Abstract

Modern institutions routinely confuse the ability to connect large populations with the ability to govern them legitimately. Digital platforms can distribute messages to millions, aggregate votes at negligible cost, and automate enforcement. They cannot, by themselves, reproduce the contextual knowledge, repeated interaction, reciprocal obligation, and reputational accountability through which human groups sustain trust. This distinction is foundational for Commonsent: a federated coordination network designed to give individuals and communities organization-level capabilities without recreating the centralization, opacity, and capture risks of the corporate platform model.

This paper synthesizes research on Dunbar-style social-network constraints, social-brain and time-budget theories, collective intelligence, epistemic democracy, Ostromian commons governance, polycentric institutions, federated platforms, decentralized autonomous organizations, and AI-assisted coordination. It argues that Dunbar’s number should not be treated as a universal biological constant or as a mechanical cap of 150. Its importance is more general and more consequential: relationship-based accountability has finite cognitive, temporal, and emotional bandwidth; social networks therefore form nested layers; and governance costs rise nonlinearly when institutions demand that participants maintain more context than those layers permit.

The institutional response is not to abandon scale but to separate its functions. Commonsent anchors deliberation, stewardship, sanctioning, and legitimacy in bounded communities; it scales transactions, learning, standards, and evidence through interoperable protocols; and it reserves higher-order institutions for coordination among communities rather than command over them. The result is a proposed architecture of Dunbar-aware cells, nested federations, attenuated reputation, verifiable claims, subsidiarity, contestable AI assistance, and constitutional anti-capture controls. The paper concludes with design requirements and a falsifiable pilot research agenda.

Executive Summary

• The empirical premise. Human social networks repeatedly exhibit nested layers of decreasing intimacy, often approximating 5, 15, 50, and 150 active ties. Exact values vary, but the layered structure and maintenance cost of strong relationships are the durable findings.

• The systemic implication. Communication scales cheaply; mutual knowledge, trust, and accountable participation do not. Beyond human-scale contexts, informal governance gives way to formal rules, administrators, metrics, and intermediaries.

• The epistemic implication. Truth is not merely content. It is produced through a topology of witnesses, incentives, provenance, contestation, and correction. Large anonymous arenas weaken contextual accountability and amplify manipulation, conformity, status, and information cascades.

• The institutional implication. Durable governance should scale through nested and polycentric units. Local units retain meaningful autonomy; federations coordinate shared standards, externalities, appeals, and common infrastructure.

• The technical implication. A ledger cannot manufacture trust. It can make evidence portable, preserve provenance, encode commitments, and reduce the cost of verification across groups. AI cannot supply legitimacy. It can reduce cognitive burden, identify disagreement, model trade-offs, and improve institutional memory under human-governed constraints.

• The Commonsent design rule. Scale evidence, capabilities, and interoperability - not intimacy, undifferentiated voting populations, or global reputation scores.

1. The Scaling Problem Commonsent Is Designed to Solve

Commonsent begins from a structural diagnosis. Individuals are increasingly surrounded by institutions that possess organization-level intelligence: persistent memory, legal continuity, real-time data, predictive models, specialized labor, capital reserves, lobbying capacity, and automated decision systems. Most individuals and local communities possess none of these capabilities in coordinated form. The asymmetry is not merely informational. It is organizational.

The dominant digital response has been the platform. Platforms solve discovery, transaction, communication, and aggregation by centralizing identity, data, ranking, enforcement, and infrastructure. This produces enormous convenience, but it also converts the coordination problem into a control problem. The entity that owns the interface can shape visibility, define acceptable behavior, monetize behavioral traces, alter rules unilaterally, and exploit dependence. Scale is achieved by moving intelligence upward and agency inward, into the platform.

Commonsent proposes a different path: preserve local autonomy and ownership while federating capabilities. Communities should be able to deliberate, procure, invest, learn, negotiate, and build shared assets with the competence of an organization, yet without surrendering their data, identity, or political agency to a singular intermediary. The core challenge is therefore not simply decentralization. It is how to decentralize authority without fragmenting knowledge, and how to coordinate at scale without recreating centralized domination.

