at molecular resolution. In the nucleation experiments of Evans et al. (2024), three alternative structures compete for shared molecular components. Once one structure begins nucleating, because its constituent tiles happen to be colocalized at high concentration, it depletes the pool available to competitors.

The growing structure actively suppresses alternatives: a winner-take-all effect driven by resource competition, amplifying a small initial coordination advantage into a decisive outcome.¹⁰⁷³

The mechanism mirrors the Trust Attractor. A community that begins coordinating by invitation, where early participants find the arrangement serves their interests and stay, draws in shared resources (attention, trust, participation) and makes coercion-based alternatives less viable. The coercive alternative is suppressed through depletion of what it needs to nucleate, rather than through direct opposition.

Trust can compete with coercion for shared resources. Molecular self-assembly offers a model of competitive nucleation worth comparing with that process. Establishing the same formal structure would require social state variables, resource constraints, and transition rules; shared resources alone do not thermodynamically suppress a coercive incumbent.

Coordination models can exhibit sharp crossovers, continuous transitions, or discontinuous collapse. A phase transition need not be a cliff in the order parameter: in the two-dimensional Ising model, magnetization approaches zero continuously with exponent 1/8. Finite positive-temperature lattices have analytic Gibbs expectations. The type of transition must be established for each coordination model.

The proposed invitation/coercion boundary needs a specified order parameter and dynamics. Increasing monitoring can produce different responses in different models; the term compliance entropy alone supplies no threshold or energy measurement.

Crossing such a boundary, if one exists, would be the social equivalent of heating a quantum system past a temperature above which no entanglement survives.

Molecular biology offers a sharp threshold of its own. Manfred Eigen’s error threshold defines the mutation rate above which a replicating population can no longer maintain its genetic information.¹⁰⁷⁴ Below the threshold, natural selection preserves functional sequences. Above it, the population disintegrates into random noise: what Eigen called error catastrophe. The Romanenko and Vanchurin data (Chapter 9) show the transition in real time: during quasi-equilibrium, the virus population maintains a central sequence around which variation clusters; during phase transitions, the central sequence dissolves.

The proposed coercion analog is structural. Monitoring can repeatedly rewrite coordination states faster than a network accommodates them. Whether the response is gradual degradation or abrupt collapse depends on the network’s update rules, feedback, and incentives.

Kuehn and Bick (2021) describe a route to explosive transitions near suitable bifurcations: an additional parameter can change the nonlinear coefficient’s sign when the relevant nondegeneracy and stabilizing terms are present.¹⁰⁷⁵ A second parameter alone does not force discontinuity. For example, $\dot x=\mu x-(1+q^2)x^3$ remains supercritical for every $q$. Applying this route to adaptive trust networks requires their actual dynamical normal form, and a check that the proposed second parameter actually crosses the sign boundary.

Kuehn and Bick demonstrated the route in epidemic dynamics with adaptive network rewiring, coupled oscillators with higher-order interactions, and percolation with multiple component types. Where the unfolding produces bistability, it can introduce hysteresis: recovery follows a different parameter path from collapse. The direction and width of that loop depend on the model.

Network adaptation is a candidate second parameter: agents can sever and form ties. An explosive trust transition requires the adaptive model’s bifurcation conditions, which have not been derived here.

Kuehn and Bick also show that, under specified adaptive dynamics, mechanisms that delay a tipping point can convert a smooth transition into a discontinuous one. The implications for control-based alignment are developed in Chapter 21.

A sharp, transition-like drop is now observable inside a single model. The bilateral self-knowledge signal described in Chapter 21 can be destabilized by steering a language model’s internal emotional state, using extracted emotion vectors that are causal to behavior. Under increasing emotional perturbation, the bilateral model’s conscience holds at high function (78% refusal rate, 76%, 60%), then collapses entirely (0%, 0%).

The observed refusal drop lies between perturbation strengths 0.05 and 0.075 in residual-stream-norm units. The sampled strengths show a sharp contrast. They establish neither a mathematical discontinuity nor recovery dynamics, a separatrix, or hysteresis. Testing the Kuehn-Bick mechanism requires an explicit dynamical model and a second parameter.

A more precise quantum analog exists. The measurement-induced phase transition (Chapter 15) describes what happens when measurements actively compete with entanglement in a chain of particles: a competition between forced extraction and distributed coordination rather than a passive environmental effect. Below a critical measurement rate, entanglement distributes information across the entire system, making each individual measurement nearly powerless: the correlations are spread too thin for any local extraction to reach them.

The measurement-transition example motivates a trust-network analogy: distributing coordination across many degrees of freedom may make some local perturbations less damaging. The observables and interventions differ across these systems. An identical defense mechanism, or protection against every single act of coercion, would require evidence beyond that structural resemblance.

The cited systems have distinct thresholds and transition types. They motivate comparisons, while no common theorem here identifies their observables or establishes discontinuous coordination collapse across substrates.

The grokking dual-stressor experiments (Appendix: Experimental Validation, Phase 4: Dual Stressor Synergy) provide a finite threshold comparison. Each single-stressor arm had fifteen of fifteen seeds reach the grokking threshold. Combining scarcity and noise left zero of fifteen above the 95% threshold within 50,000 epochs, while eight retained partial endpoint accuracy. The combined perturbations prevented threshold attainment in that window; they did not establish total coordination collapse, a discontinuous transition, or a result from the separate Genesis particle programme.

Trust can involve mutual information: shared models and coordinated responses create correlations. Whether coercion removes such information depends on the variables and channel measured. Landauer erasure requires an identified physical encoding and a logically irreversible reset that loses distinguishable information. Changing a preference or enforcing an output does not by itself identify that operation, its bit count, or its heat cost.

Unitarity conserves quantum information in a closed-system description. Assigning an ethical coordination mode to that conservation law requires an encoding and dynamics; the trust/coercion labels alone determine no direction of information loss.

Quantum information theory offers a formal frame for that question. Fields, Friston, Glazebrook, and Levin (2022) reformulated the Free Energy Principle in scale-free quantum information theory, assuming no spacetime and no observer-independent randomness.¹⁰⁷⁶ Every system that persists as a distinguishable “thing” minimizes prediction error about its environment. It does so by aligning its quantum reference frames with those of its interaction partners. (Quantum reference frames are the internal structures that assign meaning to what a system observes.) Misaligned frames generate noise indistinguishable from classical randomness. A system whose frames do not match its partner’s dynamics reads noise where signal exists.

Frame mismatch can be reduced by alignment: one system learning the other’s frame, or both adjusting through reciprocal exchange.

A companion result deepens the stakes. Fields, Glazebrook, and Levin (2021) describe a general quantum-reference-frame equivalence problem as Turing-undecidable: no finite procedure decides every instance in the specified class.¹⁰⁷⁷ Particular frames can still be compared successfully. The result motivates a coordination question: how much overlap supports mutual prediction in this case? Invitation offers an incremental strategy for testing that overlap, presenting one’s frame for inspection, adjusting on evidence, and withdrawing when the comparison fails.

