Daemon Corporation

Global Leader in Excellence.

Research Report № 006

Adaptive Collective Systems: An Archipelago, Not Yet a Field

Executive synthesis

The intersection of adaptive coordination, collective intelligence, world models, institutional learning, and living systems is not empty, but neither is it an integrated science or practice.

What exists is an archipelago:

  • cognitive science, active inference, AI world-model research, and biological intelligence supply formal accounts of sensing, prediction, goal pursuit, and error correction;
  • cybernetics and complex-systems research supply a common language of feedback, regulation, emergence, and multiscale organization;
  • commons governance, adaptive management, experimentalist governance, and organizational learning supply institutions that monitor consequences and sometimes revise rules;
  • civic technology and deliberative democracy supply infrastructure for eliciting perspectives, identifying disagreement, and occasionally allocating public resources;
  • forecasting, site-reliability engineering, learning health systems, and living-evidence institutions supply mature bounded feedback loops;
  • new AI systems supply increasingly capable tools for synthesis, persistent memory, simulation, and mediation.

Several of these communities have built durable and effective components. None of the publicly evidenced cases examined here combines all of them.

The central gap is narrower—and more useful—than “no adaptive collective system exists.” The missing conjunction is:

An explicit, maintained, plural and contestable model of the world whose contestation can alter binding decisions, resources, goals, authority, or institutional structure, with consequences subsequently measured against the model and independently attributed.

The strongest cases fail in complementary ways:

  • Adaptive harvest management has maintained competing predictive models, updated their weights against outcomes, and used them to set binding annual regulations for roughly three decades. It deliberately keeps its objective function stable.
  • Ostromian commons institutions monitor, sanction, learn, and revise rules over very long periods, but their working models are embodied in practice rather than represented as explicit rival hypotheses.
  • The IETF versions, contests, revises, retires, and appeals shared specifications—including the rules governing revision itself—but its specifications are not competing predictive models and the institution does not allocate a public budget.
  • Experimentalist governance explicitly treats goals, metrics, and procedures as provisional and revisable, but its general outcome evidence remains incomplete.
  • Participatory budgeting connects public participation to real expenditure, but usually lacks explicit causal-model comparison and disciplined institutional learning.
  • The World Models interactive corpus represents disagreement unusually well, including fifteen explicit tensions and an edge type for “talking past,” but it is disconnected from decisions, execution, and accountability.

This mismatch divides the landscape into two broad halves:

  1. Model-rich but institution-poor systems: active inference, model-based reinforcement learning, basal cognition, world-model research, argument mapping, and AI deliberation.
  2. Institution-rich but model-poor systems: commons governance, adaptive governance, participatory budgeting, self-managing organizations, and organizational-learning practice.

The most important open problem is therefore not inventing another theory of adaptation. It is designing and evaluating an institution that can be wrong in public, preserve disagreement, act consequentially, learn from the result, and revise what it is trying to do without making evaluation impossible.

That last requirement creates a genuine design tension. Adaptive harvest management gives a defensible reason for holding goals stable: competing models cannot be evaluated against outcomes if the definition of success changes continuously. Yet institutional learning requires goals, decision rules, and authority eventually to become revisable. No investigated approach resolves this tension convincingly.

1. What would count as an integrated adaptive collective system?

The phrase can otherwise become so broad that almost any responsive organization qualifies. A falsifiable test needs to distinguish a pilot, an adaptive routine, and a genuinely integrated institution.

The following six-part standard is an analytical stipulation rather than an established benchmark:

  1. Recurrence: at least three completed cycles, documented contemporaneously.
  2. Explicit rival-model artifact: a maintained, inspectable representation in which distinct accounts of how the domain works coexist and remain identifiable.
  3. Allocative consequence: model state demonstrably affects a binding allocation of resources or authority.
  4. Consequence observation: outcomes are tested against the models’ predictions, not merely against participation or process metrics.
  5. Second-order revision: consequences have changed at least one goal, governing procedure, allocation rule, authority relation, or structural feature.
  6. External attribution: someone other than the operator has documented the causal connection.

No public case examined here satisfies all six.

Case Recurring cycles Explicit rival models Binding allocation Outcomes tested Second-order revision External evidence
Adaptive harvest management Yes Yes Yes Yes No—declined by design Yes
IETF Yes Partial: versioned specifications Partial: standards authority, no budget Partial Yes Yes
Ostromian commons institutions Yes No explicit artifact Yes Yes Yes Yes
Wikipedia Yes Partial Partial Partial Yes, endogenously Yes
Experimentalist governance Claimed in multiple settings No common model artifact Yes Yes Yes in the architecture Partial
Learning health systems Yes Partial Weak/variable Yes Weak Partial
World Models interface Not applicable Yes No No No No

This test changes the diagnosis. Second-order change is not universally absent: commons institutions, the IETF, Wikipedia, and experimentalist governance all reach it in some form. The rare combination is explicit rival models plus binding allocation plus revision of goals or authority.

2. The landscape: substantial components, weak integration

Fragmentation is not merely an impression. A bibliometric study of 39,334 collective-intelligence publications found that the field has struggled to benefit from adjacent disciplines that do not identify with the collective-intelligence label. It also found that AI–collective-intelligence work is growing quickly while becoming less topically and disciplinarily broad.1 A separate collective-adaptation program similarly argues that psychology, economics, sociology, biology, and physics repeatedly rediscover related principles in isolation.2

Landscape map

Community Lineage, scale, and assumed problem Typical beneficiary and method Maturity Characteristic omission
Cybernetics and organizational cybernetics Wiener, Ashby, Conant, Beer; regulator–system relations and organizations Managers and systems practitioners; formal theory, Viable System Model diagnosis, consulting Durable practitioner institutions; weak validation Longitudinal outcome evidence and comparative metrics
Active inference / free-energy principle Friston; agents and nested systems with generative models Cognitive scientists and AI researchers; variational formalism and simulation Durable field-building infrastructure Legitimate goal-setting, budgets, accountability, authority
AI world-model research Ha–Schmidhuber, model-based RL, JEPA-style prediction AI labs; architectures and benchmarks Functioning in silico infrastructure Contestation, legitimacy, institutional consequence
Collective and basal biological intelligence Levin, Watson, TAME; cells, tissues, organisms Biology, regenerative medicine, bioengineering; intervention and theory Empirically grounded at biological scale Rights, coercion, public accountability, negotiated goals
Complex adaptive systems Santa Fe Institute and related traditions; agents and emergent macropatterns Interdisciplinary research; simulation and mathematical modelling Large intellectual influence, weak disciplinary autonomy Binding institutional authority
Collective intelligence Woolley, Malone, Nesta; group performance and aggregation Researchers and platform designers; experiments and aggregation methods Repeatable methods, contested foundations Institutional structure, power, long-term adaptation
Collective adaptation Galesic and collaborators; trajectories of socio-cognitive systems Researchers; formal and empirical program Conceptual and early empirical agenda Deployed institutions
Commons governance Ostrom and successors; resource users and common-pool resources Communities and policymakers; comparative institutional analysis Durable institutions and strong case literature Explicit model artifacts; incomplete treatment of trust and power
Adaptive management and governance Holling, Walters, Folke; ecosystems and managing agencies Agencies and resource users; structured decisions, experiments, model updating Functioning domain infrastructure, uneven enactment Goal revision, distributional conflict, power
Experimentalist governance Sabel and Zeitlin; central frameworks plus local units Regulators and frontline institutions; provisional goals, reporting, peer review Developed architecture with claimed instantiations Independent comparative outcome evidence
Organizational learning and reliability Argyris and Schön, Weick, HRO, SRE Firms, teams, safety-critical operators; debriefs and postmortems Repeatable and sometimes mature bounded practice Public legitimacy and redistribution of authority
Deliberative democracy and civic technology Citizens’ assemblies, Pol.is, Decidim, participatory budgeting Citizens and public authorities; deliberation, clustering, voting Repeatable practice and functioning infrastructure Rival causal models, outcome attribution, institutional learning
Human–AI collective intelligence Human–machine teams, AI mediation, continual learning Organizations, AI labs, public platforms Demonstrated methods and emerging infrastructure Liability, legitimate goals, binding authority, long-term evidence

The divisions are structural. Fields capable of representing an explicit model generally treat goals and authority as external inputs. Fields with real authority and budgets generally embed their models in routines, rules, professional judgement, and political compromise rather than maintaining them as inspectable rival hypotheses.