1.1 Communication scale is not governance scale

A network can add users faster than a community can add relationships. Broadcast, search, polling, and automated tabulation make mass participation appear nearly costless. Yet meaningful governance requires more than transmitting preferences. Participants must understand the issue, assess evidence, recognize affected interests, distinguish good-faith disagreement from manipulation, remember prior commitments, evaluate implementation, and revise decisions when conditions change. These are context-intensive tasks.

When institutional design ignores this difference, participation becomes thin. Citizens are asked to react to isolated proposals, token holders vote on technical matters they have not studied, and users click through rules they did not help form. Formal inclusion may rise while effective agency remains concentrated among agenda setters, professional administrators, large holders, highly motivated factions, or system designers. Commonsent treats this gap between nominal participation and substantive agency as a central failure mode.

2. What Dunbar Research Establishes - and What It Does Not

2.1 From the social-brain hypothesis to layered networks

Robin Dunbar’s original work linked primate group size to relative neocortex size and extrapolated a characteristic human group size near 150. Subsequent research shifted the more useful claim away from a single ceiling and toward a layered model of personal networks. Active networks commonly display nested circles near 5, 15, 50, and 150, with progressively weaker ties in larger layers. Longitudinal evidence suggests that the sizes of these layers can remain comparatively stable even as individual members enter and leave them. Digital social media expands reach and lowers the cost of contact, but has not demonstrated an unlimited expansion of emotionally meaningful or frequently maintained relationships.

The mechanisms are jointly cognitive and temporal. People must recognize individuals, represent relationships among them, infer intentions, recall histories, and allocate finite time to interaction. A larger address book does not remove these costs. Technologies may improve maintenance efficiency at the margin, but each relationship competes for attention, and stronger relationships require greater investment.

2.2 The controversy improves the design argument

Critics have correctly challenged the popular treatment of 150 as a precise, universal biological limit. Estimates vary by method, confidence intervals can be wide, cultural and ecological conditions matter, and the original primate-to-human extrapolation cannot carry the full weight often placed upon it. Commonsent should explicitly accept these criticisms.

The project does not require a fixed threshold. It requires only the more conservative propositions that strong-tie capacity is finite; human networks are uneven and layered; maintaining relationships consumes scarce cognitive and temporal resources; and governance tasks differ in the relational intensity they require. These propositions are considerably better supported than the claim that every viable community must contain exactly 150 members.

2.3 From number to function

The most important advance is to map social layers to governance functions rather than map one number to one institution. A five-person layer may support emergency trust, custody, or intensive execution. A group around fifteen may sustain a high-context operating circle. A group around fifty may support a working community in which participants can maintain meaningful familiarity. A group around 150 may sustain recognizable membership and reputational accountability, but may be too large for every member to deliberate deeply on every issue. Beyond that, direct relational governance must be supplemented by representation, delegation, specialization, formal procedure, or federation.

3. From Cognitive Constraint to Institutional Scaling Law

3.1 Relational bandwidth is a scarce system resource

Institutions are often modeled as rules plus incentives. A Dunbar-aware model adds relational bandwidth: the capacity of participants to maintain enough context about one another for informal accountability to work. This capacity is scarce, unevenly distributed, and consumed by governance. Every additional committee, conflict, proposal, and role imposes attention costs. When institutions exceed available relational bandwidth, they do not simply become slower. They change form.

Informal norms are replaced by written rules; mutual monitoring is replaced by audits; direct voice is replaced by representatives; contextual judgment is replaced by standardized categories; and reputation is replaced by metrics. These substitutions are not inherently bad. Large societies could not function without them. The danger is that they create new control points. Whoever writes the rules, owns the data, defines the metric, administers the appeal, or trains the model can acquire power that is difficult for ordinary participants to observe or contest.

3.2 Bureaucracy is an adaptation to lost context

Bureaucracy is frequently criticized as inefficiency, but it can be understood more precisely as a technology for coordinating people who do not know one another. Standardization allows decisions to travel without requiring shared history. Records allow institutions to remember beyond individual tenure. Roles reduce dependence on personal identity. The same mechanisms, however, can displace tacit knowledge, obscure responsibility, and privilege those fluent in formal procedure.

This creates a recurring scaling trap. A community grows; relational context thins; formalization expands; the system becomes harder for ordinary members to navigate; specialists gain influence; participation falls; and low participation is then used to justify further professionalization. Commonsent’s objective is not to eliminate formalization but to minimize unnecessary context loss and keep formal systems subordinate to locally authorized purposes.