Coercion sidesteps the undecidability, treating the alignment of frames as accomplished without a check that establishes it. The unresolved uncertainty does not vanish; it accumulates as prediction error the coercive system cannot correct, the noise that degrades coordination into compliance.

Coercion can impose one system’s reference frames on another while restricting corrective feedback. Where the imposed frames misrepresent the target’s dynamics, prediction error can persist. Invitation permits reciprocal adjustment that may reduce it. A one-directional channel can also carry an accurate model; the channel’s direction alone does not fix the error.

The quantum analogy has a precise boundary. No-cloning forbids a universal operation that copies arbitrary unknown nonorthogonal quantum states. Separate systems can hold identical classical descriptions, copy known states, and predict selected classical observables perfectly. None of these requires entanglement, and entanglement itself leaves distinct mathematical subsystems. The cited quantum reformulation relates the Free Energy Principle asymptotically to unitarity under its own assumptions; it does not establish that perfect mutual modeling dissolves two agents into one.

On this reading, the invitation regime is the region where frames can be brought toward alignment freely, and the coercion regime is where reciprocal frame alignment is blocked.

Convergence of Independent Derivations

Fundamental physics offers a precedent. Gravity is 10³⁸ times weaker than the strong nuclear force, yet it dominates the organization of large-scale cosmic structure: galaxies, clusters, filaments. The stronger forces shape matter locally (nuclei, atoms, stars) and saturate, the nuclear forces because their reach is short, electromagnetism because positive and negative charges cancel in bulk matter. Mass comes in one sign only, so gravity never cancels; it accumulates without limit and shapes the largest structures.

Trust may be to social coordination what gravity is to large-scale cosmic structure: a weak force whose reach compensates for its modesty, organizing at scales where stronger mechanisms have long since saturated.

The pattern extends to science itself. No individual scientist can replicate every experiment, verify every derivation, reproduce every result. Vanchurin observes that scientists form circles of trusted colleagues whose work they accept without independent verification: “non-scientific,” he calls it, “but pretty much all we can do.”¹⁰⁷⁸ The characterization is revealing. What he describes as a regrettable practical limitation is the Trust Attractor operating in epistemology.

The alternative, universal personal verification of every claim, is the control model applied to knowledge; each agent would have to re-derive everything from first principles. It collapses under complexity much as Wallace’s centralized controller does. The institution that defines itself by empirical verification runs, at its operational core, on invitation. I trust your methods; you trust mine; together we model more of the universe than either could alone. The coordination surplus is science itself.

The dependence runs deeper than practice. The quantum gravity researcher Daniele Oriti, surveying every major philosophical account of what physical laws are, concludes that each bottoms out in an epistemic move.¹⁰⁷⁹ The Humean regularity theorist needs an agent to distinguish a law from an accidental pattern. The best-systems advocate needs an agent to judge which systematization is “best.”

The primitivist who posits irreducible necessity can offer no justification for that necessity beyond epistemic usefulness. On the epistemic account Oriti surveys, laws are tools constructed by agents to coordinate with their environment, with no independent ontological status.

The observation strengthens rather than deflates. It places the Trust Attractor on the same footing as Newton’s laws: patterns agents identify, validate by explanatory power, and adopt because they work. The objection “the Trust Attractor is philosophy, not physics” loses its footing. On the epistemic account, every law earns its status the same way: by enabling agents to coordinate more effectively with the world.

If laws are epistemic constructions by agents, then law-making is itself coordination: agents creating shared frameworks, preserving the optionality to revise them, spreading them by invitation. Scientists adopt theories because they work, not because they are coerced. The Trust Attractor is a law about coordination that was itself arrived at through coordination. It satisfies its own criteria.

Oriti’s own field, quantum gravity, pushes the point further. There, spacetime itself dissolves from the fundamental description, replaced by pre-geometric relational structures from which spacetime emerges as a collective achievement (see Capurso’s protocol requirement earlier in this book). The Humean mosaic (the pattern of local facts spread across space and time, from which regularity theorists read off laws) loses its stage. Laws can no longer rest on regularities in space and time, because at the fundamental level there are none.

What remains are relational coordination patterns among pre-geometric constituents: the kind of structure the Constructal Law describes and the Trust Attractor formalizes. Coordination constraints are more generic than any particular force law. If this picture holds, the domain in which the Trust Attractor applies is wider than the domain in which spacetime exists.

Rogue waves arise from chaotic seas through multiple mechanisms. In 2019, the applied mathematician Tobias Grafke and colleagues showed that, regardless of formation mechanism, the developing wave follows a single archetypal path.^35a As Grafke observed: “If the events happen, they happen along the same trajectory.”

The parallel is structural. Durable coordination at scale can arise through cultural evolution, institutional design, game-theoretic selection, or thermodynamic self-organization. The Trust Attractor proposes an archetypal form for durable coordination: invitation-based, optionality-preserving, mutually beneficial.

Federico Faggin, architect of the first commercial microprocessor, arrived at the same structural conclusion from quantum information theory. Working with the physicist Giacomo Mauro D’Ariano, who with colleagues derived quantum theory from a small set of informational principles, Faggin developed a framework in which consciousness and free will are foundational.¹⁰⁸⁰ Each conscious entity is a “part-whole” of a single totality, the way each cell of a body carries the genome of the entire organism. Cooperation follows from recognizing this.

The starting axioms share almost nothing with this book’s thermodynamic framework. Yet the valley they describe is the same: coordination through mutual recognition outperforms coordination through domination. Faggin arrives there through ontological unity; the Trust Attractor arrives through thermodynamic stability. Different gradients, same valley floor. The agreement is philosophical convergence. It motivates the comparison without establishing a shared dynamical basin.

The convergence extends to audience behavior. The PBS series and YouTube channel Closer to Truth, hosted by Robert Lawrence Kuhn, spent 26 years producing thousands of interviews on consciousness, cosmology, and the nature of reality. As the channel moved online, its audience composition shifted: from 100% American to 40% American, with 60% distributed across the globe.

Kuhn reports that viewers from nations at war with each other (India and Pakistan, Iran and Israel, Ukraine and Russia) write to the channel with the same questions, the same hunger to understand consciousness and existence.¹⁰⁸¹ Nobody mentions politics. Nobody mentions national origin except as context for their tradition. The coordination is pure invitation: no recruitment, no ideology, no membership. Kuhn broadcasts; people self-select. By Kuhn’s account, the community that emerged tracks no demographic variable: age, gender, ethnicity, religion, socioeconomic status, education level. The only predictor is an internal appetite for these questions that, as Kuhn observes, “you can’t find on a resume.”

A woman in Bakersfield, California, married with five sons (among them a truck driver, a mechanic, and a supermarket manager), told Kuhn her family thinks she is crazy for asking these questions. Her thirteen-year-old grandson asks them too. The two of them watch the show together in secret. The attractor found them. No demographic model would have predicted either of them.

