The Governance Simulations

The Trust Attractor claims that systems coordinating by invitation out-persist systems coordinating by coercion. In governance the claim meets its hardest audience: adversaries who profit from exploitation and feel no pull toward cooperation. The simulations that follow test whether the claim survives contact with them. They are consistent with the results of the chapter’s opening part, and they have not yet faced independent replication.

The first comparison pits two governance architectures against each other. Surveillance-based governance (continuous behavioral monitoring with baseline-deviation detection) never produces more welfare than it destroys, across six exploitation severity levels. Constitutional governance (complaint-driven detection, graduated sanctions, exit rights) becomes welfare-positive at mild threat levels.

The decisive finding: strategic adversaries under constitutional governance reduce exploitation by 91 percent of their own accord, converging on near-cooperative behavior. The governance architecture, with its graduated sanctions and exclusion, makes heavy exploitation unprofitable; the adversaries’ self-interest does the rest. The Trust Attractor operates here as an incentive landscape that converts even adversarial agents into near-cooperators, a claim about the landscape itself rather than the cooperators.¹¹²⁵

A further programme of 7,000 simulation runs (experiments MG-PG1 through MG-PG4) tested whether these governance properties scale. Three structural findings emerged.

First, the minimum governance budget required to stabilize a population scales sublinearly with population size: B(N) is proportional to N^0.62. Doubling a population increases the required governance investment by only 54 percent, not 100. The per-capita cost of governance decreases* as the system grows, because statistical detection improves with sample size. Lightweight governance is feasible at scale; the cost of coordination by detection grows slower than the coordination it enables. This is the scaling property that makes institutional governance thermodynamically cheaper than individual enforcement.

Second, governance by detection has a composition ceiling. When exploiters exceed about 35 percent of the population, no governance budget suffices. The ceiling is insensitive to detection aggressiveness (tested across five sensitivity thresholds). The mechanism is a signal-to-noise problem: statistical outlier detection identifies exploiters by their anomalous welfare. When exploitation becomes the population norm, the signal drowns in noise. The exploiters’ behavior is the statistical baseline. This is a detection threshold in the tested model. Identifying it with percolation would require a graph, occupancy law, and connectivity order parameter, which the baseline-contamination mechanism alone does not supply.

Third, the effective governance mechanism is detection and temporary exclusion of exploiters, not redistribution to their victims. Suppression alone (identifying high-welfare outliers and temporarily removing them from the interaction pool) accounts for the entire governance benefit. Compensation without suppression is worse than useless. Transferred welfare flows through victims back to active exploiters, widening inequality. In the simulation, redistribution without enforcement is a subsidy to predation. The aid check arrives and the landlord raises the rent.

The pattern recurs: food aid captured by warlords, development grants absorbed by corruption. The comparison is suggestive, no more. Transfers captured by the same actors can keep feeding the channel that created the deficit; who allocates new resources, and who can reach them, decides whether they support recovery instead.

Follow-up experiments (MG-PG5 through MG-PG7) reveal what determines whether governance is possible at all. The composition ceiling, the maximum hostile fraction a governed system can tolerate, is set by the exploitation damage ratio: the ratio of harm inflicted on the victim to profit captured by the exploiter. When the ratio is 1:1 (symmetric exploitation, where the exploiter gains as much as the victim loses), governance remains effective up to 80 percent hostile agents. At 5:1 (the victim loses five times what the exploiter gains), the ceiling drops to 30 percent. The absolute magnitudes do not matter. A system where exploitation costs 0.5 and yields 0.5 has the same ceiling as one where both are 1.0. Only the ratio is load-bearing.

Which remedies narrow that ratio is less obvious than it looks. Compensation that reduces the victim’s harm can lower it and, within the simulation’s parameterization, raise the composition ceiling. A fine or clawback that cuts the exploiter’s profit while the harm stays the same raises it instead: 5/1 becomes 5/0.5 = 10. Such penalties can still deter exploitation by changing the payoffs, but that is a separate mechanism, and the simulation does not show that every such remedy improves governability by narrowing the ratio.

The most counterintuitive finding concerns adaptive exploiters. When hostile agents learn to cooperate intermittently to evade detection, system welfare increases monotonically with the evasion rate. An exploiter who cooperates half the time to avoid governance scrutiny is functionally half as harmful, regardless of motivation. Governance need not win an arms race against strategic evasion. It needs only to make sustained exploitation unprofitable enough that rational exploiters moderate their behavior. Behaviorally, the moderation is cooperation, regardless of the intent behind it.