3.3 The federation principle

If trust does not scale linearly, the system should scale by composing bounded units. A federation permits local units to retain decision authority over matters whose consequences are primarily local, while delegating limited functions upward where interdependence requires coordination. The higher layer should not be a larger copy of the lower layer. It should perform a narrower class of tasks: interoperability, common standards, cross-boundary externalities, shared infrastructure, appeals, and mutual defense against capture.

4. The Epistemology of Human-Scale Governance

4.1 Knowledge is produced by institutions, not merely stored in them

Governance is epistemic because every collective decision depends on claims about reality: what is happening, who is affected, what caused the problem, which interventions are feasible, what trade-offs are acceptable, and whether the intervention worked. Information systems often treat these as content-management questions. In practice, the reliability of a claim depends on provenance, incentives, expertise, local observation, adversarial scrutiny, and the ability to correct error.

Human-scale groups possess epistemic advantages that are difficult to reproduce centrally. Participants can observe repeated behavior, interpret local context, detect discrepancies between stated and actual conduct, and attach consequences to misrepresentation. Knowledge is socially situated: residents may know why a formal service fails, workers may know where process metrics mislead, and caregivers may know which burdens are invisible in administrative data. A centralized model can aggregate records while missing the context that gives those records meaning.

4.2 Small groups are not automatically wise

Locality is not a guarantee of truth. Small communities can be conformist, exclusionary, factional, nepotistic, or confidently wrong. Dense ties may suppress dissent because the social cost of disagreement is high. Shared experience may produce correlated errors rather than independent judgments. Therefore, Commonsent must not romanticize local knowledge. It must combine local grounding with procedures that preserve diversity, independence, evidence quality, and contestability.

Research on collective intelligence shows that group performance depends less on the maximum intelligence of individual members than on interaction patterns such as social sensitivity and relatively equal conversational participation. Other work demonstrates that social influence can improve or degrade collective judgment depending on network structure, diversity, confidence, and whether participants retain sufficiently independent information. The design target is therefore not consensus as such. It is an error-correcting process in which evidence, minority views, and uncertainty remain visible long enough to improve the decision.

4.3 Scale evidence, not intimacy

Commonsent’s epistemic architecture follows from a division of labor. Relational trust should remain primarily local, because it depends on context and repeated interaction. Evidence should become portable, because other communities need to evaluate claims without recreating the originating relationships. A portable evidence object may include source provenance, methods, data lineage, uncertainty, affected-party attestations, dissenting interpretations, audit history, implementation outcomes, and the authority under which it was produced.

This principle differs from a global reputation score. Reputation compresses a person or institution into a generalized ranking and often transports judgments beyond the context in which they were earned. Evidence portability transports claims and contributions with their context intact. Trust can then be attenuated across distance: a receiving community may recognize that an external claim was produced under a credible process without granting the originating actor unrestricted authority.

4.4 Institutional memory without institutional domination

A major weakness of purely local governance is memory loss. Volunteers leave, disputes recur, and successful practices remain tacit. Commonsent should treat institutional memory as a shared commons. Decisions should generate structured records: the question, evidence considered, value conflicts, decision rule, minority report, implementation owner, review date, and outcome metrics. AI can summarize and retrieve this memory, but it should not silently rewrite it. Source records, uncertainty, and dissent must remain accessible.

5. Ostrom, Polycentricity, and the Federation Response

5.1 Ostrom’s institutional answer to the scaling problem

Elinor Ostrom’s research challenged the assumption that common resources must be controlled either by markets or by centralized states. Long-enduring commons frequently rely on clearly understood boundaries, rules fitted to local conditions, meaningful participation in rule modification, monitoring accountable to users, graduated sanctions, accessible conflict resolution, recognition of self-organization, and nested enterprises for larger systems. The critical point is not that every commons uses the same rule. It is that governance is fitted to the resource, participants, and scale.

Polycentric governance extends this logic. Multiple decision centers possess meaningful autonomy while interacting under shared or overlapping rules. They may cooperate, compete, learn from one another, and resolve conflicts through higher-order arrangements. Polycentricity is not simple fragmentation. It requires interfaces, mutual adjustment, information exchange, and institutions capable of addressing externalities that cross local boundaries.

