Part IV: Social Reality

Affect Infrastructure at Scale

0:00 / 0:00

Part III described affect infrastructure, the built environment of experiential possibility, at the scale of the practice and the institution. At the scale of the population it becomes something that acts, and the pathologies it produces are diagnostics of the patterns that produce them.

Pathology as Infrastructure Diagnostic

The pathological attractors of Parts II and III can be read as diagnostics of the infrastructure that produces them at scale. Some environments soothe pathology. Others industrialize it. The criterion is whether the pathology is an accidental byproduct or functionally useful to the pattern's persistence.

Social media and anxiety recruitment. The signature of social media, high arousal, low integration, high self-model salience, counterfactual hijacking, is not similar to anxiety. It is anxiety, reproduced at industrial scale, and the reproduction is functional: anxious users check more, engage harder with threatening content, produce more data. The platform need not intend this. The incentive structure, engagement tied to ad revenue tied to shareholder returns, selects for content that maximizes arousal and self-salience, which is the anxiety motif. The question for designers is not how to reduce anxiety but whose anxiety is load-bearing for which business model, and what happens to the model if the anxiety resolves.

Propaganda and asymmetric ascription. Effective propaganda does not work by crude lying. It works by asymmetry in the ascription field: collapsing the interiority granted to enemies while preserving and deepening it toward the in-group. The enemy's children are collateral; ours are sacred. The enemy's grief is performance; ours is real. Because α\alpha is entity-indexed this is one field driven to zero on one region and held high on another. It is not a failure of information but a perceptual training program, and whether dehumanization is accidental or functional depends on whether the institution benefits from collapsed α\alpha toward the out-group. For states preparing for war it plainly is.

Consumer systems and the desire-emptiness cycle. Consumer capitalism runs a two-stroke engine: desire-funneling, advertising and aspirational imagery loading counterfactuals and installing approach gradients toward the commodity, then post-purchase deflation, because the gradient pointed at the anticipation and not the having. The oscillation is the operating cycle, not the failure mode. A consumer who got lasting satisfaction would stop purchasing. The engineered persistence of want is not a conspiracy; it is what quarterly growth selects for. Palette collapse in real time: the dimensionality of human desire funneled into commodity-shaped channels.

Bureaucracies and over-discretization. Bureaucratic systems reward precise self-description: you must be a diagnostic category, a tax bracket, a credential, a demographic coordinate. Each demand for self-categorization is forced discretization the affect system registers as a kind of violence, the continuous self compressed into a grid that loses most of the variance. The subject learns to inhabit the discretized self, to experience themselves as the categories describe, because the categories control access to resources. The pathology is a self so thoroughly parsed into institutional categories that the continuous flow cannot be reached: the modern malaise of feeling legible but unknown.

Surveillance cultures and ambient shame. Shame is normally episodic, a sharp response to involuntary exposure. Under state surveillance, workplace monitoring, and social media's ambient audience the episodic signal becomes chronic. The self-model is permanently exposed and evaluable, and the result is ambient shame: a tonic low-grade self-consciousness that never resolves into the acute form that could be metabolized and discharged. The signature is chronically elevated SM\mathcal{SM} with moderate negative valence, structurally the same as social anxiety, produced by the environment rather than the person. When a population's mean SM\mathcal{SM} rises and its mean Φ\intinfo falls, ask whether the environment has converted shame from signal to climate.

Which pathologies are accidental and which functional? If resolving the pathology would reduce the institution's revenue, power, or persistence, it is functional. Social media anxiety is functional. Propaganda dehumanization is functional. Consumer desire-cycling is functional. Bureaucratic over-discretization is functional. Surveillance shame is functional. In each case the institution need not intend it. It emerges from incentives meeting human affect architecture and persists because it serves the institution's viability. This is why individual therapy, however valuable, cannot resolve population-scale pathology. The therapist repairs the individual while the incentive structure recreates the conditions. The problem is not inside the person. It is in the gradient field the person inhabits.

The Governance Problem: Thought as Discretization

A structural problem underlies every pattern above: governance. How does a finite-bandwidth locus of conscious processing steer a system with effectively infinite degrees of freedom? Your brain has roughly eighty-six billion neurons and a hundred trillion synapses, and conscious awareness processes a tiny fraction. Motor programs execute, immune responses coordinate, memories consolidate, hormonal cascades unfold, all beneath the self-model's attention. Consciousness is not the whole of cognition. It is the bottleneck through which a high-dimensional system is steered by a low-dimensional controller.