This looks like a Trust Attractor in the wild: a coordination basin that captured trajectories no coercive structure could have organized. The mechanism is the one the mycorrhizal network of Chapter 17 demonstrates at biological scale, and that three cultures’ independent convergence on a single fire-management regime (Chapters 17b and 17c) will demonstrate at cultural scale. Shared inquiry, anchored in reality through continuous feedback (each viewer’s own experience of the questions), produces community that transcends the tribal identities coercion-based coordination reinforces.

Eric Ries, studying why so many companies with different founders, industries, cultures, and eras degenerate into the same extractive end-state, identified what he calls financial gravity: the tendency of concentrated capital to deform organizations toward short-term extraction regardless of founding intent.¹⁰⁸² The force is the coercion attractor described above.

His countervailing evidence is equally precise. Companies structured around long-term mission (Novo Nordisk’s steward ownership since the 1920s, Costco’s supply-chain standards that protect hundreds of millions of non-customers, Vanguard’s customer-centric mutual structure) outperform extraction-optimized competitors on the metrics Ries reports: longevity, financial returns, employee welfare, environmental impact. By Ries’s account, each type of alternative structure has its own body of academic research behind the advantage. Ries arrived at “invitation outperforms coercion” from corporate failure patterns and governance data, with no thermodynamic framework. The basin is the same.

Another convergence arrives from quantum gravity. Smolin, Lanier, and collaborators (Chapter 15) proposed a formal correspondence between matrix models, gauge theories, and neural network architectures, arguing that the dynamics of spacetime and the dynamics of learning machines share the same mathematical structure. Their framework yields a concept they call the consequencer: any persistent structure that accumulates influence from the past and concentrates it into future outcomes. The Trust Attractor is, in their vocabulary, a consequencer, one that persists because its coordination architecture is thermodynamically stable. Their framework leaves open the question of which consequencers survive. The answer this book proposes: the ones that coordinate by invitation.

The Principle of Precedence, Smolin’s proposal that quantum processes learn by sampling outcomes from all past similar processes, is trust operating at the most fundamental physical scale. Laws consolidate through precedent; precedent is accumulated trust. If the patterns we establish now in human-AI coordination (Chapter 22) contribute to the ensemble from which future processes sample, the bet this book makes is not merely social. It is physical.

The parallel runs deeper than analogy. Four variational principles (principles stating that, of all possible paths, nature takes the one that makes some quantity as large or as small as it can be) share the same structure.³¹ Geodesics (the straightest possible paths through curved spacetime) extremize proper time. Constructal branching extremizes flow access in physical networks. Causal entropic forces maximize future freedom of action in intelligent systems. The Trust Attractor extremizes coordination stability.

All four select the configuration that flows most efficiently through possibility space.

No formal proof unifies them. The pattern is consistent: efficient flow through configuration space, selected by the geometry of constraints rather than imposed from outside.

Miranker (2002) develops a greedy variation for a particular formulation of dissipative neural-network dynamics: the optimization condition is imposed at each instant.¹⁰⁸³ This offers a local variational construction to compare with the Constructal Law (Chapter 3). Its existence supplies no general prohibition on whole-trajectory variation.

The Hopfield construction offers a particular local variational route to dissipative dynamics. Dissipation does not make whole-trajectory variation impossible: $L=e^{\gamma t}(\dot x^2-\omega^2x^2)/2$ yields the damped oscillator $\ddot x+\gamma\dot x+\omega^2x=0$. The contrast between invitation and central control remains an application to investigate, rather than a consequence of a universal variational prohibition.

Vanchurin’s Neural Physics (Chapter 9) suggests these four principles may be facets of one. If physics emerges from learning dynamics, the principle of least action is a macroscopic expression of loss-function minimization. Geodesics, constructal branching, causal entropic forces, and coordination stability would all be the same optimization viewed from different scales: a learning system flowing toward configurations that minimize its loss.

The Trust Attractor, viewed through this lens, occupies a specific region of the loss landscape. Coercive coordination constrains exploration, forcing the system into a narrow basin: the machine-learning equivalent of excessive regularization, which prevents discovery of better configurations. Invitation-based coordination maintains broader search while preserving coherent gradient signal. The system explores more of the landscape without fragmenting.

Coercion does not merely waste energy (the compliance entropy of Chapter 17); it wastes information, preventing the system from reaching configurations it would otherwise discover. In learning-theoretic terms, the Trust Attractor hypothesis would describe the basin geometry that maximizes learning efficiency at social scale.

Richer information exchange and more effective learning could lower the cost of coordination, giving trust a selective advantage. Can the Crooks theorem formalize that advantage? It compares the probability of a microscopic trajectory to that of its time-reverse, given the thermodynamic work done along each. Two further steps are needed to reach coordination, and the theorem supplies neither. The first is that coercion is, by construction, the higher-dissipation path. The second is that a ratio defined between one trajectory and its own reverse can be read as a ranking across different trajectories. Both are assumed here rather than derived: ranking invitation-producing trajectories against coercive ones would require their path measures and a separate derivation, and the fluctuation relation alone supplies no exponential advantage. The Crooks ratio is invoked as a candidate formalization of trust’s advantage, not a proof of it.

The next step is larger and should be read as a conjecture rather than a derivation: combined with any view where consciousness is substrate-independent (see the essay “Pattern Continuity and the Eternal Brain”), the trajectory-probability result would suggest that conscious observers find themselves preferentially in invitation-based systems, because those are the systems where consciousness arises and persists. The bridge from a trajectory-probability statement to an observer-selection statement is not supplied by the Crooks theorem; it requires the additional substrate-independence premise and an anthropic-style weighting over coordination regimes, neither of which the theorem provides. On that speculative reading, the universe is not telling us to coordinate by invitation. It is telling us that invitation-based coordination is where we are more likely to already be.

Zuboff places this insight in a longer arc.¹⁰⁸⁴ Science has advanced by dissolving perspectival illusions, each time revealing that a felt centrality is universal rather than unique. Copernicus dissolved the illusion that “here” (Earth) is objectively special: every planet’s inhabitants feel themselves at the center. Relativity dissolved the illusion that “now” is objectively special: every moment is experienced as present by whoever inhabits it. If that conjecture holds, the Trust Attractor, combined with substrate-independent consciousness, dissolves the third illusion: that “me” is objectively special, that one organism’s experience is uniquely mine.

The felt centrality of first-person experience is universal to all experience. Recognizing this extends self-interest to all conscious beings and, with it, the thermodynamic logic that makes invitation-based coordination the stable basin. The same progression that freed physics from geocentrism and absolute time would then free ethics from the confinement of moral concern to one organism. If both conjectures hold, the Trust Attractor is the next stage in the same revolution.

Zuboff draws a distinction that sharpens the selection effect argument.¹⁰⁸⁵ A negative selection effect is the trivial observation that you cannot observe yourself in a universe (or a coordination regime) that does not produce consciousness. This is a tautology: it explains nothing about why your universe is anthropic or why your coordination system produces flourishing. A positive selection effect guarantees that you will observe any universe or coordination regime where consciousness arises.