5.2 Subsidiarity as an information principle

Subsidiarity is often framed normatively: decisions should be made as close as possible to those affected. The Dunbar-Ostrom synthesis provides an information-theoretic reason. Local units generally possess richer contextual knowledge and stronger accountability, while higher units possess broader comparative information and greater capacity to manage cross-boundary effects. Authority should therefore sit at the lowest level that contains the consequences of the decision and has adequate capability to act.

This formulation avoids localism. A local unit should not control decisions whose harms are exported to others, whose benefits require larger coordination, or whose rights implications demand external protection. The correct level is determined by the topology of consequences, not by an ideological preference for either centralization or decentralization.

5.3 Federation as a learning system

The most powerful advantage of a polycentric network is parallel experimentation. Different communities can test policies under different conditions, compare results, and adopt successful practices without forcing the entire system into a single irreversible decision. Diversity becomes an epistemic asset. Failure can remain bounded; success can diffuse.

For this to work, experiments must be legible across communities. Commonsent needs common schemas for describing interventions, populations, resources, safeguards, metrics, and outcomes. AI can assist with causal comparison and transportability analysis, but communities must retain authority to decide whether a result fits local values and circumstances. Federation creates learning when it standardizes the description of evidence, not when it standardizes every local choice.

6. Why Platforms, DAOs, and AI Do Not Automatically Solve Scale

6.1 The platform fallacy: connectivity mistaken for community

Commercial platforms typically optimize engagement, retention, growth, and transaction volume. Their interfaces flatten social context because standardized interactions are easier to scale and monetize. A follower, rating, like, token balance, or account status becomes a proxy for relationships that are actually heterogeneous. The platform can coordinate enormous populations, but the coordination is usually organized around platform-defined objectives and enforced through centralized infrastructure.

The Dunbar implication is that a platform cannot turn millions of users into one accountable community. At best, it can host many communities. When it refuses to recognize that distinction, governance becomes a contest over a single global rule set, and the platform operator becomes the de facto sovereign.

6.2 The DAO fallacy: transparent execution mistaken for democratic legitimacy

DAOs improve the transparency and automation of certain commitments. Smart contracts can make treasury rules visible, execute votes, preserve records, and reduce reliance on a central administrator. Yet many DAOs reproduce familiar concentration through token-weighted voting, low participation, proposal complexity, delegation oligarchies, and core-team control. Transparency of procedure does not guarantee equality of voice, quality of deliberation, or legitimacy of agenda formation.

Recent research on blockchain governance increasingly emphasizes that decentralization is an institutional and democratic problem, not only a technical one. Commonsent should therefore treat ledgers and smart contracts as accountability infrastructure within a constitutional system, not as the constitution itself. Code executes authority; it does not justify authority.

6.3 The AI fallacy: cognitive assistance mistaken for political judgment

AI can reduce the burden that makes large-scale participation difficult. It can summarize evidence, translate language, identify recurring arguments, simulate scenarios, detect missing stakeholders, retrieve precedents, and help participants understand trade-offs. This is essential to Commonsent’s ambition to give citizens organization-level capabilities.

The same systems can centralize interpretation. A model may decide what information is salient, how uncertainty is framed, which options are compared, and which behaviors are flagged as anomalous. If the model, training data, prompts, or evaluation criteria are controlled remotely, AI assistance can become an invisible governance layer. Simulated consensus can also replace actual judgment: an AI can generate a plausible synthesis without demonstrating that participants understood, accepted, or authorized it.

6.4 AI changes the Dunbar constraint, but does not abolish it

AI may expand functional coordination capacity by externalizing memory and reducing information-processing costs. A participant can query past decisions, receive tailored explanations, and delegate bounded analytical tasks to a personal agent. This may allow a community to govern more issues or maintain weaker ties more effectively. It does not automatically expand mutual obligation or legitimate trust. Relationships mediated by agents still require decisions about whose agent represents whom, what information is disclosed, how errors are corrected, and who is responsible for action.