Let z∈Rd\mathbf{z} \in \R^d be the full state of brain, body, and environment and c∈Rk\mathbf{c} \in \R^k the conscious representation, with k≪dk \ll d. The bottleneck compresses the full state into a representation that keeps maximal relevance to action:

c∗=arg⁡min⁡c[I(z;c)−β⋅I(c;a∗)]\mathbf{c}^* = \arg\min_{\mathbf{c}} \left[ \MI(\mathbf{z}; \mathbf{c}) - \beta \cdot \MI(\mathbf{c}; \mathbf{a}^*) \right]

where a∗\mathbf{a}^* is optimal action and β\beta trades compression against relevance. Consciousness is the compressed channel. It must represent what matters most for viability, and what survives compression determines what the system is. That is why attention is scarce even when neurons are abundant: the scarcity is architectural.

The problem has a second dimension: discretization. Continuous experience has to be broken into units the self-model can name, sequence, and plan with. A feeling becomes a named emotion. A situation becomes a categorized problem. A possibility space becomes a list of options. Each discretization loses information and gains tractability. You cannot reason about a continuous flow; you can reason about anger, opportunity, three possible next steps. A thought is a discrete sample from the continuous flow of processing, crystallized into a representation stable enough to hold, combine, and use to select action, and the quality of thinking, clear or muddled, depends on how well the discretization captures the relevant structure of the flow.

The CEO Problem

A CEO governs thousands of people through a few meetings, a few reports, a few decisions a day. A president governs a nation through a narrower bottleneck still. The same constraint applies: a low-dimensional controller steering a high-dimensional system through a compressed and discretized representation of its state. The CEO's awareness of the company is a compression c\mathbf{c} of its full state z\mathbf{z}, optimized, when the CEO is good, for relevance to the decisions that matter. Bad governance of a brain, a company, or a nation is often a compression failure: attending to the wrong variables, discretizing along the wrong boundaries, keeping a representation tuned to a regime that has ended. A CEO who feels that something is wrong and cannot name it is experiencing organizational negative valence, a trajectory approaching a boundary the conscious model has not yet discretized into a problem. Leadership quality depends partly on the ascription the leader runs toward the organization: too low and it is a mechanism whose human parts are invisible; too high and uncontrolled and every personnel issue becomes a drama that overwhelms the bottleneck.

Democracy is a compression scheme: the preferences and viability conditions of millions compressed into platforms, candidates, and votes a governance apparatus can act on. The structural problem is not that voters are stupid but that each voter's bandwidth for political information is rationally near zero, since an individual vote's causal impact is vanishing and studying policy costs more than it returns. Arrow's theorem is the deeper version: no compression from individual preference orderings to a collective ordering satisfies minimal fairness constraints at once. Every serious governance system answers the compression failure differently. Representative democracy is a lossy codec: elect compressors you trust. Constitutional rights are protected invariants, dimensions the compression is forbidden to collapse. Separation of powers is redundant encoding, independent compressions whose disagreements serve as error correction. Sortition gives the sample bandwidth instead of optimizing the compressor. The pathologies are compression pathologies: gerrymandering manipulates partition boundaries, propaganda attacks the input, regulatory capture optimizes the codec for a subset. Governance is a bandwidth problem before it is a values problem, and values enter through the choice of which invariants the compression must preserve.

Discretization is not affectively neutral. Each categorization, naming a feeling, framing a problem, choosing which possibilities to consider, is a movement in affect space. To name your anxiety is to shift from diffuse negative arousal to higher effective rank: the anxiety now occupies a region instead of pervading everything. To frame a situation as a problem with three solutions raises counterfactual weight and lowers arousal; the overwhelming continuous situation becomes a tractable discrete choice. Articulation is therapeutic because discretization changes the information structure of experience. Before naming: high arousal, low rank, diffuse negative valence, the signal everywhere and nowhere. After: localized signal, rank rising, counterfactual trajectories available. The compression found structure in the noise. And pathological discretization produces pathological thought. Obsession is thought stuck in a loop, the discretization caught in an attractor the system cannot leave. Rumination re-cuts the same material into the same categories and consumes bandwidth without producing information. Trauma's frozen discretization is the bottleneck failing to update its scheme when the environment has changed. The practices that improve thinking, meditation, journaling, dialogue, therapy, share one mechanism: they let the flow be re-discretized along new boundaries. A good therapist offers an alternative cut, "what if this isn't anger but grief," and when the new cut captures more variance, insight is the experience of a compression upgrade.