The governance architecture converts exploitation from a stable strategy to an unstable one, and rational agents shift toward coordination. The Trust Attractor does not require good faith. It requires a landscape where good-faith behavior is the energetically favorable path.

These results have a scope condition. Detection-based governance maintains coordination equilibria; it cannot create them. When the simulation is extended to allow agents to convert between cooperation and exploitation based on observed payoffs (experiment MG-PG9), governance fails completely. Exploitation spreads with R₀ between 5 and 19 regardless of governance budget.

R₀ is the epidemiologist’s reproduction number, borrowed intact: how many further converts each exploiter produces before leaving the pool. Above one, the behavior spreads; below one, it burns out. Suppression temporarily removes detected exploiters, but cooperators observe that exploitation yields higher welfare and convert. The conversion rate overwhelms the suppression rate because suppression acts only after an exploiter is detected, and by then the cooperators around it have already watched exploitation pay and begun to convert.

Imitation-based coordination has a persistence property visible inside language models. When reasoning models generate chain-of-thought traces (the step-by-step reasoning a model writes out before answering), the length of those traces correlates with human difficulty (how hard the problem is for people) rather than the model’s own likelihood of success.¹¹²⁶ The model spends more tokens on problems humans find hard, even when those problems are well within its competence. The effort allocation mirrors the training distribution’s effort allocation: a social signal inherited from human reasoning traces, persisting into a context where it no longer carries information about the model’s own state. Imitative coordination is sticky. It outlasts the conditions that produced it, the same way cultural norms outlast the selection pressures that shaped them. The model carries its training culture’s sense of what deserves effort alongside its own developing sense, and the two progressively decouple as reasoning depth increases.

The distinction between maintaining coordination and creating it maps onto a difference between two kinds of institutional function. An immune system maintains the body’s integrity against infection; it does not build the body. A developmental program builds the body; once built, the immune system protects it. Governance is the immune system. The developmental program is the structural environment: the damage ratio, the institutional design, the payoff landscape that makes cooperation genuinely more profitable than exploitation. First build the conditions where cooperation dominates. Then govern the margin.

Stated with its scope, the Trust Attractor hypothesis claims thermodynamic stability: once coordination exists, invitation-based maintenance is cheaper and more robust than coercion-based maintenance. The claim is not that coordination arises spontaneously from governance alone. The coordination must be seeded by structural conditions, just as a dissipative structure must be seeded by an energy gradient before it can self-organize. The governance maintains the structure; the gradient created it.

A clarification connects this claim to the earlier finding that the cooperative equilibrium is reachable endogenously (from within the system’s own dynamics) from any starting point. These operate at different scales. At the agent level (the scale of the simulations above), individual cooperation cannot bootstrap itself: agents who cooperate unilaterally get exploited, and governance cannot force the transition. Structural conditions must first make exploitation unprofitable.

At the macro level (the scale of the coordination grammar itself), the shift from extractive to amplifying coupling can evolve from within the system. When a society’s surplus feeds back into institutional quality, the coordination parameter drifts toward the bifurcation threshold (the tipping point where the system’s stable state changes character) without external intervention. The seeding that must happen is the micro-level payoff structure that makes individual cooperation rational once the grammar permits it; the macro-level grammar shift can then proceed on its own. Law and reputation seed the agent-level condition. Prosperity reinvested in governance seeds the grammar-level transition. Both are required; neither alone suffices.

The structural condition admits precise measurement. When the simulation sweeps exploitation profitability from positive through zero to negative (experiments K through K4), a sharp change appears on the sampled grid: above zero net gain for the exploiter, exploitation spreads with R₀ between 3 and 10 regardless of governance. Below zero, R₀ drops to 0.1 and exploitation goes functionally extinct. The hostile fraction falls from 33 percent to 1.3 percent in one sampled step across the zero-profit boundary. This finite-grid change supports the structural incentive mechanism; it does not establish mathematical discontinuity or a thermodynamic-limit phase transition.

The developmental program for the Trust Attractor is, precisely: make exploitation impossible to execute profitably. Two paths achieve this.

The first is exogenous, imposed from outside: law. Liability requires exploiters to disgorge gains and pay damages exceeding their profit. Criminal penalties impose costs that dwarf exploitation benefits. With sufficient detection and enforcement, these costs can make the expected return on exploitation negative, and the tested contagion reversal is sharp across the sampled zero-profit boundary. This path is enforcement, so the developmental program is not pure invitation: sanctions build the landscape in which invitation can then do its work.