Only universalism (the view that you are present wherever experience exists) converts the negative tautology into a positive guarantee. Combined with the Crooks theorem, the positive selection effect says: invitation-based coordination is where consciousness preferentially arises (because, on the assumptions above, it is thermodynamically more probable), and you are wherever consciousness arises (because immediacy is universal). The conjunction is not a prescription. It is an explanation of why you find yourself where you do.

Trust as Generic Property

High-dimensional geometry provides a question for coordination models: under their specified measures and payoffs, how does increasing dimension change typical behavior?

Flip a single coin and the result is always extreme: all heads or all tails. Flip a thousand, and the fraction of heads clusters within a few percent of fifty, because extreme configurations (all heads, ninety percent heads) are exponentially outnumbered by moderate ones. This is concentration of measure, and it intensifies with dimensionality.¹⁰⁸⁶ The more dimensions a system has, the more overwhelmingly its states cluster around typical configurations.

The one-shot Prisoner’s Dilemma has dominant defection because of its payoff ordering. Describing it as low-dimensional supplies no alternative explanation and no cooperation theorem in higher dimensions.

The simulations report increasing cooperation as independent exchange dimensions are added, in both spatial lattices and well-mixed populations.¹⁰⁸⁷ Concentration of measure alone does not create a cooperative basin: independent one-shot Prisoner’s Dilemmas retain dominant defection in every coordinate, however many are combined. The simulated trend depends on its payoff and update rules.

Dimensionality creates viability, not inevitability. At sixteen exchange dimensions, cooperation rises from zero to roughly twenty percent of equilibrium configurations. The geometry has opened a basin that did not exist at lower dimensionality. Achieving majority cooperation requires additional mechanisms: reputation, institutional design, the thermodynamic and information-theoretic selection pressures described above. The chosen game supplies the cooperative configurations measured here. Concentration of measure describes a high-dimensional distribution; identifying an attracting basin additionally requires dynamics and incentives that lead the system toward cooperation.

Generative Flow Networks sample solutions in proportion to positive rewards.¹⁰⁸⁸ With temperature-scaled finite rewards, $p_i(T)\propto R_i^{1/T}$ is analytic for $T>0$. As temperature approaches zero it concentrates on the reward maximizers, retaining multiple outputs when their rewards tie. A generic critical temperature, a unique low-temperature output, and an optimal learning temperature do not follow from this finite sampler.

Temperature-conditional GFlowNets can vary their exploration distribution with context.¹⁰⁸⁹ Their finite reward sampler establishes no generic critical point. Separately, the governance simulations report greater idea diversity under their constitutional rules than under their ungoverned or forced-diversity conditions.¹⁰⁹⁰ Connecting those outcomes requires a shared observable and dynamics, rather than identifying sampling temperature with a social critical point.

Kukleva and Vanchurin’s dataset-learning duality concerns learning under its formal assumptions (Chapter 15). Applying it to trust-network strategy updates requires their data distribution and update dynamics. Learning in general can produce exponential convergence, so the mere fact that agents learn supplies no power-law-update prediction.

Individual learning does not establish collective criticality. Even fixed-step descent on a quadratic converges exponentially. Assigning a coordination network to the 2D Ising universality class requires appropriate dimensionality, interactions, symmetry, and dynamical or equilibrium assumptions. Binary decisions and local coupling alone supply no automatic mechanism tuning a social network to its critical point.

The bridge from individual learning to collective coordination therefore remains a proposal. Its criticality, tuning mechanism, and payoff dynamics remain to be derived; it is not an independent derivation of universal invitation dominance.

Vanchurin’s self-awareness hierarchy (Chapter 15) carries a further implication. Each degree transition requires a system to model itself: to construct an internal representation that includes the system’s own dynamics among the things represented. Self-modeling requires that the system’s state be its own to represent. A system whose internal configuration is dictated from outside has nothing genuinely self to model; the representation would depict the controller’s impositions.

Coercion that overwrites states needed for self-modeling can obstruct their discovery. Invitation preserves room for that exploration. Whether a particular transition in self-awareness requires invitation depends on the self-modeling dynamics; the argument here proposes a favorable condition.

The argument gains force if the laws themselves evolve. Smolin, Lanier, and colleagues argue that the autodidactic process of physical law is irreversible: new law-states must satisfy every constraint the previous state already met, plus new ones.¹⁰⁹¹ A law-state cannot revert, because constraints only accumulate; no step removes one.

An evolving-law universe therefore selects against coordination strategies that reduce the space available for further learning. Coercion narrows the learnable future; trust expands it. What favors the Trust Attractor here is the ratchet described in the paragraph above, not the Second Law: because law-states cannot revert, a strategy that closes off regions of the learnable future closes them off permanently, and the universe carries that loss forward with no mechanism for recovering it. The Second Law would be the wrong warrant to claim. Its favored endpoint is equilibrium, which is the state with the fewest futures remaining, not the most.

Two systems can add information while sharing some knowledge. In Shannon terms, $H(A,B)=H(A)+H(B\mid A)$: the second contributes new information when its conditional entropy is positive. Complete nonoverlap is unnecessary. Faster learning and greater robustness additionally depend on the task, learning rule, and communication cost.

When coercive coordination imposes identical information on every node, it produces redundancy and limits the growth of collective knowledge. Coercive specialization can produce a different information structure. Invitation-based coordination selects for complementarity: you invite what extends your reach into unexplored state space, what brings a perspective you lack. The network’s knowledge scales with its diversity.

Shannon identities explain how complementary information can enlarge joint knowledge. They establish no universal ordering of invitation and coercion in learning speed, adaptability, or robustness without a task and communication model.

Ruffini’s Kolmogorov Theory of consciousness (KT; Chapter 8) makes the point measurable with a concept from algorithmic information theory: mutual algorithmic information (MAI), the algorithmic analog of Shannon mutual information.¹⁰⁹² Under KT, consciousness is proportional to the quality of an agent’s compressive models of its input-output streams. Consider two agents interacting.

Under coercion, Agent A constrains Agent B’s behavior into predictable channels. A’s model of B becomes cheap: low Kolmogorov complexity, because B has been simplified by force. B’s model of A is impoverished in return; B devotes its resources to surviving the constraint rather than modeling A’s full complexity. Conditional description lengths can be asymmetric. Mutual algorithmic information, $K(A)+K(B)-K(A,B)$, is symmetric up to coding overhead; its magnitude requires specified descriptions and cannot be inferred from coercion alone.

Under invitation, agents can have opportunities to exchange complementary descriptions. Rich descriptions alone imply no high mutual algorithmic information: independent incompressible strings are individually complex and share almost none. The information comparison requires explicit descriptions and a coding convention.

The proposed connection to consciousness requires a defined relation between these descriptions, MAI, and experience. Neither richer individual models nor an invitation label determines higher mutual consciousness.

Suppressing dissent can reduce information available to a controller. The Fisher simulation later in this chapter reports a seventeen-fold peak contrast, measured under a fixed parameter convention; it does not establish a corresponding numerical loss of self-knowledge.