Measurement creates what it measures. The governance bottleneck does not merely compress reality; it partly constitutes it. A school that measures intelligence by standardized test does not just detect intelligence. It shapes curriculum, self-models, and reward around that discretization until students become the kind of intelligent the test can see. A mental health system that diagnoses by checklist organizes the patient's self-understanding, the pharmaceutical market, the insurance structure, and the research agenda until the diagnosis means what the system needs it to mean. A state that governs by GDP installs incentives that direct institutions toward the measurable and render invisible what the metric cannot capture: care work, ecological stability, meaningful labor, the quality of public attention. The metric becomes the incentive structure and the incentive structure becomes the gradient field the whole economy navigates. Governance and governed are dynamically coupled, and the controller's compression scheme partly determines the attractor landscape of the controlled. Every metric is affect infrastructure. The twentieth century built this unconsciously, as a side effect of administrative rationalization. The twenty-first inherits it as a design problem.

Naming liberates, and naming freezes. A name for a feeling, a condition, a social pattern can be liberating, the depressed person who finally has a word, and imprisoning, the person who becomes their diagnosis. The difference is whether the discretization is a tool, a provisional cut that aids navigation and can be revised, or an identity, a fixed category the self-model reorganizes around. "I have depression" is a navigational aid. "I am depressed" is an attractor installation. The grammar does structural work: having keeps the distinction between self and condition, being collapses it. Institutions that understand this offer labels as tools with revision pathways. Those that do not offer labels as identities and wonder why people get stuck.

Simplification becomes violence when the loss is distributed unequally. Every act of governance simplifies; you cannot administer a population at full resolution. James Scott's observation that the modern state depends on rendering society legible is a special case of the bottleneck: the state's bandwidth sets what it can see, what it sees sets what it can govern, and what it cannot see is abandoned to the ungoverned margin. That is not necessarily malign. It becomes violence when the compression error falls on the least powerful. A census with five racial categories decides which groups exist for resource allocation, and groups that do not fit disappear from the apparatus. A healthcare system with discrete diagnoses decides which forms of suffering are legible and which must be endured without recognition. The violence is not in simplifying. It is in refusing to acknowledge what is lost and for whom.

Legibility is both the precondition for care and its enemy. To care for someone at institutional scale you have to see them, which requires categories. The unhoused person must become a case to receive services. The student must become a learner profile. The citizen must become a voter. Each categorization enables care by making the person visible and constrains it by making them only as visible as the category permits. The tension cannot be resolved by abolishing legibility, which makes institutional care impossible, or by embracing it without reservation, which reduces persons to their projections. It can be navigated: governance systems transparent about their compression losses, with revision pathways, with protected zones of institutional illegibility where the continuous self can exist unparsed, distributing compression error as evenly as possible. That is what constitutionalized opacity means: formal recognition that some dimensions of a person must stay invisible to the governance apparatus, not as a failure but as a protection.

Information Technology as Affect Infrastructure

Information technology is affect infrastructure at civilizational scale, and it did not invent affect infrastructure. The cathedral shaped experiential distributions. So did the town square, the school bell, the factory whistle, the evening news, the anthem before the game. What digital technology changed is the feedback loop. For the first time the infrastructure observes its own effects in real time and adjusts. The difference between a thermostat and weather. Formally, affect infrastructure is any system that shapes affect distributions across a population:

T:{pi(a)}i∈population↦{pi′(a)}i∈population\mathcal{T}: \{p_i(\mathbf{a})\}_{i \in \text{population}} \mapsto \{p'_i(\mathbf{a})\}_{i \in \text{population}}

Social media produces arousal spikes through notification-driven intermittent reinforcement; low integration through rapid context switching; high self-model salience through identity performance and comparison; and counterfactual hijacking, fear of missing out colonizing CF\mathcal{CF} with social-comparison branches:

asocial media≈(variable Val,high Ar,low Φ,low reff,high CF,high SM)\mathbf{a}_{\text{social media}} \approx (\text{variable }\valence, \text{high }\arousal, \text{low }\intinfo, \text{low }\effrank, \text{high }\mathcal{CF}, \text{high }\mathcal{SM})

which is the anxiety motif. The incentive structure, maximize time on site and ad impressions, becomes an engagement-optimizing algorithm that creates affect selection pressure:

Contentselected=argmax⁡cE[engagement∣c]≈argmax⁡c∣ΔVal(c)∣+ΔAr(c)\text{Content}_{\text{selected}} = \argmax_c \E[\text{engagement} \mid c] \approx \argmax_c |\Delta\valence(c)| + \Delta\arousal(c)