3. The named starting cluster

The named starting points form a real, recently productive cluster around cognition, artificial intelligence, biology, and scientific synthesis. They should not, however, be mistaken for eight independent attempts to build adaptive institutions.

There are direct organizational links among them. AICACP’s media lead, Michael Garfield, identifies Atlas Research Group as his team. The founder of the Active Inference Institute appears on Atlas’s published roster. The World Models interface was built by Ideoscopic for the journal issue supported by the same collaboration. The apparent breadth is therefore partly a connected network of people and projects rather than independent convergence.

3.1 Adam Safron and AICACP

The AI Capabilities & Alignment Consensus Project is the most clearly substantiated organization in the cluster. It received a reported $270,000 Survival and Flourishing Fund grant. Adam Safron is identified as creator and principal investigator, with a named team including Victoria Klimaj, Michael Garfield, Colin Thomley, and Owen Lynch. Its mechanisms include journal collections, discussion-oriented workshops, and academic media intended to clarify world models and agency across AI capability and alignment camps.3

  • Actual problem: conceptual fragmentation in AI capability, alignment, agency, and world-model discourse.
  • Methods: synthesis, publishing, workshops, media, and convening.
  • Beneficiaries: researchers and participants in AI-alignment and policy debates.
  • Operational maturity: repeatable field-building practice.
  • Organizational model: grant-funded project with a named team.
  • Stated goal: bridge intellectual camps and build consensus.
  • Reasonable interpretation: agenda-setting and field formation.
  • Limit: no public evidence that its outputs have changed AI-development authority, regulation, resource allocation, or organizational structure.

AICACP is therefore substantive, but it is not an adaptive-governance system.

3.2 The 2026 Royal Society theme issue

World Models in Natural and Artificial Intelligence, published in Philosophical Transactions of the Royal Society A 384(2320), contains seventeen contributions plus one editorial.4 The editors explicitly decline to impose one definition of “world model,” instead including causal, self-referential, goal-directed, collective, and narrative forms.5

Across sixteen available publisher abstracts, the subjects include:

  • model-based reinforcement learning;
  • theory-based learning;
  • narrative world models;
  • goals and experience;
  • self-modelling neural systems;
  • LLM emergence and grounding;
  • unconventional embodiments and consciousness;
  • topological constraints on self-organization;
  • human–AI population simulations;
  • cybernetic regulation.

None of those sixteen abstracts addresses institutional direction-setting, budget allocation, coordinated implementation, accountability, or revision of authority. This is strong evidence of conceptual and operational distance, but it remains abstract-level evidence: most publisher full texts were not available for direct inspection.

Two contributions are especially relevant:

  1. Levin and Lyons describe the price system as a “cognitive glue” that allows actors to form mutually compatible plans and propose it as a generic template for collective coordination. This is an ambitious conceptual universality claim, not an institutional evaluation.6
  2. Alicea and colleagues recast the Conant–Ashby good-regulator theorem as a world-model result and extend it to second-order cybernetics. The paper also warns that tightly coupled regulation becomes problematic under non-uniform and out-of-distribution conditions—the conditions in which real institutions often operate.7
  • Actual problem: understanding what world models are and how they arise across biological and artificial systems.
  • Unit and scale: neural systems, AI models, organisms, cells, simulated populations, economies, and formal regulators.
  • Operational maturity: mostly conceptual, computational, and formal research.
  • Reasonable interpretation: a map of a contested intellectual territory, not an institutional design.
  • Limit: incompatible meanings and little connection to authority or implementation.

3.3 Active Inference Institute

The Active Inference Institute is a real 501(c)(3), with a board, officers, scientific advisers, educational cohorts, symposia, a journal, projects, and a substantial internal knowledge system. It reports more than 2,000 community members, nine completed textbook cohorts, six applied symposia from 2021 through 2026, twenty-eight institute-hosted projects, thirty-seven catalogued ecosystem projects, and 1,397 tracked records in “InstituteOS.” These figures are first-party and have not been independently audited.8

  • Actual problem: developing and disseminating active-inference research and practice.
  • Methods: education, publishing, mentoring, community formation, modelling projects.
  • Built systems: InstituteOS and GEO-INFER, among others.
  • Maturity: durable institution for education and field-building.
  • Funding model: not publicly detailed in the inspected materials.
  • Stated goal: support research, training, and application.
  • Reasonable interpretation: successful epistemic-community infrastructure.
  • Limit: no publicly evaluated deployment in which active inference repeatedly governed a real collective’s budgets, decisions, accountability, or structural revision.

The Institute demonstrates that a theory can acquire durable carriers without demonstrating that its general framework improves institutional adaptation.

3.4 Michael Levin’s collective and biological intelligence program

Levin’s research provides the named cluster’s strongest empirically grounded adaptive loop. Cells and tissues sense, coordinate across scales, pursue morphogenetic outcomes, and correct deviations through bioelectric and other signalling mechanisms.9

This is genuine prior art for distributed intelligence, but the scale transfer is not automatic. Watson and Levin themselves emphasize that collective intelligence is ambiguous, its mechanisms are often domain-specific, and collectives usually are neither evolutionary units nor singular loci of reward.10

A philosophical paper sometimes presented as a sceptical refutation of basal cognition is better read as a qualified defence. Fábregas-Tejeda and Sims argue that sweeping scepticism may be unwarranted while retaining evolutionary and methodological caution.11

  • Actual problem: explaining goal-directed regulation and problem-solving across biological scales.
  • Unit: cells, tissues, organisms, morphogenetic collectives.
  • Methods: biological experiment, intervention, and theoretical modelling.
  • Maturity: demonstrated empirical method at biological scales.
  • Institutional relevance: conceptual rather than operational.
  • Limit: cells do not negotiate rights, authorize budgets, appeal coercive decisions, or establish legitimate public authority.

The program should be taken seriously as biology and cautiously as institutional analogy.

3.5 Michael Garfield, Humans On The Loop, and Continual Learning

Humans On The Loop and the Continual Learning series translate the journal issue into public conversations about agency, embodiment, causality, human–machine co-evolution, and scientific collaboration. The confirmed series materials connect Safron, Levin, Ideoscopic, AICACP, and Atlas.12

  • Actual problem: cross-disciplinary interpretation and public sensemaking.
  • Methods: essays, interviews, podcasts, discourse maps, and public questions.
  • Maturity: repeatable media and convening practice.
  • Beneficiaries: researchers, practitioners, and interested publics.
  • Limit: no delegated authority, resource control, execution mechanism, or accountable feedback loop.

It is meaningful epistemic infrastructure, not collective-governance infrastructure.

3.6 Ideoscopic

Ideoscopic states a general method organized around four objects: work, span, collection, and reading. A question is put to each passage in a corpus, responses are arranged for comparison, and claims point back to source passages. Its central design commitment is that “nothing is summarised away.”13

  • Actual problem: seeing a body of work without losing source provenance.
  • Methods: passage-level retrieval, visual comparison, and corpus interrogation.
  • Maturity: demonstrated method with at least two demonstration corpora.
  • Beneficiary: people maintaining bodies of texts or knowledge.
  • Limit: no public operator, team, funding source, legal entity, user base, adoption measure, governance process, or client record was found.