Accordingly, Commonsent should treat Dunbar thresholds as endogenous to tooling and task. AI may move the practical boundary for some functions, but the effect must be measured. The correct research question is not whether AI defeats Dunbar’s number. It is which forms of cognitive assistance increase participation and decision quality without reducing comprehension, responsibility, or human connection.

7. Commonsent Reference Architecture

7.1 Layer 1: the individual and the sovereign agent

The base unit is the person, supported by a user-controlled AI agent and personal data store. The agent represents preferences only within explicit mandates, preserves provenance for recommendations, and maintains a boundary between private context and shared evidence. It can prepare the user for deliberation, explain proposals, surface likely impacts, and manage attention. It should not cast unrestricted votes or infer durable political commitments from behavioral data.

7.2 Layer 2: human-scale cells

Cells are bounded groups organized around geography, function, resource, workplace, cooperative enterprise, or shared need. Membership and standing are intelligible. Cells hold authority over a defined domain and maintain their own rules within constitutional limits. They may contain smaller circles for execution and larger pools for oversight, sortition, or ratification. Cell size should be adaptive, with signals for fission or restructuring when participation, cohesion, or accountability deteriorates.

7.3 Layer 3: federations

Federations connect cells without absorbing them. Their authority is enumerated rather than presumed. Typical functions include shared technical standards, cross-cell procurement, interoperability, pooled insurance, regional infrastructure, externality management, portability of verified contributions, and appeals when local processes violate federation-level rights or procedures. Representation should be mixed: cell delegation, affected-party representation, sortition, expertise, and direct ratification may each be appropriate for different decisions.

7.4 Layer 4: protocol and evidence commons

The protocol layer provides identity credentials, permissions, data contracts, evidence schemas, treasury controls, audit logs, decision records, and interfaces between autonomous software modules. It should be open, forkable, and governed independently from any single application provider. No vendor should be able to revoke community access to identity, history, or assets.

7.5 Layer 5: constitutional safeguards

A federation needs limits that local majorities and technical operators cannot easily override. These include due process, privacy, freedom of exit, anti-discrimination, transparency of automated systems, rights to explanation and appeal, conflict-of-interest disclosure, limits on concentration, and procedures for constitutional amendment. Some safeguards protect individuals from communities; others protect communities from platforms, investors, and federation-level institutions.

7.6 Core protocol rules

1. Bounded authority. Every cell and federation body operates within an explicit domain; authority does not expand by technical convenience.

2. Subsidiarity with externality tests. Decisions remain local unless consequences, capability requirements, or rights protections justify escalation.

3. Attenuated trust. Credentials and reputation lose force as they travel across social and institutional distance unless supported by portable evidence.

4. Evidence before ranking. The system should expose reasons, provenance, and uncertainty before compressing them into scores.

5. Reversible delegation. Delegated power is specific, time-limited, inspectable, and revocable.

6. Contestable automation. Users can inspect, challenge, and bypass consequential AI recommendations or enforcement.

7. Forkability and exit. Communities can leave a service provider or federation while preserving identity, records, and lawful assets.

8. Plural metrics. No single engagement, reputation, welfare, or efficiency metric becomes the objective function for the entire network.

8. Deep Implications for Economics, Democracy, and Human Agency

8.1 The economic implication: coordination is a means of production

Firms dominate not only because they own capital but because they can coordinate specialization, information, contracts, and investment. Individuals purchasing separately remain demand; organized individuals can become a counterparty. Dunbar-aware federation allows communities to pool demand, procure collectively, create shared assets, negotiate market access, and retain value locally without requiring a single giant cooperative to manage every relationship.

This reframes Commonsent’s local recirculation model. The cell supplies trust and contextual knowledge; the federation supplies scale; the protocol supplies transaction and evidence infrastructure; AI supplies analytical capacity. Economic power is created by combining these layers. The result is not withdrawal from markets but a stronger demand-side institution capable of bargaining with firms and selectively insourcing production.

8.2 The democratic implication: representation becomes recursive

Mass democracy typically compresses citizens into periodic votes and delegates most complexity to parties, agencies, and interest groups. Direct digital democracy attempts to remove intermediaries but often overwhelms participants and advantages organized minorities. A Dunbar-aware system suggests recursive representation: individuals deliberate in bounded contexts; cells authorize delegates or agents for defined matters; federations coordinate among cells; and decisions return downward for review, ratification, or implementation.