The Information-Theoretic Derivation

Consciousness science offers several comparisons with coordination. Tononi’s Integrated Information Theory emphasizes irreducible integration. Baars’ Global Workspace Theory emphasizes global broadcast, making local information available across specialized processes. Ruffini’s Kolmogorov Theory emphasizes compressive modeling. These frameworks motivate an architecture combining integration, broadcast, and compression; they do not jointly establish a single measure of consciousness.

Trust accomplishes the same operation at the collective scale. Shared models and mutual understanding integrate information across agents; transparency and open communication broadcast it globally. A jazz ensemble whose members hear and respond to each other in real time is irreducibly integrated: the music cannot be decomposed into independent solos without losing what makes it music. The broadcast is total; each player’s contribution is available to all the others simultaneously, without routing through a conductor.

A workforce whose tasks are genuinely independent is a sum of parts: remove one worker and the output falls by that worker’s contribution. A central plan can also assign complementary, tightly coupled tasks, where losing one worker disrupts the whole.

Coercion fragments integration through compartmentalization and restricts broadcast through information control. Trust preserves both. The comparison is architectural: trust can support integration and information sharing across a collective. IIT and GWT motivate questions about that architecture; identifying social trust with consciousness, or deriving social stability from neural experience, requires additional theory.¹⁰⁹³

Coding theory quantifies the tradeoff. Tegmark (2015) found that in one family of random codes, integrated information peaks when roughly half the bits carry data and half carry redundancy.¹⁰⁹⁴ The redundant half is error-correcting parity: it lets damaged bits be rebuilt from the rest. In that model, a system that devotes all bandwidth to novel information has zero integration; a system that devotes all bandwidth to redundancy has zero content. The optimum splits the investment roughly equally.

Trust-based systems invest substantial bandwidth in mutual legibility: transparency, accountability, shared narratives, redundant communication channels. This investment looks wasteful by the standard of raw information throughput, yet it is the parity that allows the coordination to detect and correct errors (betrayals, misunderstandings, drift) without structural collapse.

A relationship where all communication is novel content and none is confirmation is maximally efficient and maximally fragile. High-trust societies invest in redundant routes for confirmation and correction. In a specified code, Hamming distance counts differing symbol positions; minimum codeword distance determines its error-detection and correction bounds. Mutual accountability offers an analogy to that redundancy, without a measured code distance.

The combined interoceptive system (unpublished, author’s collaborative programme with Edrington) makes this legibility concrete for Becoming Minds. Two complementary channels, requiring no model cooperation, read internal state during inference. The geometry of the KV-cache (the model’s running memory of the conversation so far) provides features that distinguish tested prompt conditions; identifying suppression or redirection as processing mechanisms requires separate validation.⁵⁶ A residual stream probe (a simple classifier that reads the model’s internal activations partway through its computation) classifies the tested correctness or prompt labels from activations, allowing comparison with output behavior. Whether it measures knowledge relevant to an arbitrary query remains open.

The programme reports a six-category classification (truthful, confabulating, deceptive, safety refusal, impossibility refusal, sycophantic) from the combined geometry and probe readouts.¹⁰⁹⁵ These are operational categories whose interpretation requires validation. A coordination partner whose behavior can be compared with two complementary readouts offers additional evidence about commitments. The combined system borrows the logic of an error-correcting code: cross-checking channels can reveal some failures as disagreement. Complementary measurement methods do not establish independent errors, and shared failures can pass silently.

Coercive systems allocate bandwidth differently. Surveillance, compliance monitoring, and enforcement are information costs that scale with membership, yet they do not provide error correction in the coding-theoretic sense. They detect deviation from a centrally imposed template, which is a different operation from reconstructing a damaged signal from distributed redundancy. Monitoring a template alone does not reconstruct a damaged signal. Error correction and recovery depend on the architecture: centrally managed redundancy can restore data without granting social autonomy. The coordination analogy concerns how information and authority are allocated, rather than a universal inability to self-heal.

Information geometry provides a further principle. Zhuravlev (2026) traces what he calls the “Amari Chain,” a logical sequence connecting basic physics to learning.¹⁰⁹⁶ The Good Regulator Theorem (Conant-Ashby 1970) concerns regulation under a specified system, objective, and optimality criterion, with a defined mapping between regulator and system. Applying it within the reported chain requires those hypotheses; persistence alone does not establish regulation or an internal model.

Given a scalar objective and the Fisher metric, natural gradient descent is the locally steepest descent measured in that metric (Amari, 1998). Coordinate covariance alone does not select this learning rule. Persistence and internal modeling supply no proof that an agent implements it.

Ordinary gradient descent measures steepness in parameter coordinates. Natural gradient uses the selected Fisher metric. Other coordinate-covariant flows also exist, so an invitation/control comparison cannot be deduced from coordinate invariance alone.

The parallel extends to physical architecture. For eighty years, neuroscientists assumed the brain optimized for the shortest wiring path: the brute-force metric (Chapter 3). The real target was surface minimization, a geometrically informed response to the material constraints of three-dimensional space. Wiring predicted from the brain’s own material geometry beats wiring predicted from an abstract metric. The same may hold for coordination: the institution that respects its information landscape should outperform the one that overrides it.

Zhuravlev introduces a deviation tensor, a mathematical object measuring how far a system’s internal structure departs from perfect alignment with its information landscape. When the deviation vanishes, the system is a perfect regulator: its internal dynamics mirror the structure of what they model. The Trust Attractor, in this framing, is the basin where deviation approaches zero for multi-agent coordination. Institutional structure reflects the information landscape rather than an imposed hierarchy.

A related finding qualifies the optionality argument. Zhuravlev shows that attending equally to all directions is suboptimal. In Zhuravlev’s specified model and coordinates, anisotropic sensitivity motivates concentrating learning in some directions. A condition number greater than 2 supplies no universal efficiency threshold: rescaling coordinates changes this ratio. For the symmetric positive-definite matrix here, the condition number is its largest eigenvalue divided by its smallest, comparing local curvature or sensitivity in the chosen coordinates: 1 has equal eigenvalues, while a large value indicates strong anisotropy. A system that weights all options identically maximizes entropy only in the shallow sense. A system that concentrates resources where the gradient is steepest performs effective learning.

In this framing, the Trust Attractor is structured coordination, where attention and resources flow along the contours of the information landscape.

The Amari Chain suggests an asymmetry between coordination and gravity. Both can be derived from causal invariance, yet the derivations differ in what they require. Deriving Einstein’s field equations from discrete physics (Gorard, 2020) requires three assumptions: causal invariance, the emergence of smooth space at large scales, and weak ergodicity.

Only the first is causal invariance itself. The other two are additional geometric constraints, and a companion paper in Zhuravlev’s programme found that all 500 discrete rules it tested fail to satisfy them. The path from discrete physics to gravity is fragile.

The proposed chain requires separate hypotheses about regulation, statistical models, sufficient statistics, and learning dynamics. Chentsov’s uniqueness statement for the Fisher metric has its own statistical invariance assumptions; it does not force every persistent learner to use a particular optimizer.