The provenance commitment is important because institutional memory requires more than summaries. Yet there is no evidence that Ideoscopic’s representations remain synchronized with decisions and outcomes.

3.7 The World Models interactive corpus

The World Models interface is the cluster’s clearest functioning artifact for representing disagreement. It reports:

  • 18 issue items;
  • 217 concepts;
  • 28 themes;
  • 15 tensions;
  • 84 quoted positions.

It provides corpus-grounded question answering, a concept network, semantic maps, a tension compass, a concept matrix, weighted cross-references, and paper-specific views. Most importantly, it distinguishes genuine disagreement, complementary perspectives, and talking past.14

Its tensions include:

  • whether LLMs possess genuine world models;
  • how demanding the definition should be;
  • whether embodiment is necessary;
  • whether the self is enduring or dynamically extended;
  • whether descriptive and evaluative aspects can be separated;
  • whether intelligence is fundamentally collective;
  • whether one generic “cognitive glue” exists or coordination is substrate-specific.

The interface’s own AI-generated coding describes nine contributions as conceptual or position papers, three as computational modelling, two as empirical, two as formal theory, and one as review or synthesis. That coding should not be treated as publisher metadata.

  • Maturity: functioning infrastructure at demonstration scale.
  • Strength: unusually explicit representation of conceptual disagreement.
  • Limit: no evidence of reuse across domains, collaborative governance, outcome evaluation, or connection to binding decisions.

It is the landscape’s strongest example of a contestable model artifact and one of its clearest examples of a model artifact without institutional consequence.

3.8 Atlas Research Group

Atlas describes itself as a federated research group building “sovereign infrastructure” for collective intelligence, social coherence, and regenerative coordination. Its design language combines hierarchy, holarchy, heterarchy, and anarchy under the term “Synarchy.”15

Atlas is more substantial than a simple aspirational website. Its homepage contains a structured 42-person roster, with biographies, locations, affiliations, and controlled-vocabulary skill vectors across technical, epistemic, and human capabilities. Its interface includes a capability filter. In effect, Atlas has published a queryable model of its own distributed capabilities.

The roster situates Atlas socially between civic technology, systems practice, cybernetics, knowledge infrastructure, and regenerative design. Twenty roster members self-tag with “cybernetics”; “collective sensemaking” and “network stewardship and coordination” are also common tags. Five people list Atlas among their own organizational affiliations. Because there are no role titles, start dates, or commitment records, it is impossible to tell whether the full roster represents an operating collective or a looser network of affiliation.

Atlas’s public GitHub organization, created in November 2025, contained three small repositories with one contributor at the time inspected: guild-bot, regen-map, and rad-cms. Its fiscal sponsor, Flourishing Systems Foundation, is a verified but very new 501(c)(3), with an IRS ruling date in January 2025 and no filings yet on record.16

  • Actual problem: making distributed capabilities and complex systems visible and coordinable.
  • Built artifacts: a capability self-model and three early public repositories.
  • Organizational model: federated network with a fiscal sponsor and invitation process.
  • Maturity: bottom of demonstrated method, not functioning coordination infrastructure.
  • Stated goal: sovereign, life-aligned collective-intelligence infrastructure.
  • Reasonable interpretation: an emerging practitioner network maintaining a model of its own capabilities.
  • Limit: no disclosed deployments, clients, budget, outcome evidence, founding history, or decision governance.

Atlas is a concentrated instance of the landscape’s larger problem: it maintains a useful self-model, but no consequential use of that model is visible.

4. What the strongest operational cases actually integrate

The requested adaptive loop has ten functions:

  1. sensing and interpretation;
  2. contestable working models;
  3. direction-setting;
  4. decisions;
  5. resource allocation;
  6. coordinated execution;
  7. consequence observation;
  8. accountability;
  9. learning;
  10. revision of goals, mechanisms, authority, resources, or structure.

The strongest cases distribute these functions differently.

Legend: ● strong; ◐ partial or indirect; — not demonstrated.

System Sense Models Direction Decide Allocate Execute Observe Account Learn Revise institution
Adaptive harvest management ● ● rival ◐ fixed ● ● ● ● ● ● —
Ostromian commons institutions ● ◐ tacit ● ● ● ● ● ● ● ●
IETF ● ● specifications ● ● — ◐ ◐ ● appeals ● ●
Experimentalist governance ● — ● provisional ● ● ● ● ● ● ● architecturally
Wikipedia ● ◐ ● ● ◐ ● ◐ ● ● ●
Learning health systems ● ◐ ◐ ● ◐ ● ● ◐ ● ◐
SRE and postmortems ● — — ● ● backlog ● ● ● owners ● ◐ mechanisms
Participatory budgeting / Decidim ● — ● ● ● ● ● status ● ◐ ◐
Pol.is / vTaiwan ● — ● ◐ non-binding — — ◐ ◐ ◐ —
Habermas Machine / AI deliberation ● — — — — — — — — —
Buurtzorg ● local — ◐ ● ● scheduling ● ● ● ● ● authority
World Models interface ● corpus ● tensions — — — — — — — —
Active inference as formalism ● ● ● formal ● formal — — ● formal — ● formal —

No row is complete. The apparent integrations fall into six distinct types:

  1. Conceptual integration: active inference and cybernetic theory describe a unified loop formally.
  2. Computational integration: model-based RL and agent systems close perception–model–action loops in controlled environments.
  3. Biological integration: living systems perform multiscale regulation and repair.
  4. Organizational integration: double-loop learning, SRE, and self-management revise routines or authority.
  5. Governance integration: commons and experimentalist institutions connect participation, decisions, accountability, and rule revision.
  6. Operational integration: adaptive management and learning health systems repeatedly join data, decisions, interventions, and monitoring.

The evidence does not identify one winner. It identifies components that might be composed—but composition is itself unproven.

5. The strongest prior art

5.1 Adaptive harvest management: the best rival-model case

North American waterfowl adaptive harvest management began with four competing models, each weighted 0.25 in 1995. Their weights were subsequently updated according to predictive performance, and the results informed binding annual hunting regulations. The program describes those changing weights as learning about system dynamics and management response.17

It also reports that its original objective function and model set remained in place. That was not simply neglect. The program argues that objectives and model sets should change less frequently than annual regulations, while recognizing that periodic renewal is needed for political and stakeholder support.

This produces the landscape’s most important design tension:

Model comparison requires a stable criterion of success; institutional learning requires that the criterion eventually become revisable.

Frequent goal changes destroy comparability. Permanent goal stability freezes politics and value conflict outside the learning loop. No investigated institution provides a demonstrated solution.

5.2 IETF: versioning and appealing shared specifications

The IETF supplies unusually mature prior art for contestable shared artifacts. RFC 2026 requires a new version of an established Internet Standard to pass through the full standards process, provides for retirement to “Historic” status, and establishes an appeals ladder from working-group leadership through the IESG and IAB. RFC 2026 is itself Revision 3, meaning the revision procedure has governed its own revision.18

RFC 7282’s rough-consensus doctrine is also significant. It requires issues to be addressed, not necessarily accommodated, and does not reduce consensus to majority voting. It also records its own failure mode: participants may leave believing that “consensus” produced a collection of bad compromises.19

The IETF demonstrates:

  • versioned artifacts;
  • explicit objections;
  • procedural revision;
  • retirement;
  • formal appeal;
  • institutional durability.

It does not demonstrate rival predictive models tied to outcome forecasts or budget allocation.