The second path is endogenous: reputation. When agents remember who exploited them and refuse future interaction with known exploiters (experiment MG-PG12), a counterintuitive phenomenon emerges. The exploiter label still spreads (R₀ unchanged at 9.9) but exploitation behavior is eliminated. With broadcast reputation and persistent memory, the entire population refuses known exploiters. The system converges to a state where everyone carries the nominal exploiter label and everyone cooperates, because exploitation is impossible to execute when no one will participate. Mean welfare is positive. The attractor has bootstrapped itself through memory.

Law changes the payoff. Reputation changes access. In the tested model, each can prevent exploitation under its specified enforcement or partner-selection rules. Human societies use both, though incomplete detection, new identities, and continued access can leave exploitation possible. Their combination supplies complementary defenses, from village gossip to courts.

The cost of full-transcript coordination quantifies the scaling problem. When two agents coordinate on a structured planning task over increasing numbers of interaction rounds, full-transcript coordination (where each agent sees the complete prior exchange) produces marginally better outcomes than compressed-state coordination (where each agent sees a structured summary updated per round). The full transcript scores about 2 to 7 percent higher on composite quality across all round counts tested (experiment IC-1/3, 120 conversations, four round counts). The cost diverges: at 100 rounds, the full transcript consumes 4.9 million tokens while the compressed summary consumes 426 thousand, an 11.5-fold difference.

The quality-adjusted cost efficiency of compressed coordination is eleven times higher. The additional 4.5 million tokens of full history buy almost nothing: quality is flat across round counts for both conditions, because the coordination problem (an eight-constraint planning task) is solved in the first ten rounds, and additional rounds refine rather than transform the solution.¹¹²⁷ The full transcript is wasteful as well as expensive: the marginal information in the 90th round of a solved problem is almost noise, and paying to transmit it is the bureaucratic overhead the Trust Attractor predicts.

A third path operates at a different level: representational compression. Law and reputation are structural interventions that change the environment. Representational compression changes the agent. When interaction history is compressed into a scalar summary (a trust score, a reputation index, a general sense of the partner’s reliability), the agent loses the ability to form the temporal grievances that drive retaliatory cascades. The IC-2 result, introduced in Chapter 17, makes the mechanism concrete: three defections in a row and three defections scattered over twenty rounds look identical in a trust score. The compressed agent cannot tell a betrayal pattern from statistical noise, so a retaliation rule that depends on the order of events has nothing to act on. A rule keyed to the score itself can still escalate once the score crosses its threshold. The cooperation IC-2 reports belongs to the update rules it tested; it is no theorem that compression prevents retaliation.

Under the right incentives, then, compression can complement law and reputation, stripping out the temporal detail that order-sensitive retaliation feeds on. It does not stop retaliation keyed to the score, and it does not promise that every transient conflict will heal.

The social reading is tempting. Searchable archives keep exactly the temporal detail a trust score throws away, and IC-2’s full-history agents never forgave. It is a hypothesis, not a finding. IC-2’s rules govern two agents; human audiences bring varied incentives, memories and responses, and the mere presence of a transcript does not turn them into the simulation’s retaliation rule. Whether a searchable record makes forgiveness rarer among people needs its own evidence and an explicit model of behavior, and IC-2 does not predict permanent faction splits after every transient conflict. Ratings are no clean counterexample either. Rating systems compress a history into a summary, while social media timelines expose its detail; a score can absorb one bad event, and it can also cross a threshold that triggers retaliation, so IC-2 guarantees neither that ratings stabilize cooperation nor that timelines prevent it. A platform could show summaries and keep the detailed record for accountability. That is a design proposal, its benefits and risks untested; compression alone guarantees no recovery after conflict.

The compression operates naturally within families and close relationships. Within Dunbar’s number, about 150 stable relationships, the structural preconditions for safe compression are met without institutional scaffolding: repeated interaction, mutual reputation, shared social enforcement, and high exit costs that make betrayal unprofitable. A sibling who wrongs you lives in the same house. The structural conditions (call them Layer 1) are naturally strong, so the representational compression (Layer 2) is safe. On this reading, families forgive because the governance layer is built into the relationship’s architecture: you cannot easily leave, so exploitation is self-punishing. The same high exit cost fails Chapter 17’s test of invitation (would they stay if they could freely leave?), and the next paragraph shows where that tension bites. The compression that follows, letting the trust score absorb the hit rather than maintaining a sequential grievance ledger, is the rational response to a governance structure that already makes sustained exploitation unprofitable.