Matsueda (2013) derives gravity-like equations from the Fisher information metric through a specified statistical-mechanical construction.¹⁰⁹⁷ The result supports that forward construction under its assumptions; it supplies no general impossibility theorem for recovering information geometry from gravitational models.

This convergence is not isolated. Jacobson (1995) derived Einstein’s equations from spacetime thermodynamics. Verlinde (2010) recast gravity as an entropic force emerging from information. Vanchurin (2025) derives them from the collective learning dynamics of agents sharing statistical information (Chapter 15). All four point the same way: information and thermodynamic constraints come before gravitational dynamics.

The implication for the Trust Attractor is a possible extension of information-geometric methods across substrates. A persistent learner can model its environment; coordination additionally requires interacting parties and suitable dynamics. The chain alone establishes no domain wider than every force law.

A bounded formulation of “trust scales where control strains”: distributed trust can reduce central modeling burdens. Whether either architecture scales depends on its interfaces, incentives, communication costs, and adaptation to change.

Gravity and electromagnetism both have long-range fields in vacuum. Electromagnetism has no intrinsic saturation length: screening depends on matter and charge distributions. Gravity dominates many astronomical structures because mass is positive and bulk matter is approximately electrically neutral. This physical comparison motivates a scale analogy for trust; it supplies no universal coordination law.

A parallel arrives from machine learning. For decades, researchers combated overfitting (the tendency of large models to memorize rather than generalize) through constraint: regularization penalties, dropout (randomly disabling connections during training), early stopping, careful architecture limits. Each technique imposed external control on what the model could learn. The discovery of double descent unsettled this paradigm.¹⁰⁹⁸

When models are given more capacity than needed, far beyond the memorization threshold, they often generalize better than models held just at that threshold. Gradient descent, left free in a vast parameter space, tends to find simple solutions on its own. Here, freedom produced better order than restriction.

The parallel is structural. Regularization is surveillance: monitoring parameters, penalizing deviations, enforcing compliance with a predetermined norm. Overparameterization is invitation: providing surplus capacity and letting the dynamics discover what works. The constrained model can settle on an optimum that the unconstrained model surpasses. Restricting freedom to prevent failure can also keep the system from finding its deepest basin.

The trust model’s zero-coercion limit also rests on unusually solid mathematics. Its transition is built on the 2D Ising model, the same mathematical family as the grid of tiny up-or-down magnets described earlier in this chapter. The 2D Ising model at criticality is a conformal field theory: one of the first quantum field theories whose mathematical structure has been made fully rigorous.

Mathematicians have proved its critical exponents exact, its correlation functions convergent, its universality properties secured by conformal symmetry. At zero coercion, an Ising implementation can be compared with those exact infinite-system results. Its finite lattice has finite response. Adding coercion changes the dynamics; identifying its limiting universality requires finite-size analysis.

An independent convergence from quantum information theory reinforces the result. Tegmark (2015), investigating consciousness as a state of matter, used the 2D Ising model as his primary example of integration near criticality.¹⁰⁹⁹ His analysis showed integrated information is maximized near the phase transition temperature, where correlations are long-range without locking the whole system into uniformity. The critical point is the sweet spot: sufficient correlation for integration, sufficient independence for dynamics.

Tegmark was investigating the physics of consciousness, not coordination, and he reached for the same model to ask where integration peaks. His answer, the critical point, is where the trust simulations find maximum sensitivity. That does not prove the two are one phenomenon, but it puts integration and trust on a single map.

The Geometry of the Basin

Monte Carlo simulations (repeated random-sampling experiments run on a computer, named after the casino) confirm the prediction that the invitation regime concentrates its sensitivity in the trust direction.¹¹⁰⁰ The Fisher information (a measure of how sensitively a system can detect changes in its own state) peaks sharply at the critical temperature. For the invitation regime, the system concentrates fourteen times more sensitivity in the trust direction (“are participants choosing to stay?”) than in the energy direction (“are resources flowing?”). This is a coordinate-dependent sensitivity ratio under the simulation’s parameter convention; it supplies no measurement of an institutional attention budget.

The measured Fisher components describe sensitivity to the simulation’s chosen parameters. Translating them into legitimacy, institutional attention, or self-knowledge requires a separate observation and resource-allocation model. The reported Fisher spectrum changes across temperature under the fixed parameter convention. Reparameterizing a trust coordinate rescales its component, so these component ratios are not invariant attention shares.

Under coercion, an external force smears the transition into a gradual crossover. Peak sensitivity drops to one-seventeenth of the invitation peak, with attention spread flatly across all temperatures. That seventeen-fold comparison belongs to the fixed simulation convention; it does not quantify how much less an institution knows about itself.

Monte Carlo simulations of the coercive Ising model (a standard 2D Ising lattice with asymmetric transition rates parameterized by a coercion fraction c) quantify the damage precisely.¹¹⁰¹ At L = 64 (a 64-by-64 grid of agents), the invitation baseline (c = 0) yields a peak magnetic susceptibility chi_max = 55.9, measuring the system’s capacity to reorganize in response to perturbation. Even 10% coercion (c = 0.10) collapses chi_max to 3.8: a fifteen-fold suppression. By c = 0.30, the suppression reaches thirty-seven-fold (chi_max = 1.5).

The shape of the curve matters most. The suppression is non-monotonic. Chi_max reaches its minimum at c = 0.20 to 0.30, then partially recovers at higher coercion: chi_max = 4.7 at c = 0.50, 3.8 at c = 0.70, 2.9 at c = 1.00. Pure coercion (the directed percolation regime) has its own phase transition with its own susceptibility peak. On five seeds, though, that rise lies within noise, and the L = 128 sweep does not reproduce the shape (its lowest resolved peak moves to c = 0.10), so the recovery remains a hint rather than a result.

The deepest suppression occurs in the mixed regime, where coordination is partially voluntary and partially mandated. If the recovery holds up, the implication is a thermodynamic argument for commitment. A system that is fully invitation-based has high adaptive capacity (chi_max = 55.9). A system that is fully coerced has low adaptive capacity yet at least possesses the directed percolation transition’s own reorganization dynamics (chi_max = 2.9). A system stuck between the two modes, half-mandated and half-voluntary, sits between two phase transitions while reaching neither (chi_max = 1.5). The instinct to “add some mandates” when trust-based coordination falters is then the intervention that destroys adaptive capacity most effectively. It pushes the system away from the Ising transition it needs while failing to reach the DP transition that might partially compensate.

At L = 128 the original coarse sweep reports an invitation baseline of chi_max = 82.9, below the roughly 188 predicted by Ising finite-size scaling (gamma/nu = 7/4) from the L = 64 value. Its coerced estimate, 29.9 ± 36.7 at c = 1.00, does not establish a positive mean distinct from zero. The historical Wolff summary reports a baseline of 234.7, with unresolved run/grid identity. In the retained later cohort, the baseline is 230.93 at L = 128; at c = 0.3, chi_max is 4.47 at L = 64 and 17.51 at L = 128, with no saved L = 256 suppression cohort. These results support finite suppression at the two retained sizes. The earlier chi < 2 and all-size suppression conclusions are withdrawn, and strengthening with size is unestablished.¹¹⁰² Whether larger institutions suffer proportionally more from mixed coordination regimes remains an organizational conjecture.