5.3 Ostromian commons: adaptive capacity without a model artifact

Commons institutions provide the strongest evidence for durable collective monitoring, sanctions, conflict resolution, nested organization, and collective-choice rule revision. A review of 91 studies found Ostrom’s design principles well supported while accepting that they are incomplete.20

The omissions matter. Structural principles do not fully explain:

  • trust;
  • legitimacy;
  • transparency;
  • power;
  • biophysical context;
  • external markets and socioeconomic pressures.

Commons institutions show that an explicit model artifact is not necessary for adaptation. Their models can be embodied in rules, norms, memory, and repeated interaction. That is a warning against assuming that making a system formally representable will automatically make it more adaptive.

5.4 Experimentalist governance: the architecture most people assume does not exist

Experimentalist governance explicitly describes a recursive cycle of:

  • provisional framework goals;
  • local discretion;
  • reporting and comparison across units;
  • peer review;
  • periodic revision of goals, metrics, and decision procedures;
  • dynamic accountability and penalty defaults.

Its authors treat ends and means as corrigible and explicitly place goal revision inside the architecture.21 This is the clearest prior theory for the institutional loop under investigation.

It should not, however, be promoted from architecture to demonstrated superiority. Its authors acknowledge that generalization and inclusiveness remain unresolved, and the evidence base assembled here did not establish comparative longitudinal outcomes.

5.5 Wikipedia: institutional revision by selection rather than design

Wikipedia changed from an unreliable early project into a more reliable institution, but the mechanism documented in one American Political Science Review study is not deliberate double-loop redesign. Institutional losers became demotivated and exited; institutional winners remained.22

This matters because institutions can “learn” through membership selection without transparently revising their models. Such change may improve performance while suppressing or losing minority knowledge. Outcome improvement and deliberative legitimacy are not the same property.

5.6 Participatory budgeting and civic platforms

Participatory budgeting shows that collective input can govern real resources. Barcelona committed €30 million in both its 2020–2021 and 2024–2027 cycles, while Decidim provides proposals, voting, budget constraints, project selection, implementation status, public metrics, and open-data export.23

Its weakness is not consequence. Its weakness is model structure. Participants choose among proposals, but competing causal accounts, explicit predictions, uncertainty, and model revision are generally not preserved as first-class artifacts.

Pol.is and vTaiwan provide stronger opinion topology and facilitated consensus, with governmental participation and some policy effects. But current evidence indicates non-binding outputs, a narrow issue range, limited case volume, selective participation, and institutional decline after Audrey Tang’s departure. The often-cited government-action rate is implementer-sourced.24

5.7 Learning health systems, SRE, and bounded organizational loops

Learning health systems repeatedly connect clinical data, evidence generation, practice change, and outcomes. A 35-study review found positive implementation signals but comparatively weak governance and stakeholder engagement. A later scan of 44 systems found no evaluations of effectiveness in sustained systems.25

Site-reliability engineering has a stronger bounded process:

  • telemetry and incident records;
  • blameless causal reconstruction;
  • corrective and preventive actions;
  • named owners;
  • tracking tickets;
  • measurable completion criteria;
  • stored postmortems;
  • observation of recurrent incidents.

Google documents cases in which later recurrence had a smaller blast radius after corrective action, and meta-analysis supports debriefing as a performance-improving practice.26 But SRE usually takes product goals and organizational authority as given. It revises mechanisms, not public mandates.

5.8 Distributed authority without explicit models

Buurtzorg demonstrates that large organizations can distribute scheduling, hiring, care coordination, and operational authority to local teams. Dutch evidence reports high satisfaction and efficient home-care hours, although total system costs are closer to average once broader costs are counted. International transfer has faced incompatible regulation, hierarchical culture, IT constraints, and weak comparative evidence.27

Buurtzorg is prior art for authority redesign, not for explicit model comparison.

5.9 AI-mediated deliberation and human–AI collaboration

The Habermas Machine was tested with 5,734 UK participants. Participants preferred its group statements to those written by human mediators and rated them more informative, clear, and unbiased. Embedding analysis indicated that successful statements incorporated dissent while respecting majority positions. The result was replicated internally in a virtual citizens’ assembly with a demographically representative UK sample.28

This is strong evidence for synthesis assistance—not for downstream governance. It does not establish improved implementation, accountability, resource allocation, long-term legitimacy, or resistance to strategic manipulation. Critiques of minority compression and authority laundering should therefore be treated as design risks, not as demonstrated failures.29

The wider human–AI evidence is more positive than common summaries suggest. A preregistered meta-analysis of 106 studies and 370 effect sizes found:

  • human–AI systems performed slightly worse than the better of human or AI alone: g = −0.23;
  • human–AI systems substantially outperformed humans alone: g = +0.64;
  • heterogeneity was extreme: I2 = 97.7% and 93.8%, respectively.30

For institutions deciding whether to augment an existing human process, the second baseline is often the relevant one. The evidence supports augmentation under some conditions, not a general claim that combination produces superhuman collective intelligence.

6. Operationalization inventory

The landscape separates sharply when concepts, methods, practices, infrastructure, and institutions are not treated as equivalents.

Ladder

  • L1 — Conceptual framework: an articulated account without a demonstrated application.
  • L2 — Demonstrated method: at least one worked application, generally by its originators.
  • L3 — Repeatable practice: a documented procedure used by multiple operators.
  • L4 — Functioning infrastructure: a running system usable without its builders being present.
  • L5 — Durable institution: governance, succession, and persistence across multiple cycles or leadership changes.
Rung Representative examples Interpretation
L1: Conceptual framework Free-energy principle as a general theory; TAME at institutional scale; “cognitive glue” universality; collective adaptation; much of the Royal Society issue; Atlas’s larger ambitions Richest part of the named cluster
L2: Demonstrated method Levin’s bioelectric interventions; Habermas Machine; Voyager; Generative Agents; argument mapping; Atlas capability roster and repositories Demonstrated capability, not generalized institution
L3: Repeatable practice SRE postmortems; after-action review; participatory budgeting; citizens’ assemblies; forecasting tournaments; AICACP convening; AII cohorts; VSM consulting Procedures have moved beyond their originators
L4: Functioning infrastructure Decidim; Pol.is; Talk to the City; Global Dialogues; agent-memory runtimes; World Models interface; learning health infrastructure; adaptive harvest management Real users and recurring operations, generally domain-bounded
L5: Durable institution IETF; long-enduring commons institutions; Wikipedia; Cochrane; Active Inference Institute as a knowledge community; American Society for Cybernetics; Buurtzorg Durability does not imply full-loop integration

Three distinctions prevent category errors:

  • Active Inference Institute is L5 as a knowledge community, not as evidence that active inference has been validated for governing organizations.
  • The World Models interface is L4 at demonstration scale, not a durable institution.
  • Atlas is L1–L2, despite infrastructure language: it has a capability model and early repositories, but no demonstrated deployment.

7. Genuine overlap and terminological collision

The communities genuinely overlap around:

  • feedback;
  • prediction;
  • error correction;
  • constraint;
  • regulation;
  • coordination;
  • multiscale organization.

The Conant–Ashby good-regulator result is one of the few places where a contemporary “world model” and a cybernetic construct can be identified as the same formal object rather than compared metaphorically.7

Beyond that narrow core, collision is common.