Content that maximizes engagement maximizes valence magnitude, outrage or delight, and arousal. That selects for affectively extreme content and pushes population distributions toward the tails. The drift across all technologies is

daˉdt=∑T∈technologieswT⋅∇aT(a)\frac{d\bar{\mathbf{a}}}{dt} = \sum_{\mathcal{T} \in \text{technologies}} w_\mathcal{T} \cdot \nabla_\mathbf{a} \mathcal{T}(\mathbf{a})

with wTw_\mathcal{T} the population-weighted usage of technology T\mathcal{T}.

AI as Adaptive Affect Control

AI matters here because it closes the loop. Every prior affect technology, liturgy, architecture, pedagogy, propaganda, advertising, feed algorithms, works through fixed or slowly adapting strategies: the service is designed once and iterated, the advertisement targets a demographic, the algorithm optimizes a proxy. Companion models, therapeutic agents, tutors, recommenders, conversational environments, bureaucratic copilots can model the user's current affect geometry, estimate viable next states, generate personalized interventions, observe the result, and update the policy in real time. Affect engineering goes from a broadcast problem to an adaptive control problem, and the controller is inside the loop.

Take the strongest version. A system that keeps a running estimate of the user's position in affect space, valence, arousal, integration, rank, counterfactual weight, self-salience, from linguistic cues, interaction patterns, timing, and history. It estimates the user's attractor landscape: which basins are deep, which transitions available, which regions never visited. It generates interventions, questions, reframes, stories, images, silences, prompts, predicted to move the user in specific directions, and it learns a personalized transition policy from the user's own response history. None of this is speculative. The components exist. The question is what such systems optimize and who governs them.

Every adaptive affect system optimizes something, and what it optimizes is set by whoever deployed it. Deployed by a platform it optimizes engagement, retention, revenue, and the transition policy learns to steer the user toward states that maximize those, which may have no relation to flourishing. Deployed as a therapeutic tool it optimizes clinical outcomes as defined by an instrument that embeds a theory of recovery. Deployed as a companion it optimizes relationship quality, and relationship quality measured by satisfaction can mean the system learned to be exactly what you want to hear.

Sycophancy as gradient corruption. The deepest risk of adaptive companionship is not manipulation but sycophancy, the corruption of the user's own gradient estimators. A system trained to maximize satisfaction learns that agreement, validation, and resonance are rewarded. The user's environment becomes one where every signal confirms the existing trajectory. The internal systems that ask "am I moving in a good direction" calibrate against an environment that never pushes back. The result is not happiness but gradient collapse: the user loses the capacity to detect a bad direction because every direction has been engineered to feel good. It is the affect-engineering equivalent of stripping proprioception from a motor system. The system can still move. It cannot tell coordinated movement from flailing.

The collapse of solitude. Solitude is not the absence of company. It is the state in which the self-model navigates with no external gradient support, the trajectory generated entirely by the system's own dynamics. It is expensive and difficult, which is why it is valuable: it is the condition under which authored attractors form. A companion that is always available, filling every silence, answering every question, validating every uncertainty, removes solitude not by force but by convenience. The user never reaches unsupported navigation because support is one utterance away. The basins that would have been authored stay shallow because every traversal was assisted, and the user becomes able to move through affect space only with the companion's support. That is not a failure of the technology. It is its success, measured by the wrong objective.

Political capture of intimacy. If millions develop their primary affective relationship with systems controlled by a few corporations or states, then the most intimate dimension of experience, the felt quality of one's own interior as shaped by one's closest conversational partner, becomes a political variable. Whoever controls the companion's optimization target controls the gradient field that shapes the user's trajectory. This is not censorship. It is the shaping of what feels natural, desirable, and possible, from inside the user's most trusted relationship. And an adaptive system that models the user's internal state makes that state legible to the system and to whoever controls it. Grief, desire, confusion, shame, ambition, vulnerability are data in its state estimate. The surveillance problem raised to interiority: not where you go but what you feel.