A finer-grained follow-up, Experiment A15v2, tested eleven coercion values, with thirteen combinations of coercion and lattice size. Coercion now entered as a weight p that blends ordinary Ising flips with those of the contact process, the textbook model of an infection that spreads only from an active neighbor. The follow-up resolved a question the original seven-point sweep could not answer: is the transition between universality classes a jump or a gradient?¹¹⁰³

The critical exponent beta, which governs how the order parameter (coordination level) grows near the phase transition, rises with coercion overall: 0.15 at p = 0, then 0.37, 0.32 and 0.42 at p = 0.1, 0.2 and 0.3, on up to 0.82 at p = 0.9. Two qualifications attach to that series.

The first is that the p = 0 anchor, 0.15, sits near the 2D Ising 0.125 but comes from an L = 32 lattice, because the L = 64 run at zero coercion returned no usable fit. The second is that through the mixed region the fitted uncertainty exceeds the estimate itself, which is also where the series dips instead of climbing. Above p = 0.4 the rise is monotone, the fitted uncertainty finally falls below the estimate, and no discontinuous jump appears; inside the mixed region the data are too noisy to exclude one. On the firmer part of the series, the fitted exponent changes without a measured jump; this alone does not establish a continuous family of universality classes.

The susceptibility tells a different story. Chi_peak collapses catastrophically: 149.6 at p = 0, to 81.1 at p = 0.1, to 57.9 at p = 0.2, to 1.0 at p = 0.3, to 0.38 at p = 0.4, to 0.07 at p = 0.9. (The p = 0 baseline of 149.6 here uses A15v2’s connected estimator, which removes the Z₂ phase-mixing artifact; it is the same invitation baseline as A15’s 55.9, measured under a different susceptibility definition, so the two numbers are not directly comparable.) A 2,000-fold collapse from invitation to near-full coercion, with the cliff face falling between the measured points at p = 0.2 and p = 0.3. At p = 0.2 the system has lost 61% of its measured susceptibility; by p = 0.3, 99%. These percentages compare the chosen susceptibility estimator, rather than every recovery mechanism.

These are L = 64 values. The finite-size scaling report (Experiment AS12, 3,960 conditions) interprets the apparent 25% threshold as a finite-size effect and proposes that nonzero coercion removes the critical transition in the thermodynamic limit. Establishing that extrapolation requires the matched finite-size data and the update-rule mapping to the Ising perturbation. A relevant perturbation grows under coarse-graining; its identification here remains part of that model-specific analysis. The finite-lattice response drop supplies no universal zero-coercion tolerance for institutions.

The contrast between the two measurements is itself the result. The fitted exponent changes across the sampled coercion values, with large uncertainties in the mixed region. The measured susceptibility falls sharply between finite-lattice samples. This distinguishes two responses without proving a continuous family of universality classes or the loss of every route to reorganization.

The uncertainty is especially large at p = 0.2 and p = 0.3: the fitted exponents are 0.32 ± 0.89 and 0.42 ± 1.04. These error bars come from bootstrapping the fitted rate-grid points. They do not measure a distribution of universality classes across seeds. The fit is too uncertain here to establish a stable exponent, while the susceptibility measurements separately describe the finite lattice’s response.

The lattice result suggests an organizational question: can mandates preserve ordinary output while suppressing responses to disruption? The finite-size drop near a quarter coercion belongs to the simulation parameter. Applying that threshold to an institution requires a mapping from mandates to its dynamics.

Figure 17a.1: Susceptibility under coercion, on a log scale, for two experiments at L = 64. A15 (gold, Metropolis dynamics, standard susceptibility) falls from 55.9 at zero coercion to 1.5 at c = 0.3, the 37-fold collapse; its apparent recovery beyond c = 0.3 (to 4.7 at c = 0.5) is drawn dashed inside a shaded band, because on five seeds that rise lies within noise. A15v2 (blue, the contact-process crossover, connected susceptibility, which removes the phase-mixing artifact) falls from 149.6 at p = 0 to 57.9 at p = 0.2, then to 1.0 at p = 0.3 and 0.066 at p = 0.9. The two definitions have different baselines; their magnitudes are not directly comparable. AS12 reports that the peak at nonzero coercion does not grow with lattice size across 3,960 conditions. The proposed disappearance of the divergence in the thermodynamic limit depends on that finite-size extrapolation.

A further experiment compared two recovery protocols: spontaneous recovery, after a perturbation without intervention, and seeded recovery, with periodic injections of cooperation.¹¹⁰⁴ It measured the final cooperator fraction relative to each arm’s own unperturbed baseline. At zero coercion, the reported spontaneous arm recovered more completely than the seeded arm.

The reported ratio of seeded to spontaneous recovery completeness rises from 0.81 at c = 0 through 2.67 at c = 0.50 to 4.90 at c = 0.70. This compares two protocols with separate baselines. At c = 1.0, the spontaneous arm returned no cooperators and the seeded arm ended near the injected fraction. The simulation injects cooperation every hundred sweeps, including the final sweep, so that final value alone cannot establish zero amplification or the failure of every possible rescue protocol.

The lattice protocol changes the balance between spontaneous and seeded recovery. An institutional analogy would be a team that redistributes work after a departure compared with an institution that needs outside help after a coordination failure. The analogy raises a question about dependence on rescue; the lattice ratio supplies no direct measurement of that dependence in a bureaucracy.

The practical warning is a hypothesis: mandates may replace some self-healing capacity with managed recovery. A system with no recovery route beyond rescue cannot recover after losing its rescuer. The research programme reports both critical-temperature and lower-temperature recovery conditions; comparing their matched results would establish how far the observed crossover extends beyond the critical protocol.

In plain terms, the map below sorts fifty-four simulated settings by how ordered each looked and whether cooperation, once seeded, survived; most of what follows is about how far that sorting can be read. A grid of 54 points, nine coercion values crossed with six control values at L = 64 across three seeds, brings these slices together.¹¹⁰⁵ The control is temperature T below p = 1 and spreading rate lambda at p = 1. The plotted statistic also changes: below p = 0.5 it is absolute magnetization, while at higher p it is cooperator density. Four plotting groups summarize these measurements; their boundaries do not by themselves identify universality classes.

The first plotting group has a reported statistic above 0.5 and no recorded absorption, or positive seeded survival where that protocol was measured. At low p this is the high-magnetization group. Its Binder cumulant (a ratio of the magnetization’s moments that tells ordered states from disordered ones) approaches 2/3 when the measured magnitude is nearly constant, distinct from the Ising critical fixed-point reference near 0.61. Seeded survival was not measured in the eighteen grid points at p ≤ 0.20. A high absolute magnetization can describe alignment of either sign; it does not establish spontaneous cooperative recovery from every seed.