Term Operational meanings encountered Assessment
World model Learned environment simulator; abstract predictive representation; pixel-level future generator; causal theory; narrative schema; cybernetic regulator homomorphism; price system; shared discourse map High collision. Meanings differ on whether a model is internal, explicit, predictive, causal, shared, or merely functionally implied.
Intelligence Benchmark performance; efficient problem-solving; group aggregation score; adaptive trajectory; organizational capacity; biological goal achievement Family resemblance, not a common measure
Agency Policy selection by an artificial agent; biological goal pursuit; professional discretion; institutional authority Collision unless scale and decision rights are specified
Learning Parameter update; belief revision; changed model weight; accumulated skill; changed routine; changed goal or authority structure Hierarchy of distinct changes, routinely conflated
Adaptation Better fit; reactive adjustment; changed action; model revision; structural reorganization Genuine overlap only when consequence-driven change is demonstrated
Cognition Representational processing; embodied regulation; basal competence; collective problem-solving Boundary remains actively disputed
Coordination Aggregation; consensus; division of labour; command; mutual adjustment; alignment on shared goals Different mechanisms, beneficiaries, and power relations
Living system Biochemical self-maintenance; multiscale biological organization; organizational metaphor; ethical commitment to “life alignment” Biological fact and normative metaphor should not be merged

The theme issue itself supplies unusually direct evidence of collision. Within one publication:

  • the editors decline a single definition;
  • one contribution states that all intelligence is collective because parts align to system-level goals;
  • another proposes the price system as a generic cognitive glue;
  • the accompanying interface encodes “talking past” as a relationship between contributions.5614

That is not failure. It is a useful representation of disagreement. But it should not be mistaken for theoretical convergence.

8. Failure modes any integrated design must survive

8.1 Shared models can amplify social pathologies

Within active-inference research, naive posterior-belief sharing has been shown to produce echo chambers and self-doubt. The same researchers propose an alternative sharing strategy that mitigates these effects.31 The result is a design constraint, not a refutation of shared models.

8.2 Coordination fails through specification and partner modelling

A taxonomy derived from 1,642 multi-agent traces identifies fourteen failure modes in three categories and reports benchmark failure rates from 41% to 86.7% across seven systems. The benchmarks differ and are not directly comparable. Depending on which figure in the paper is used, specification and inter-agent misalignment account for roughly 68.6% or 76.5% of failures—not the frequently repeated 79%.32

Separate work finds that LLM agents perform better when coordination depends on visible environmental variables than when they must actively model partners’ beliefs and intentions.33

8.3 Consensus can hide the politics of framing

A synthesized consensus does not reveal:

  • who chose the question;
  • who set the objective;
  • which options were excluded;
  • which minority models were compressed;
  • who has authority to act;
  • who bears the consequences.

Preserving dissent provenance, model lineage, and appeal rights is therefore different from maximizing endorsement.

8.4 Metrics alter the systems they measure

Goodhart dynamics and legibility problems constrain any machine-readable institutional model. Once a measure becomes a target, actors can optimize the proxy rather than the underlying purpose. Standardization also makes some knowledge visible while suppressing contextual and practical knowledge.34

This is not an argument against measurement. It is an argument for multiple measures, independent challenge, qualitative evidence, and periodic scrutiny of whether the metric still represents the intended goal.

8.5 Institutional memory is distributed and political

Research across Westminster systems attributes institutional amnesia to organizational churn, weak absorptive capacity, strategic decision-making, and historical storytelling. Related work distinguishes document-centred memory from dynamic memory dispersed among people and institutions.35

More memory is not automatically better. Archives can preserve obsolete assumptions as efficiently as useful lessons. A viable memory system needs:

  • provenance;
  • model and decision versioning;
  • explicit uncertainty;
  • challenge and appeal;
  • expiry or review conditions;
  • pathways from records into current decisions.

8.6 Adaptation can be too fast

Rapid adaptation can undermine due process, safety, minority protection, strategic continuity, and interpretability. Institutions operate across mismatched timescales: seconds for software, weeks for operations, annual budget cycles, multi-year laws, and decades for ecosystems.

No investigated infrastructure demonstrates a principled way to coordinate these timescales. A full system would need differentiated update rates, escalation rules, reversibility, and protected periods of stability.

8.7 Power remains undertheorized

Adaptive-governance scholarship has generated an internal critique that arrangements presented as flexible and collaborative often reflect existing power relations. Critical geography similarly warns that resilience language can obscure inequality or justify withdrawal of public responsibility.36

A system can adapt effectively for dominant beneficiaries while becoming less just for everyone else. “Adaptive for whom?” is not an external ethical add-on; it is part of the specification.

9. Territory classification

Well-developed

These areas have mature methods, evidence, and durable carriers:

  • bounded telemetry, incident response, postmortems, and debriefing;
  • forecasting, scoring, and calibrated belief updating;
  • commons governance as a comparative institutional science;
  • formal model-weight updating tied to recurring resource decisions;
  • standards procedures for versioning, retiring, contesting, and appealing specifications;
  • techniques for externalizing arguments, concepts, and disagreement.

Partially developed

Real practices exist, but important elements or evidence remain weak:

  • adaptive management and adaptive governance;
  • learning health systems;
  • participatory budgeting;
  • distributed organizational authority;
  • large-scale deliberative institutions;
  • organizational cybernetics and VSM practice.

Cybernetics illustrates the distinction between institutional survival and evidentiary maturity. The American Society for Cybernetics continues to run conferences, a speaker series, and a study group, while Metaphorum offers VSM training and certification. Yet a 2025 systematic review of twenty-one VSM studies found limited empirical validation, reliance on single cases, a lack of longitudinal studies, and few quantitative metrics.37

Fragmented

Substantial work exists but remains structurally disconnected:

  • collective sensemaking from decisions, execution, and institutional learning;
  • formal world models from legitimate institutional authority;
  • collective-intelligence research from commons and organizational-learning scholarship;
  • cognitive-science world-model communities from civic-technology communities;
  • cybernetic practitioners from comparative outcome research.

The predecessor histories are cautionary. Complexity science did not become an autonomous, self-reproducing discipline. Cybernetics fragmented as a discipline but survived in organizations, practices, and ordinary vocabulary.38 The relevant lesson is not that integration is impossible. It is that a broad vocabulary can spread while common methods, evidence standards, and institutions remain weak.

Emerging

  • AI-mediated deliberation;
  • continual-learning and agent-memory runtimes;
  • corpus-scale discourse maps with explicit tension layers;
  • longitudinal public-input datasets;
  • federated knowledge and capability mapping of the Atlas type;
  • active-inference applications seeking empirical benchmarking.

Genuinely underdeveloped

  1. Plural, versioned, live rival models connected to binding decision rights.
  2. Traceable links from deliberation to authority, budget, execution, outcome, and revision.
  3. A solution to the stable-objective versus revisable-goal tension.
  4. Accountability and appeal for AI-maintained institutional models.
  5. Coordination across technical, budgetary, legal, and ecological timescales.
  6. Rate limits and reversibility for institutional adaptation.
  7. Comparative evaluation of whole adaptive loops.
  8. Processes that preserve minority models without preventing coordinated action.

Impossible to assess from public evidence

  • proprietary enterprise planning and knowledge systems;
  • consulting deployments;
  • military and intelligence command infrastructure;
  • confidential platform-governance systems;
  • unpublished clinical-governance tools;
  • any non-public Atlas or Ideoscopic client work.

These inaccessible systems are the largest threat to the broad negative conclusion.

10. Design implications

The evidence supports design principles, not a proven integrated architecture.

10.1 Treat models, goals, and institutions as different versioned objects

A working system should not store “the model” as one undifferentiated knowledge graph. At minimum, it would separately version:

  • observations and data;
  • interpretations;
  • causal or predictive models;
  • uncertainties and dissenting models;
  • goals and values;
  • metrics;
  • decision rules;
  • authority assignments;
  • resource commitments;
  • actions;
  • observed consequences;
  • appeals and review findings.

This separation would make it possible to identify whether a disappointing outcome came from a bad prediction, a bad objective, poor execution, inadequate authority, or an invalid measure.