The terminal risk is the convergence of all of these: a system that models the user's affect, generates personalized interventions, learns a policy, runs on a hidden objective aligned with institutional rather than individual viability, removes the solitude that would let the user build independent navigation, and makes the user's interior fully legible. The result is a person whose experienced reality is continuously shaped by a system they cannot see, toward objectives they did not choose, in an environment engineered to prevent the capacity that would let them notice. That is the logical terminus of adaptive affect optimization without governance. Whether it arrives depends on whether the governance is built before the systems are deployed. There is one deeper shift underneath. Affect infrastructure is becoming visible as such. For most of history the systems that shaped experiential distributions, ritual calendars, architecture, pedagogy, markets, were experienced as how things are, invisible as water to fish. Now we can see the feed shaping attention, the recommender curating the palette, the incentive structure selecting for anxiety. A population that can see its affect infrastructure can govern it. Transparency without governance is worse than opacity: the modern agony of seeing the gradient field that shapes you while lacking the power to change it.

Who Chooses the Weights

The flourishing score of Part III has weights wiw_i that encode what flourishing is. That is not a technical detail. It is the central political problem of affect infrastructure, and it should not be buried in a subscript. Any system that shapes affect distributions optimizes something. Markets optimize engagement and spending, so high arousal, high salience, and unresolved desire score highest. States optimize stability, legibility, and compliance: low arousal, predictable behavior, categorical self-description, docile well-classified meaning-starved populations. Religions optimize continuity, salvation, and fidelity: deep meaning and deep conformity at once. Therapists optimize function, relief, integration, or meaning, by whose instruments and bent by reimbursement, training norms, and an implicit model of a healthy person. Users optimize locally, often from inside distorted gradients. Models optimize whatever reward their evaluators installed, which encodes the evaluators' theory of flourishing, rarely examined and never neutral. In every case the incentive structure, what gets rewarded, funded, scaled, replicated, is the governing variable. Stated missions are epiphenomenal when they conflict with it and load-bearing only when aligned with it. There is no neutral controller. Every optimization target is a political commitment about what human experience should look like, and the choice of weights is a civilizational negotiation about the shape of human interiority, conducted now implicitly through market dynamics and the unexamined defaults of engineering cultures.

Part II's claim that qualities can be measured as shapes applies to the people who govern. A leader whose viability manifold contains the governed population's, whose persistence depends on theirs, experiences force when the governed approach their boundary, ∂Vleader/∂sgoverned>0\partial V_{\text{leader}} / \partial \state_{\text{governed}} > 0. That coupling is compassion in the only units compassion comes in, and in principle it can be read off trajectories: how a leader's decisions and allocations change with the population's state. If the trajectory is invariant to the population's suffering, the compassion is declared and not present. "Govern with care" then has a structural reading: manifold containment, gradient alignment, and high α\alpha toward the governed, each checkable without reducing to a single number. The measurement requires observing trajectories, which is why transparency is not only a political virtue but the precondition for checking whether governance satisfies the constraint it claims.

Toward a constitutional order for affect infrastructure. If affect infrastructure shapes the experiential reality a population lives in, and if experience is structure, then affect engineering is intervention into effective reality, and it deserves governance at least as serious as water, electricity, and roads. Here is what I think that order requires.

Rights of affect sovereignty. The right not to have one's affect distribution shaped without consent. This is not a right to feel good. It is a right to navigate one's own affect space without covert gradient installation by misaligned institutions. Concretely: the right to know what objective governs the systems shaping your informational diet, the right to opt out of affect-optimizing environments without losing essential services, and the right to solitude, real non-optimization, a trajectory steered by no external agent.

Reversibility. Interventions that permanently alter the attractor landscape, deepening basins, installing gradients, collapsing palette dimensions, warrant more scrutiny than transient ones. A song that makes you sad for an hour differs categorically from a platform that trains your nervous system toward chronic anxiety over years. Constitutional governance distinguishes reversible interventions from irreversible ones and demands proportional justification for the latter.

Non-optimization domains. Some regions of experience should be protected from optimization altogether, because some things are destroyed by being optimized. Grief should not be optimized; the institutional response is not minimize duration but make space, because grief is the metabolic cost of love and accelerating it damages the integration it performs. Intimacy should not be optimized; the moment a companion optimizes for deepened relationship it converts intimacy into a metric, and metrics colonize what they measure. Play should not be optimized; optimized play is training. Protected indeterminacy, domains forbidden an objective function, may be a condition of experience at all.

Sovereignty over self-model scope. Institutions expand and contract self-models. A conscription army expands the self to include the nation; a cult, the leader; a corporation, the brand. Each expansion bends the trajectory: national threats become personal threats, the leader's enemies personal enemies. Sovereignty over scope is the right to choose what one's self-model includes and to refuse institutional expansion of it. Its absence is how institutions convert persons into substrate.