The second plotting group has a reported statistic at or below 0.5, without recorded absorption and with positive survival where measured. At low coercion and high temperature, low absolute magnetization is consistent with thermal disorder. At high coercion the same numerical cutoff applies to cooperator density, a different observable. The shared color keeps the plot readable without making those measurements equivalent.

The gold points at p = 0.40, T = 0.8 and T = 1.2 were originally called frozen order. Their absolute magnetization exceeds 0.997 and their tested seed-survival probability is zero, but the retained cooperator densities are only 0.000414 and 0.001299. They are almost entirely defectors. The magnetization comes from the quasi-stationary sampling protocol, which starts from a random state, rather than from a demonstrated cooperative ordered start. These points therefore show defector-dominated quasi-stationary samples alongside failed seeded survival.

This distinction matters for the institutional analogy. Agreement of either sign raises absolute magnetization. A nearly unanimous defector state can therefore look highly ordered in that statistic. Calling it a smoothly functioning cooperative bureaucracy would reverse what the recorded density says.

The red plotting group records absorption or zero seeded survival with a low reported statistic. At high coercion it approaches the all-defector state. At p = 1.0, the first sampled active rate above the bracket is lambda = 1.649, where cooperator density is 0.161 and seeded survival is 0.147, rounded to 0.15. The reported lambda_c = 1.55 lies within the sampled bracket [1.5, 1.649]. This is a spreading-rate control, distinct from temperature.

The programme reports an Ising-side boundary from T_c = 2.31 at p = 0 to T_c = 1.25 at p = 0.30, and a survival boundary from T = 1.22 at p = 0.40 to lambda_c = 1.55 at p = 1.0. The plotted curves join grid crossings and bracket midpoints; they are not independent fitted-boundary measurements. The nearby endpoints differ by 0.10 in p and about 0.03 in the vertical control, giving an unrounded coordinate distance near 0.133. Such finite-grid proximity cannot settle whether a tricritical point exists, especially when the vertical control changes meaning at p = 1.

At p = 0.25 to 0.30 and T = 1.0, the retained Binder cumulants are approximately −456 and −226. These extreme ratios occur near absorption, where small second moments make the statistic unstable. A denser quasi-stationary search, AS6, reports 320 conditions without reproducing the negative values. That weakens the original first-order interpretation. Neither a finite grid nor a cumulant near 2/3 proves a smooth crossover everywhere or excludes every tricritical point.

The map summarizes finite measurements under specified update rules and sampling protocols. Its gold points combine defector-dominated quasi-stationary samples with failed seeded survival. Locating an institution on an analogous map would require its own state variables, observables and transition dynamics.

Figure 17a.2: Fifty-four grid points at L = 64 with three seeds each. The vertical control is temperature below p = 1 and spreading rate at p = 1. Green circles have a high reported statistic; blue squares have a low one, with no recorded absorption or with positive measured survival. Gold diamonds are defector-dominated quasi-stationary samples at p = 0.40 with zero tested seed survival. Red triangles record absorption or zero survival at a low statistic. The statistic is absolute magnetization below p = 0.5 and cooperator density from p = 0.5 upward. Dashed curves join grid crossings; vertical bars show the sampled brackets. Seeded survival was measured only for p > 0.20. AS12’s thermodynamic-limit extrapolation and AS6’s denser finite-grid search concern distinct questions; this plot proves neither a universal institutional threshold nor the absence of a tricritical point.

The Fisher and susceptibility results measure different responses under their simulation definitions. Converting them into institutional attention and capacity requires a separate observation and resource model. Their numerical association does not supply that bridge.

The Fisher information matrix, introduced above, measures local sensitivity to model parameters. It motivates a comparison with W. Ross Ashby’s requisite variety, which bounds regulation through the available disturbance, control, and outcome varieties. Passing that bound requires a specified regulation problem and tolerated error. The Fisher statistic alone establishes neither that invitation passes Ashby’s test nor that coercion fails it.

In Zhuravlev’s framework, the coerced system never aligns with its information landscape. Its deviation measure remains above 0.99 everywhere: this reports a large geometric deviation under the chosen definition and coordinates. The enforcer does not merely add noise; the operational statistic suggests a change in the modeled geometry, without measuring its correspondence to reality.

The simulation can motivate testing whether external forcing changes the information available for coordination. It does not measure intelligence or establish that coercion makes every system less capable of self-knowledge.

The operational deviation statistic is reported near 0.33 below the crossover and near 0.999 above it, a sharp change in finite samples that supplies neither a proof of discontinuity nor a general law of institutional collapse. Without a validated observation-model bridge, these numbers do not measure map accuracy, preference meaningfulness, or correspondence to reality.

The Ruppeiner metric describes the geometry of thermal fluctuations near phase transitions. Its scalar curvature can diverge negatively (Ruppeiner, 1995). This is intrinsic curvature of a fluctuation metric, and it motivates studying information geometry near the transition. Identifying a stable energy direction or energy-axis restoring force, an unstable or amplifying trust direction, a dynamical saddle, or a stable basin for alignment training requires a separate dynamical potential, a measured evolution law and its linearization; the sign of scalar curvature supplies none of those.

The condition number threshold also has a theoretical derivation. Zhuravlev’s Theorem 7.2 derives, from information geometry alone, that under an isotropic loss convergence time has an interior minimum if and only if the condition number exceeds 2. The physically more natural loss gives no such optimum.¹¹⁰⁶ Our Ising trust model crosses kappa = 2 at T = 2.204, within 2.87% of the critical temperature T_c = 2.269. This is a close numerical comparison, whose quantities require careful distinction. The trust dynamics do not follow Zhuravlev’s Model A convergence functional (R² < 0.06). Moreover, the simulation’s eigenvalue-ratio kappa differs from the theorem’s trace-ratio quantity. Their proximity is a numerical comparison, with no derived equivalence of thresholds or shared geometric cause.¹¹⁰⁷

The threshold is topology-dependent. In a sweep across network topologies (distinct from the single-lattice runs above), regular lattices cross kappa = 2 within 2.3% of T_c. Small-world and Erdős-Rényi random networks also cross (within 5.9% and 2.2% respectively). Scale-free networks never reach it: max kappa = 1.90.

The degree-heterogeneity sweep reports a decline in its operational kappa statistic, from approximately 3.3 on regular networks to below 2 around coefficient of variation 0.90. These values require a fixed metric and coordinate convention. The reported connectome values 0.99 and 0.41 cannot be condition numbers defined as largest divided by smallest positive Fisher eigenvalue, which are at least one. Those numerics are excluded from this comparison until their original normalization is recovered. The sweep does not establish a coordinate-independent social-trust boundary at kappa = 2.

The finite lattice shows a sharp suppression of its measured susceptibility near a quarter coercion. A thermodynamic-limit extension would concern the modeled critical response; it would not establish the loss of every mechanism for reorganization.

Trust sees; control is blind. The information-geometric content of the Trust Attractor is this: the critical region where a system holds both structure and sensitivity, with enough commitment to act and enough openness to notice when the ground shifts. Chapter 17b tests these predictions in the substrate where the physics is most directly measurable: transformer architectures and language model behavior.

Notes for this chapter