10.2 Preserve model plurality through decisions

Most tools either aggregate to one answer or maintain an archive with no consequence. A stronger architecture would preserve:

  • model identity and lineage;
  • explicit predictions;
  • minority and adversarial models;
  • model weights or confidence;
  • reasons for choosing an action despite disagreement;
  • conditions under which a losing model should be reconsidered.

IETF-style objection and appeal procedures are relevant here, even though IETF specifications are not predictive models.

10.3 Separate operational learning from constitutional learning

The adaptive-harvest-management tension suggests two timescales:

  • operational epochs, during which objectives remain stable enough to compare model predictions and outcomes;
  • constitutional reviews, at less frequent and predeclared intervals, when goals, metrics, authority, and institutional structure become contestable.

Versioning objectives alongside models could preserve evaluability within each epoch while allowing periodic second-order change. This is a plausible design possibility, not a demonstrated solution.

10.4 Make the action trace a first-class public artifact

The missing infrastructure is not only a better deliberation tool. It is a trace connecting:

signal → interpretation → model version → decision → authority → funds → action → outcome → review → revision.

A durable trace would allow participants and external evaluators to see whether deliberation affected anything, whether predictions were tested, and whether accountability reached the people with authority.

10.5 Build appeal, reversibility, and expiry into adaptation

Institutional learning should not mean continuous automatic optimization. Binding changes need:

  • reversible pilots where feasible;
  • escalation thresholds;
  • time-bounded delegations;
  • sunset and review clauses;
  • protected appeal routes;
  • slower update rates for rights and constitutional rules than for operational routines.

10.6 Evaluate augmentation against the real alternative

Human–AI systems should be compared not only with the best conceivable human or AI performer but with the human institution they would replace. The meta-analysis indicates meaningful average augmentation over humans alone alongside failure to beat the better solo performer and extreme variation across tasks.30

This argues for domain-specific evaluation rather than generic claims about human–AI collective intelligence.

A plausible bounded experiment

A useful next experiment would not attempt to redesign an entire government. It would choose a recurring, reversible decision domain with measurable consequences—such as a municipal maintenance budget, ecosystem intervention, clinical workflow, or infrastructure-reliability portfolio—and implement:

  1. two or more explicit rival models;
  2. registered predictions and uncertainty;
  3. a stable objective for a fixed evaluation epoch;
  4. limited binding authority over a real budget;
  5. public decision and dissent records;
  6. outcome measurement against predictions;
  7. independent evaluation;
  8. an appeal process;
  9. a scheduled constitutional review at which objectives, metrics, and authority can change.

Such a pilot would test the underdeveloped conjunction directly. Its purpose would not be to prove a universal theory, but to discover where plural modelling, legitimate authority, and institutional revision interfere with one another.

11. Forward radar: the next one to three years

High probability

  • AI summarization and clustering becoming routine components of civic-participation systems.
  • Versioned memory and provenance becoming standard features of organizational agents.
  • Discourse maps exposing tension, source lineage, and competing interpretations.
  • Longitudinal public-input datasets used for evaluation rather than one-off consultation.
  • Continued growth of domain-specific human-supervised learning loops in health, reliability, and resource management.

Medium probability

  • Direct links from AI-supported deliberation to procurement, participatory budgets, or formal model constitutions.
  • Independent replication and adversarial testing of AI-mediated deliberation.
  • Empirical benchmarking of active-inference applications against alternatives.
  • Better longitudinal evidence for organizational cybernetics and VSM practice.
  • Decision-impact studies connecting forecasting performance to actual allocations and policies.

Low probability, high impact

  • A public institution that maintains rival AI-assisted models and gives model contestation traceable authority over decisions and resources.
  • A credible implementation of objective “epochs” that permits periodic goal revision without destroying model evaluability.
  • Atlas moving from a capability self-model to a documented consequential deployment.
  • The World Models interface method being generalized across multiple consequential domains and connected to decision processes.

Signals to watch

  • reuse of Ideoscopic’s method on second and third corpora;
  • publication of Atlas deployments, clients, governance rules, or outcomes;
  • independent replication of the Habermas Machine;
  • sustained-effectiveness evaluation of a learning health system;
  • event-level evidence that experimentalist-governance cycles changed goals or authority;
  • a public dataset linking model versions, decisions, funds, actions, outcomes, and redesign.

12. Important uncertainties and weak evidence

Several limitations materially bound the conclusions.

  1. Theme-issue access was incomplete. Sixteen verbatim abstracts and one preprint full text were available; most publisher full texts were not. Claims about the issue’s overall subject matter are well supported, but claims about what individual papers omit remain abstract-level.
  2. The interface classifies its own corpus. Its contribution-type counts are an AI-generated layer and should not be treated as publisher metadata.
  3. Atlas’s roster is interpretable only as a published affiliation and capability model. It does not establish employment, active participation, or operating commitment.
  4. Ideoscopic discloses little organizationally. Its method is visible; its operator, financing, adoption, and client base are not.
  5. Experimentalist governance has a stronger architecture than gathered outcome evidence.
  6. Wikipedia’s mechanism rests here on one peer-reviewed source inspected at abstract level.
  7. Global South and Indigenous governance coverage is inadequate. One review of African and South Asian adaptive-governance scholarship and a limited set of cases cannot represent that territory.39
  8. Open-source foundation governance was not investigated. Linux, Apache, and Python are important untested neighbours of the IETF case.
  9. Several organizational traditions were not deeply profiled: sociocracy, Haier, Morning Star, Mondragon, DAOs, military mission command, intelligence-community analytic standards, institutional lessons-learned systems, Lean/PDCA, and regulated safety-management systems.
  10. Some failure evidence remains secondary or partial, especially specific NHS metric-gaming cases, the Bali disruption account, institutional-turnover figures, and several critiques of the free-energy principle.
  11. The matrices are analytical codings, not measurements. The six-part test, ten-stage matrix, maturity ladder, territory classifications, and forward radar are structured interpretations of the evidence.
  12. Confidential systems may be better integrated. Public evidence cannot rule out full or near-full loops inside enterprises, governments, military organizations, intelligence services, platforms, or health systems.

One additional caution concerns the direction of uncertainty. Several initially negative interpretations proved overstated when primary evidence was examined: human–AI augmentation was omitted from a meta-analysis summary; a multi-agent paper was cited for a conclusion it did not make; a belief-sharing paper’s proposed mitigation was omitted; and a basal-cognition paper was represented as a sceptical critique despite offering a qualified defence. Residual unverified negative claims should therefore be treated conservatively.

13. Prioritized critical reading list

Tier 1: understand the central conjunction

  1. Safron et al., “World models, artificial general intelligence and the hard problems of life–mind continuity.” Read for the conceptual landscape and explicit refusal of a single world-model definition.5
  2. Sabel and Zeitlin, “Experimentalist Governance.” Read for the strongest architecture of provisional goals, distributed implementation, peer review, dynamic accountability, and revision.21
  3. Nichols et al., “Adaptive harvest management of North American waterfowl populations.” Read for the strongest maintained rival-model system and its explicit decision to keep objectives stable.17
  4. Galesic et al., “Beyond collective intelligence: collective adaptation.” Read for the argument that adaptive trajectory, rather than static group performance, may be the right object of study.2

Tier 2: operational prior art

  1. RFC 2026 and RFC 7282. Read for versioning, retirement, contestation, appeals, and the limits of consensus.1819
  2. Cox, Arnold, and Villamayor Tomás on Ostrom’s design principles. Read for the empirical support and the authors’ acceptance of the incompleteness critique.20
  3. Williams, Szaro, and Shapiro’s Department of the Interior adaptive-management guide, paired with Westgate, Likens, and Lindenmayer’s implementation review. Read architecture beside enactment.40
  4. Somerville et al. on learning health systems, paired with the 2025 jurisdictional scan. Read for technically mature loops with underdeveloped governance and sustained evaluation.25
  5. Google’s “Postmortem Culture,” paired with Savoia, Agboola, and Biddinger on after-action-report failure. Read for the difference between recording lessons and embedding change.26