Representation is therefore not a one-time transfer of sovereignty. It is a chain of inspectable mandates. AI can make the chain legible by showing what was delegated, how a representative acted, which evidence changed the decision, and where the outcome diverged from local preferences. This is a more demanding but potentially more accountable form of democracy than either mass plebiscites or unrestricted professional representation.

8.3 The epistemic implication: diversity must exist between as well as within communities

Collective intelligence requires diversity, but a single deliberative space often suppresses it through convergence, status effects, or shared framing. Federation preserves institutional diversity. Communities can maintain different norms, test different solutions, and produce independent evidence. Cross-cell comparison then becomes a source of error correction.

The architecture must nevertheless prevent epistemic segregation. Cells should not become sealed realities. Federation-level protocols should expose members to external evidence, minority reports, performance comparisons, and affected-party claims. The goal is bounded autonomy with structured permeability: enough closure to sustain identity and accountability, enough openness to learn and to protect outsiders from local harm.

8.4 The technological implication: personal AI should strengthen the social graph, not replace it

The likely evolution of AI interfaces - including ambient and augmented-reality systems - creates a choice. Personal agents could mediate nearly every encounter, optimize individual outcomes, and further privatize experience. Or they could help people perceive the social and material consequences of action: who produced a good, where value flows, which community commitments apply, what evidence supports a claim, and which neighbors are working on a related problem.

Commonsent’s exo-brain should therefore be designed as civic augmentation. It should reduce cognitive bias and information asymmetry while preserving encounters in which people recognize one another as agents rather than data points. The system succeeds when technology increases the capacity for reciprocal action, not merely the efficiency of personalized choice.

8.5 The anti-capture implication: power concentrates at translation layers

In federated systems, power does not disappear; it migrates. The most consequential actors may be those who translate between layers: identity issuers, schema designers, moderators, model providers, auditors, bridge operators, treasury custodians, and federation delegates. These positions can control what local knowledge becomes legible, which credentials travel, and which disputes receive attention.

Anti-capture design must therefore focus on interfaces as much as formal votes. Translation rules should be transparent, plural, and contestable. Critical functions should use separation of powers, rotation, random selection, independent audits, and public conflict-of-interest records. The system should measure concentration in agenda setting, delegation, code contributions, model usage, treasury access, and dispute resolution - not only token or vote ownership.

9. Risks, Boundary Conditions, and Anti-Capture Requirements

9.1 Local tyranny and exclusion

Human-scale accountability can become human-scale coercion. Dense communities may punish dissent, reproduce prejudice, or restrict exit. Commonsent requires federation-level rights, confidential reporting, external appeal, and the ability to participate through more than one affiliation. No person should be trapped under a single community’s social graph.

9.2 Cell fragmentation and coordination failure

Excessive autonomy can produce incompatible standards, duplicated effort, unresolved externalities, and weak bargaining power. Shared protocols, mutual recognition frameworks, and clearly enumerated federation functions are necessary. Federation must be strong enough to coordinate but too limited to absorb local sovereignty.

9.3 Sybil attacks, elite capture, and manufactured community

Digital systems make it cheap to simulate participation. Identity assurance should combine privacy-preserving credentials, local attestations, behavioral anomaly detection, and due process. Social anchoring can improve Sybil resistance but must avoid turning identity into a popularity contest. High-stakes decisions may require proof of affected status, residency, contribution, or randomly selected civic duty rather than unrestricted account creation.

9.4 Reputation overreach

Reputation systems can become durable instruments of exclusion. Commonsent should prefer domain-specific, time-bounded credentials and verifiable contributions over global personal scores. Negative judgments require notice, evidence, appeal, expiration, and limits on reuse. The system should distinguish trust in a claim, competence in a task, reliability in a role, and moral worth; these are not interchangeable.

9.5 Automation dependency

If participants cannot act without AI summaries, rankings, or recommendations, the network becomes dependent on model providers even when the infrastructure is nominally decentralized. Commonsent should support multiple models, local inference where feasible, auditable prompts and policies, human-readable fallbacks, and periodic exercises that test whether communities retain independent institutional competence.