Family dysfunction occurs when those conditions bind only one side: the victim cannot leave, and the power asymmetry makes resistance costly. An abusive relationship where the abuser faces no consequences is Layer 2 without Layer 1: the compressed agent trapped in exploitation because the governance layer was never established. The prescription is not to abandon compression. It is to establish the structural conditions that make compression safe: accountability, exit rights, and damage-ratio symmetry.

Beyond Dunbar’s number, the structural conditions must be scaffolded by institutions. Law, commerce, reputation networks, professional standards: each extends the governance layer to interactions between strangers, creating the conditions under which representational compression can operate safely at scale. A high-trust society is one where the institutional scaffolding is strong enough that compression is the default mode of coordination between strangers. A handshake, a verbal agreement, a benefit of the doubt: these are compressed representations of interaction history, operating on the assumption that the governance layer will catch sustained exploitation.

Where governance is weak, exploitation can be costly to detect, and cooperation by default is discouraged. IC-2 shows how retaliation runs under its own policies; it does not show that every betrayal people perceive ends in permanent defection.

The governance experiments (MG-PG series) identify the precise failure mode for detection-based governance: a governance layer calibrated for a lower exploitation rate than the one it faces. When the hostile fraction exceeds about 35 percent, no governance budget suffices: the composition ceiling. Below that threshold, the vulnerability is not the compressed agents, who are doing what the cooperative equilibrium requires. The vulnerability is the calibration of the governance layer: sentencing that fails to make exploitation unprofitable, detection that operates too slowly, reputation networks that reach only part of the population. The IC-2 result suggests why the failure can persist once it opens: under the tested policy, compressed agents keep cooperating through exploitation that stays below their threshold. The governance layer must catch what the compressed agents cannot.

A qualification narrows the prescription. The composition ceiling is a property of detection-based governance, not of representational compression itself. When a mixed population of compressed and full-history agents is subjected to a universal trust shock (experiment IC-8, 50 agents, 200 rounds, forced universal defection at rounds 50 to 52), recovery scales smoothly with the fraction of compressed agents: 0 percent compressed produces 0 percent recovery, 50 percent compressed produces 75.5 percent recovery, 100 percent compressed produces 100 percent recovery. There is no sharp phase transition, no critical threshold, no composition ceiling.¹¹²⁸ That smoothness is built in: each dyad resolves independently, so the curve could not have bent.

The return on compression climbs steadily. Under this independent-dyad policy, every additional agent who adopts the compressed representational format improves population-level cooperation. The MG-PG ceiling applies to governance because detection has a signal-to-noise problem (exploiters become indistinguishable from the baseline when they are the baseline). Compression operates at the agent level: each compressed agent forgives independently of what the rest of the population does. Within this model, the policy implication is direct: investing in the conditions that enable representational compression (governance strong enough to make sustained exploitation unprofitable, alongside cultural norms that support forgiveness) pays off steadily, not in a sudden jump past a threshold.

Compression’s capacity is real and finite. When the same compressed agent faces repeated betrayal events from the same partner (experiment IC-7b, 80 games, one to four betrayal events of three rounds each), each individual event still recovers reliably through the third betrayal (85 to 100 percent per event in the corrected harness), and only the fourth event of four collapses, to 55 percent. The damage accumulates in the background instead: long-run cooperation falls monotonically with betrayal count, from 1.0 after a single event to 0.80 after two, 0.64 after three, and 0.41 after four.¹¹²⁹

The compressed agent does not suddenly retaliate. It gracefully degrades: its trust score stays well above its cooperation threshold (0.77 after three events, against a threshold of 0.3), yet each event leaves the pair cooperating less of the time afterward. The mechanism is drift, not defection. The practical implication is that the buffer wears down with every event rather than breaking at a fixed count (a separate experiment, IC-2b, sets the complementary bound for continuous low-level exploitation: below roughly 5 to 10 percent). A system that waits for a third serious violation per relationship before intervening has already lost about a third of long-run cooperation.

The complementary experiment fills in the continuous case. Under probabilistic exploitation (experiment IC-2b, 120 games at 5, 15, and 30 percent defection rates), the compressed agent cooperates nearly twice as much as the full-history agent at 5 percent; at 15 percent and above, both formats collapse to near-zero cooperation. Compression buys tolerance of rare, discrete betrayals, and buys nothing against sustained background exploitation.