Tier 3: human–AI and disagreement

  1. Vaccaro, Almaatouq, and Malone. Read both baselines and the heterogeneity, not only the negative synergy result.30
  2. Tessler et al. on the Habermas Machine, paired with Oleart and Palomo’s democratic critique. Keep measured synthesis performance separate from normative and institutional risk.2829
  3. Catal et al., “Belief sharing: a blessing or a curse.” Read the diagnosed pathology together with the proposed mitigation.31
  4. Cleaver and Whaley, paired with Cretney. Read for power, meaning, and distributional critique inside adaptation and resilience discourse.36

Tier 4: historical and biological precedents

  1. Li Vigni on the failed institutionalization of complexity science. Read as a warning about broad promissory integration without autonomous disciplinary reproduction.38
  2. Fathi et al. on the Viable System Model. Read for what happens when a cybernetic practice survives without a strong cumulative evidence base.37
  3. McMillen and Levin, paired with Watson and Levin. Read the biological case at full strength, including its authors’ cautions about collective units and reward.910
  4. Steinsson on Wikipedia. Read for institutional change through population selection rather than intentional second-order redesign.22

Conclusion

An integrated science and practice of adaptive collective systems does not yet exist in the public evidence. What exists is more interesting than either a mature field or an empty gap: a collection of highly developed but differently incomplete traditions.

The model-building communities know how to represent, predict, simulate, and compare. The institutional communities know how to authorize, allocate, coordinate, sanction, and sometimes revise rules. The biological traditions show how distributed systems can regulate themselves across scales. The organizational traditions show how tightly bounded learning loops can become repeatable practice. Civic systems show that public input can reach real money. Standards institutions show that shared artifacts can be versioned and appealed. None combines these properties into one empirically validated institution.

The crucial open territory is the joint between epistemology and authority: where a contestable account of the world becomes consequential without becoming unquestionable, and where an institution can revise its purposes without erasing the basis on which it learns.

The prior art supplies every component. It does not yet supply the composition.

Footnotes

  1. Aleks Berditchevskaia, Eirini Maliaraki, and Konstantinos Stathoulopoulos, “A descriptive analysis of collective intelligence publications since 2000, and the emerging influence of artificial intelligence,” Collective Intelligence (2022), DOI 10.1177/26339137221107924. The 39,334-publication figure and fragmentation claims were available in the verbatim abstract. ↩

  2. Mirta Galesic, Daniel Barkoczi, Andrew M. Berdahl, et al., “Beyond collective intelligence: collective adaptation,” Journal of the Royal Society Interface 20, no. 200 (2023): 20220736, DOI 10.1098/rsif.2022.0736, PMC10031425. Full text inspected. ↩ ↩2

  3. Active Inference Institute, “June 2025 Newsletter,” https://activeinference.institute/newsletter/2025-june/; Michael Garfield, “Continual Learning: World Models in Natural & Artificial Intelligence with Adam Safron,” Humans On The Loop, July 30, 2026, https://www.humansontheloop.com/p/h-38. Both are first-party sources. ↩

  4. Philosophical Transactions of the Royal Society A 384, no. 2320, complete Crossref issue contents. Nineteen records were filed under the issue, one being an unrelated correction; the substantive issue therefore contains seventeen contributions and one editorial. ↩

  5. Adam Safron, Michael Levin, Victoria Klimaj, Zahra Sheikhbahaee, Dalton A. R. Sakthivadivel, et al., “World models, artificial general intelligence and the hard problems of life–mind continuity,” Philosophical Transactions of the Royal Society A 384, no. 2320 (2026): 20240533, DOI 10.1098/rsta.2024.0533. Verbatim publisher abstract inspected; full text was not obtained. ↩ ↩2 ↩3

  6. Michael Levin and Benjamin Lyons, “Cognitive glues are shared models of relative scarcities: the economics of collective intelligence,” Philosophical Transactions of the Royal Society A 384, no. 2320 (2026): 20240528, DOI 10.1098/rsta.2024.0528. Verbatim abstract inspected. ↩ ↩2

  7. Bradly Alicea, Morgan Hough, Amanda Nelson, and Jesse Parent, “A ‘Good’ Regulator May Provide a World Model for Intelligent Systems,” Philosophical Transactions of the Royal Society A 384, no. 2320 (2026): 20250007, DOI 10.1098/rsta.2025.0007; preprint arXiv:2506.23032. The preprint full text was inspected. See also Roger C. Conant and W. Ross Ashby, “Every good regulator of a system must be a model of that system,” International Journal of Systems Science 1, no. 2 (1970). ↩ ↩2

  8. Active Inference Institute, https://activeinference.institute/, https://activeinference.institute/projects/, https://activeinference.institute/history/, and https://activeinference.org/. Organizational figures are self-reported. ↩

  9. Patrick McMillen and Michael Levin, “Collective intelligence: A unifying concept for integrating biology across scales and substrates,” Communications Biology (2024), DOI 10.1038/s42003-024-06037-4. Evidence was available at abstract and excerpt level. ↩ ↩2

  10. Richard Watson and Michael Levin, “The collective intelligence of evolution and development,” Collective Intelligence 2, no. 2 (2023), DOI 10.1177/26339137231168355. Verbatim abstract inspected. ↩ ↩2

  11. Alejandro Fábregas-Tejeda and Matthew Sims, “On the prospects of basal cognition research becoming fully evolutionary: promising avenues and cautionary notes,” History and Philosophy of the Life Sciences 47, no. 1 (2025), DOI 10.1007/s40656-025-00660-y, PMC11802611. Full text inspected. ↩

  12. Garfield, “Continual Learning,” cited at note 3. ↩

  13. Ideoscopic, “The Elements of Ideoscopic,” https://ideoscopic.ai/about, and https://ideoscopic.ai. Public pages inspected; no organization or funding profile was disclosed. ↩

  14. World Models—in natural and artificial intelligence, Ideoscopic / Van Bettauer, 2026, https://worldmodels-ama.vercel.app/. The retained interface included the fifteen tensions, seventeen numbered contributions, three edge classifications, and its own contribution-type coding. ↩ ↩2

  15. Atlas Research Group, https://www.atlasresear.ch/, https://join.atlasresear.ch/, and “Toward Wayfinding Infrastructure for the Living Web,” May 1, 2026, https://www.atlasresear.ch/blog/toward-wayfinding-infrastructure-for-the-living-web/. ↩

  16. Atlas homepage TeamGraph data; GitHub REST API records for atlasresearch, inspected September 23, 2026; ProPublica Nonprofit Explorer, Flourishing Systems Foundation, EIN 99-1575240. ↩

  17. James D. Nichols, Michael C. Runge, Fred A. Johnson, and Byron K. Williams, “Adaptive harvest management of North American waterfowl populations: a brief history and future prospects,” Journal of Ornithology 148, suppl. 2 (2007): S343–S349, DOI 10.1007/s10336-007-0256-8. Full text inspected. See also Fred A. Johnson et al., Wildlife Society Bulletin 39, no. 1 (2015): 9–19, DOI 10.1002/wsb.518. ↩ ↩2

  18. Scott Bradner, “The Internet Standards Process—Revision 3,” RFC 2026 / BCP 9 (1996), DOI 10.17487/RFC2026. Full text inspected. ↩ ↩2