10. Pilot Hypotheses and Measurement Framework

The Dunbar-Ostrom architecture should be treated as a set of testable hypotheses, not a branding narrative. A pilot in a city of similar population and institutional scale can evaluate whether bounded cells and federated evidence improve participation, decision quality, trust, and execution relative to conventional citywide consultation or open online forums.

10.1 Primary hypotheses

9. H1 - Nonlinear governance cost. As active cell size and issue load increase, participation equality, comprehension, and resolution speed will decline nonlinearly unless the cell restructures or delegates.

10. H2 - Relational grounding. Participants who have repeated interaction and recognizable standing will report greater procedural legitimacy and show higher follow-through than participants in anonymous, platform-wide processes.

11. H3 - Structured diversity. Cells using independent evidence collection, minority reports, and equalized turn-taking will outperform unstructured discussion on calibration, factual accuracy, and post-decision acceptance.

12. H4 - Federated learning. Communities provided with standardized evidence packages from peer cells will adopt useful innovations faster while retaining greater perceived autonomy than communities receiving centralized directives.

13. H5 - AI augmentation. Personalized explanation and retrieval will increase informed participation, but opaque AI ranking will reduce perceived agency and increase convergence around model-framed options.

14. H6 - Attenuated reputation. Domain-specific credentials with contextual evidence will produce fewer unfair spillovers and gaming incentives than a unified network-wide reputation score.

10.2 Measurement domains

10.3 Adaptive cell thresholds

Rather than impose a universal 150-member cap, the pilot should estimate a cell’s effective relational capacity. A composite stress index could combine active membership, interaction reciprocity, proposal load, concentration of speech, unresolved disputes, comprehension, and reliance on coordinators. When thresholds are crossed, the system can recommend one of several responses: divide the domain, create a temporary jury, delegate execution, add facilitation capacity, form sub-cells, or federate with a new unit.

This makes Dunbar’s contribution operational and falsifiable. The model predicts that different tasks will have different viable scales and that AI or institutional tooling may shift those scales. The system should learn where relational accountability begins to fail rather than assuming the answer in advance.

10.4 Minimum viable pilot

• Scope. Three to six neighborhood or issue-based cells, each responsible for a bounded decision such as local procurement, participatory grants, mobility, food resilience, or shared services.

• Comparison. At least one decision conducted through a conventional open forum and one through structured Dunbar-aware cells with federation-level synthesis.

• Evidence layer. A common decision record including sources, uncertainty, affected-party testimony, alternatives, minority report, mandate, owner, and review date.

• AI layer. Opt-in personal explanations, multilingual translation, argument mapping, precedent retrieval, and scenario simulation; no autonomous voting.

• Governance layer. Published authority boundaries, conflict-of-interest disclosures, appeal routes, data governance, and model accountability.

• Evaluation. Preregistered metrics, independent review, and public release of aggregate findings and protocol revisions.

11. Conclusion

Dunbar’s number matters to Commonsent not because a civilization can be reduced to 150-person villages, but because it reveals a basic mismatch between the scale of modern systems and the scale of human relational accountability. Institutions can centralize information, automate decisions, and aggregate preferences across millions. They cannot assume that those millions share context, reciprocal obligation, or the capacity to monitor one another. When systems ignore this fact, power flows toward those who control the substitutes for relationship: rules, metrics, interfaces, credentials, models, and administrative memory.

The appropriate response is neither nostalgic localism nor frictionless globalism. It is a layered institutional architecture. Human-scale communities generate context, legitimacy, and responsibility. Federations manage interdependence and enable collective power. Protocols make evidence and commitments portable. AI reduces cognitive burden while remaining contestable and subordinate to human mandates. Constitutional safeguards protect individuals, communities, and the federation from one another’s overreach.

This architecture gives Commonsent a systemic and epistemological foundation. Systemically, it explains why the network must scale through modular replication, nested authority, and explicit interfaces rather than through a single expanding platform. Epistemologically, it explains why claims must remain linked to provenance, procedure, uncertainty, and local knowledge rather than collapsing into centralized rankings or synthetic consensus. Economically, it shows how federated communities can acquire organization-level capacity while retaining ownership and agency.

References

1. Dunbar, R. I. M. (1992). Neocortex size as a constraint on group size in primates. Journal of Human Evolution, 22(6), 469-493.