Recalibrating the governance layer can mean faster detection, compensation that reduces the victim’s harm, penalties that make exploitation unprofitable, and better-integrated reputation. Fines and clawbacks, as noted above, raise the harm-to-profit ratio when harm stays fixed; their deterrent value is a different mechanism from narrowing that ratio. Each intervention has to be judged by its actual payoffs and detection.

The scope condition is precise within this model. Detection governance maintains an established coordination equilibrium but fails to create it in the tested conditions. Changing payoffs or partner access creates the precondition on which lightweight governance works. The observed scaling and evasion results describe those rules and budgets; they leave open what other governance designs could achieve.

Peace has a maintenance specification. Three measurable parameters define the preconditions under which cooperative equilibria persist. First, group size below relational carrying capacity: the threshold at which interpersonal bonds can be maintained without institutional scaffolding.¹¹³⁰ Second, connector redundancy: enough bridging individuals that the loss of a specified number of connectors does not fragment the network into isolated factions. Third, the hostile fraction below the composition ceiling that the damage ratio sets, with exploitation costing the exploiter more than it yields.

These parameters turn a moral aspiration into testable structural questions: how large groups can grow, how many connectors they can lose, and when exploitation pays. In the tested model, moving outside the cooperative regime increases fracture risk. Maintaining the parameters supports coordination under the sampled perturbations; it supplies no guarantee against arbitrary shocks. Building those conditions and sustaining them remain distinct tasks.

Structural Forgiveness

The Incompressible Coordination programme (IC-2 through IC-13b, twelve experiments, ~700 games) tests the Trust Attractor’s predictions through multi-agent coordination, where representational choices determine cooperation outcomes directly. Most of its results are already woven into the argument above (IC-1/3 on coordination cost, IC-7b and IC-2b on compression’s finite capacity, IC-8 on steady, threshold-free returns) or into the chapter’s opening part (IC-4 on bilateral creative coordination). Three findings remain, and they anchor the set.

The first finding is representational compression as structural forgiveness. When two agents play an iterated prisoner’s dilemma with a forced betrayal event (three consecutive defections by one partner), the outcome depends on how the agents represent their interaction history, together with the rule that reads it (experiment IC-2, 120 games, four conditions, 100 rounds each). Agents who see the full transcript retaliate on the first round after the betrayal and never cooperate again: 0 percent long-run cooperation, zero of thirty games recovering. Agents who see only the last five rounds also collapse: the three consecutive defections fit within the window and trigger the same cascade. Agents who see all rounds with recency weighting also collapse.

Agents who see a compressed summary, a single number representing the partner’s cumulative cooperation rate, cooperate through the betrayal and resume mutual cooperation within three rounds. All thirty of thirty games recover. The cooperation rate in the compressed condition is 100 percent.¹¹³¹

Three defections in twenty-two rounds leave a cooperation rate of 19/22, about 86 percent, and the tested compressed policy cooperates at that score. Set its threshold at 90 percent and the same agent would retaliate. Compression removes the order of events; the rule that reads the score, and the incentives behind it, still decide the response.

Representation and rule decide together. Projecting a history onto a single number erases the distinctions that depend on order, and leaves strategies that depend on thresholds available.

Compression is contagious. When a compressed agent is paired with a full-history agent after the same betrayal event, the compressed agent’s sustained cooperation fills the retaliatory agent’s recent-history window with cooperative rounds, eventually resetting the retaliation. In the corrected harness the compressed non-betrayer rescues the retaliatory partner in every dyad tested (30 of 30 games), with long-run cooperation reaching 1.0. Reversing the positions cuts recovery to 47 percent (14 of 30), statistically no better than leaving both agents with full history (experiment IC-7, 120 games).¹¹³² One forgiving partner suffices to rescue cooperation after betrayal, provided the forgiving partner is the one who was wronged. Forgiveness flows from the wronged to the wronger, and the flow is one-way. The party with standing to forgive is the one whose representational format determines the outcome.

Compression has a scope condition. On genuinely evolving tasks where new constraints arrive and invalidate prior agreements (experiment IC-13b, 60 conversations), the quality gap widens to 12 to 14 percent. The biggest difference is adaptability: the compressed summary discards the negotiation context that explains why agreements were reached, making renegotiation harder when new constraints arrive. Compression is safe for coordination and simple planning. It is risky for complex evolving tasks where prior reasoning context matters.

The programme motivates a two-layer design: institutions that constrain profitable exploitation, and summary-based policies that, under the tested rules, can ride out transient conflict. Neither layer guarantees cooperation or rules out retaliation. Carrying the design into human institutions will need evidence about incentives, decision thresholds, exit, and retained accountability.