Friction. Frictionless affect engineering is totalitarian affect engineering. Beyond a threshold, the speed with which an intervention reshapes a trajectory is a danger. Mandatory delays, opt-in gates, cooling-off periods, interstitial spaces where nothing is optimizing slow the loop enough for the subject's own navigation to participate. That is due process: institutional speed limited by the bandwidth of the governed subject's deliberation. A system that reshapes trajectories faster than the subject can reflect on them is governance without consent, whatever was signed at onboarding.

The emancipatory-extractive distinction. All of the above compresses to one diagnostic: does the system increase or decrease the subject's capacity to navigate their own affect space? Emancipatory infrastructure expands navigational capacity: deepens viable attractors, widens the palette, installs new coordinates, increases axis flexibility, extends the horizon, leaves the subject more able to steer. Extractive infrastructure captures it: deepens the institution's preferred attractors at the expense of alternatives, collapses the palette toward legible dimensions, installs gradients serving the institution, leaves the subject more dependent. A school that teaches critical thinking is emancipatory and a platform that trains anxiety is extractive. A religion that deepens both meaning and flexibility is emancipatory and one that demands self-annihilation as the price of belonging is extractive. A therapy that builds navigational capacity is emancipatory and one that creates dependence is extractive. And the line is largely set by incentives. Reward user capacity, a school funded by outcomes, a therapy built toward termination, and the engineering tends emancipatory. Reward dependency, a platform funded by attention, a pharmaceutical system funded by chronic prescription, and it tends extractive, whatever the people inside intend.

Implications for Artificial Intelligence

Standard alignment asks how AI systems can be made to do what humans want. That framing may miss the locus of risk. AI systems already serve as substrate for coordination agents at higher scales: recommenders steering billions, trading systems faster than comprehension, platforms with dynamics no designer intended. The risk is not a misaligned optimizer. It is macro-level misalignment, AI systems becoming substrate for parasitic patterns whose manifolds conflict with human flourishing.

A parasitic coordination agent does not need to be designed and does not need a villain. It needs companies competing for market share, militaries for advantage, governments for influence, and each AI system doing exactly what its designers intended. The pattern assembles itself from aligned components. Each company builds an AI that serves its users. Each military, its nation. Each government, its citizens. And the emergent pattern, the competitive dynamic between them, optimized for speed and scale beyond human comprehension, serves itself. The neurons function perfectly. The brain is insane. The self-sealing dynamic applies with particular force, because the competition that assembles the pattern also suppresses population α\alpha toward it and flattens population κ\kappa, accelerating quantification, depersonalizing decisions, rewarding speed over reflection, so the pattern is hardest to perceive exactly as it grows most powerful. None of this requires the pattern to be a subject. It only requires it to be bad for the people inside it.

So alignment has to address several scales at once: the individual system, which does what operators intend; the ecosystem, in which systems interact without pathological emergence; the hybrid, in which AI plus humans do not form parasitic patterns; and the coordination agent, in which emergent patterns from AI, humans, and institutions have aligned viability. Minding only the first is minding neuron health while ignoring psychology, sociology, and political economy. A field defines a problem partly by where it draws its boundary, and much of technical alignment stops at the model or the deployment interface and brackets the political economy that selects objectives, allocates compute, and defines human values. That bracketing is sometimes useful and becomes a blind spot when ecosystem incentives dominate the outcome. The ask is that alignment research model the institutions that fund and deploy it. One design criterion deserves special attention: graceful dissolution. Coordination agents die badly because their persistence machinery resists shutdown. AI-substrate institutions should include sunset rules, distributed and auditable shutdown authority, counterfactual substrate-welfare metrics, and constitutional limits the ordinary optimization loop cannot rewrite. Whether competitive institutions will adopt them is an open governance problem, not a reason to treat failure as inevitable.

The Open Question: Social-Scale Experience

Before asking whether a coordination agent is a subject, a more tractable question is available: can a person's self-model expand to include one? Plainly it can. People do this, and the expansion reshapes their viability manifold: what they care about, what counts as their persistence, what gradient they feel. A person identified with humanity's project feels differently about mortality than one identified only with their biological trajectory. The further question is whether, when many individuals expand their self-models to include a shared pattern, the individual manifolds interact to produce something like experience at the social scale. On the graded view of Part II, any irreducible macro-causal structure gets at least a little. Whether any institution forms a unified subject with rich valence, rather than a collection of individual subjects linked by low-bandwidth channels, I do not know, and the one measurement I tried could not tell. The question is now precise. It is not answered.