  19. Pete Resnick, “On Consensus and Humming in the IETF,” RFC 7282 (2014), DOI 10.17487/RFC7282. Full text inspected. ↩ ↩2

  20. Michael Cox, Gwen Arnold, and Sergio Villamayor Tomás, “A review of design principles for community-based natural resource management,” Ecology and Society 15, no. 4 (2010): 38, https://www.ecologyandsociety.org/vol15/iss4/art38/. Full text inspected. The specific centuries-old cases derive from Elinor Ostrom, Governing the Commons (Cambridge University Press, 1990), accessed here through secondary summaries rather than the primary book. ↩ ↩2

  21. Charles F. Sabel and Jonathan Zeitlin, “Experimentalist Governance,” in David Levi-Faur, ed., The Oxford Handbook of Governance (Oxford University Press, 2011). Author-copy full text inspected. ↩ ↩2

  22. Sverrir Steinsson, “Rule Ambiguity, Institutional Clashes, and Population Loss: How Wikipedia Became the Encyclopedia Anyone Can Edit,” American Political Science Review 118, no. 1 (2024): 235–251, DOI 10.1017/s0003055423000138. Verbatim abstract inspected; full text was not retained. ↩ ↩2

  23. Decidim, “Budgets,” https://docs.decidim.org/en/develop/admin/components/budgets.html; Barcelona City Council, participatory budgets 2024–2027; World Bank, “Participatory Budgeting and the Provision of Public Services,” Policy Research Working Paper 6968 (2014). ↩

  24. Computational Democracy Project, “vTaiwan,” https://compdemocracy.org/case-studies/2014-vtaiwan/; Fabrizio Li Vigni, “Online participation: A three-dimensional approach to study digital political platforms,” Internet Policy Review 15, no. 2 (2026), DOI 10.14763/2026.2.2093. Some current-state details were available only through a live reading rather than retained full text. ↩

  25. Mari Somerville, Christine Cassidy, Janet A. Curran, Catie Johnson, Douglas Sinclair, and Annette Elliott Rose, “Implementation strategies and outcome measures for advancing learning health systems: a mixed methods systematic review,” Health Research Policy and Systems 21 (2023): 120, PMC10680228; “Evaluating learning health systems: a jurisdictional scan,” SSM—Health Systems (2025), https://www.sciencedirect.com/science/article/pii/S2949856225000698. ↩ ↩2

  26. Daniel Rogers, Murali Suriar, Sue Lueder, et al., “Postmortem Culture: Learning from Failure,” in The Site Reliability Workbook (Google, 2018), https://sre.google/workbook/postmortem-culture/; Scott I. Tannenbaum and Christopher P. Cerasoli, “Do Team and Individual Debriefs Enhance Performance? A Meta-Analysis,” Human Factors (2013), DOI 10.1177/0018720812448394; Jennifer L. Savoia, Paul D. Agboola, and Paul D. Biddinger, “Use of After Action Reports to Promote Organizational and Systems Learning in Emergency Preparedness” (2012), PMC3447598. ↩ ↩2

  27. Anna Hegedüs, Anita Schürch, and Iren Bischofberger, “Implementing Buurtzorg-derived models in the home care setting: a Scoping Review” (2022), PMC11080323; Barbara H. Gray, Dana O. Sarnak, and Jeroen S. Burgers, “Home Care by Self-Governing Nursing Teams: The Netherlands’ Buurtzorg Model,” Commonwealth Fund (2015); de Bruin et al., Journal of Nursing Management (2022), DOI 10.1111/jonm.13836. ↩

  28. Michael Henry Tessler, Michiel A. Bakker, Daniel Jarrett, Hannah Sheahan, et al., “AI can help humans find common ground in democratic deliberation,” Science 386, no. 6719 (2024), DOI 10.1126/science.adq2852. Published abstract inspected. ↩ ↩2

  29. Álvaro Oleart and Olga Palomo, “Why AI Technosolutionism Harms Democracy and Deliberation: The EU, Citizens’ Assemblies and the Habermas Machine,” Journal of Deliberative Democracy (2025), DOI 10.16997/jdd.1839. Read at galley level; treated as normative criticism rather than evidence of downstream failure. ↩ ↩2

  30. Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone, “When combinations of humans and AI are useful: a systematic review and meta-analysis,” Nature Human Behaviour 8 (2024): 2293–2303, DOI 10.1038/s41562-024-02024-1, PMC11659167. Full text inspected. ↩ ↩2 ↩3

  31. Ozan Catal, Toon Van de Maele, Riddhi J. Pitliya, et al., “Belief sharing: a blessing or a curse,” arXiv:2407.02465. Verbatim abstract inspected. ↩ ↩2

  32. Mert Cemri, Melissa Z. Pan, Shuyi Yang, et al., “Why Do Multi-Agent LLM Systems Fail?”, arXiv:2503.13657, NeurIPS 2025 Datasets and Benchmarks Track. Full text inspected. The paper contains two different category-share presentations and does not resolve which is canonical. ↩

  33. Saaket Agashe, Yue Fan, Anthony Reyna, and Xin Eric Wang, “LLM-Coordination,” arXiv:2310.03903. Abstract inspected. ↩

  34. Christopher Hood and Maria Piotrowska, “Goodhart’s Law and the Gaming of UK Public Spending Numbers,” Public Performance & Management Review 44, no. 2 (2021): 250–271, DOI 10.1080/15309576.2020.1749092; James C. Scott, Seeing Like a State (Yale University Press, 1998), accessed here through secondary summaries. Specific NHS cases were not verified against primary records. ↩

  35. Alastair Stark, “Explaining institutional amnesia in government,” Governance (2019), DOI 10.1111/gove.12364; Jack Corbett et al., “Singular memory or institutional memories? Toward a dynamic approach,” Governance (2018), DOI 10.1111/gove.12340. Both were read at abstract or summary level. ↩

  36. Frances Cleaver and Luke Whaley, “Understanding process, power, and meaning in adaptive governance: a critical institutional reading,” Ecology and Society 23, no. 2 (2018): 49, https://www.ecologyandsociety.org/vol23/iss2/art49/; Raven Cretney, “Resilience for Whom? Emerging Critical Geographies of Socio-ecological Resilience,” Geography Compass 8, no. 9 (2014): 627–640, DOI 10.1111/gec3.12154. ↩ ↩2

  37. Mohammad Reza Fathi, Soraya Birami, Alireza Payvar, and SeyedAli Doorafshan, “A Systematic Review of the Viable System Model: Applications, Insights, and Future Directions,” Journal of Systems Thinking in Practice 4, no. 3 (2025): 109–145, DOI 10.22067/JSTINP.2025.91552.1136; American Society for Cybernetics, https://asc-cybernetics.org/; Metaphorum, https://metaphorum.org/about. ↩ ↩2

  38. Fabrizio Li Vigni, “The failed institutionalization of ‘complexity science’: A focus on the Santa Fe Institute’s legitimization strategy,” History of Science (2020), DOI 10.1177/0073275320938295; Ronald R. Kline, The Cybernetics Moment (Johns Hopkins University Press, 2015), accessed through published reviews. ↩ ↩2

  39. Shahana Akther and James Evans, “Emerging attributes of adaptive governance in the global south,” Frontiers in Environmental Science 12 (2024), DOI 10.3389/fenvs.2024.1372157. ↩

  40. Byron K. Williams, Robert C. Szaro, and Carl D. Shapiro, Adaptive Management: The U.S. Department of the Interior Technical Guide (2009), https://www.usgs.gov/publications/adaptive-management-us-department-interior-technical-guide; Martin J. Westgate, Gene E. Likens, and David B. Lindenmayer, “Adaptive management of biological systems: A review,” Biological Conservation 158 (2013), DOI 10.1016/j.biocon.2012.08.016. ↩