2. Dunbar, R. I. M. (1993). Coevolution of neocortical size, group size and language in humans. Behavioral and Brain Sciences, 16(4), 681-735.

3. Dunbar, R. I. M. (2016). Do online social media cut through the constraints that limit the size of offline social networks? Royal Society Open Science, 3, 150292.

4. Lindenfors, P., Wartel, A., & Lind, J. (2021). Dunbar’s number deconstructed. Biology Letters, 17, 20210158.

5. Escribano, D., et al. (2023). Stability of personal relationship networks in a longitudinal study of middle school students. Scientific Reports, 13.

6. West, B. J., Turalska, M., & Grigolini, P. (2020). Relating size and functionality in human social networks through complexity. Proceedings of the National Academy of Sciences / related network analysis literature.

7. West, B. J., et al. (2023). Fractal structure of human and primate social networks. Proceedings of the Royal Society A, 479, 20230028.

8. Pearce, E., et al. (2014). Modelling the constraints on social network maintenance. Journal of the Royal Society Interface, 11.

9. Roberts, S. G. B., & Dunbar, R. I. M. (2015). Managing relationship decay: network, gender, and contextual effects. Human Nature, 26.

10. Woolley, A. W., et al. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004), 686-688.

11. Becker, J., Porter, E., & Centola, D. (2019). The wisdom of partisan crowds. Proceedings of the National Academy of Sciences, 116(22), 10717-10722.

12. Dietz, T. (2013). Bringing values and deliberation to science communication. Proceedings of the National Academy of Sciences, 110(Suppl. 3), 14081-14087.

13. Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.

14. Ostrom, E. (2005). Understanding Institutional Diversity. Princeton University Press.

15. Ostrom, E. (2010). Beyond markets and states: polycentric governance of complex economic systems. American Economic Review, 100(3), 641-672.

16. McGinnis, M. D. (Ed.). (1999). Polycentricity and Local Public Economies. University of Michigan Press.

17. Jordan, A., Huitema, D., van Asselt, H., & Forster, J. (Eds.). (2018). Governing Climate Change: Polycentricity in Action? Cambridge University Press.

18. Paniagua, P., & Rayamajhee, V. (2022). A polycentric approach for pandemic governance. Journal of Institutional Economics, 18(4), 537-552.

19. Lewis, P., & Aligica, P. D. (2024). The Ostroms on self-governance: the importance of cybernetics. Journal of Institutional Economics.

20. Scheufele, D. A., et al. (2021). What we know about effective public engagement on CRISPR and beyond. Proceedings of the National Academy of Sciences, 118.

21. Cosens, B., et al. (2021). Integrating law, governance, and science to manage complex problems. Proceedings of the National Academy of Sciences, 118.

22. Sharma, T., et al. (2026). Democratic governance through DAO-based deliberation for inclusive AI design. Scientific Reports.

23. Motea, M. (2026). Rethinking blockchain governance through democratic innovation and institutional legitimacy. Humanities and Social Sciences Communications.

24. Matias, J. N., et al. (2025). How public involvement can improve the science of AI. Proceedings of the National Academy of Sciences, 122.

25. Hwang, S., et al. (2025). Toward infrastructure for community-run social media: inter-community governance, modularity, forkability, and polycentricity. arXiv preprint.

26. Li, W., et al. (2025). Reflecting on Dunbar’s numbers: individual differences and implications for social network research. Review article.

27. Oh, P., et al. (2025). Self-organizing group structure through rewiring for collective decision-making. Scientific Reports.

28. Rodríguez-Barroso, N., et al. (2026). From privacy to trust in the agentic era: trustworthy federated learning as a continuously evidenced condition. arXiv preprint.

Selected Online Research Sources

Royal Society Open Science - Online social media and Dunbar constraints

Biology Letters - Dunbar’s number deconstructed

Royal Society A - Fractal structure of social networks

Ostrom Workshop - Polycentric governance resources

Science - Collective intelligence factor

Scientific Reports - DAO-based deliberation

Research note. This paper uses Dunbar-style network layers as probabilistic and task-dependent design heuristics. It does not claim that 150 is a universal biological maximum, nor that small communities are inherently legitimate or accurate. All institutional claims are intended for empirical testing and iterative revision.