From Command Centers to Coordinated Ecologies

Central finding

The cybernetic project has not simply advanced from primitive computers to intelligent ones. Its organizing ambition has changed.

Mid-twentieth-century projects such as the Soviet OGAS network and Chile’s Cybersyn imagined, in different political settings, a system capable of making an economy visible to a central coordinating apparatus. Contemporary successes are usually more modest in purpose but more durable in operation: they connect autonomous organizations, route exceptions, coordinate distributed assets, or optimize bounded physical processes through shared protocols and feedback loops.

The strongest conclusion from the evidence is therefore two-sided:

• Modern sensing, communications, computation, and interoperability have substantially expanded the scale and speed at which distributed actors can coordinate.
• They have not, by themselves, resolved who sets goals, which representations count, how authority becomes legitimate, whose interests prevail, or whether institutions can maintain and correct the system.

This is not evidence that computation was historically irrelevant. Cybersyn faced a genuine shortage of computers and communications capacity, and economy-wide optimization remains an open technical question. The defensible conclusion is narrower: in many well-documented failures, the decisive breakdown occurred upstream of the computational limit—in authorization, incentives, law, representation, institutional capacity, or political commitment.

What is newly feasible is consequently not an all-seeing machine that replaces markets, hierarchies, or politics. It is a richer repertoire of hybrid coordination: thin protocols connecting autonomous organizations; centralized optimization nested inside decentralized institutions; machine-assisted attention and exception routing; recursive coordination across operational levels; and participatory systems capable of binding decisions when public authorities choose to make them binding.

---

1. The historical shift: from total visibility to selective interoperation

OGAS: the network that could not acquire authority

Viktor Glushkov’s proposed Soviet national economic-management network, commonly known as OGAS, is often remembered as a technically premature “Soviet internet.” Its history points more strongly to an institutional problem.

Economic ministries, statistical authorities, factory managers, and reformers disagreed over what the network would do and whose power it would displace. Ministries reshaped the project into extensions of their own functions. Between 1966 and 1970, hundreds of separate management-information systems were built without a common network; later growth in ministry-level computer systems reinforced branch control rather than integrating the economy. OGAS’s value depended on coordinated management reform, but the institutions whose authority would be altered controlled whether that reform could occur.¹

This distinction is crucial. OGAS did encounter the limited hardware and networking of its era. But its documented defeat occurred in bargaining over budgets, jurisdiction, information ownership, and organizational survival. Better machines would not automatically have granted the system authority to reallocate resources or compel ministries to interoperate.

Cybersyn: a real technical bottleneck—and a political design problem

Chile’s Cybersyn supplies an important corrective to any claim that computing was never binding. Under Salvador Allende, Chile had only four mainframes available to the nationalized sector, and the intended communications architecture faced a four-node ceiling. United States vendors withdrew support after Allende’s election. These were genuine constraints against a nationalized sector of more than four hundred enterprises.²

Yet the system’s clearest operational contribution during the October 1972 strike came from its low-bandwidth telex network, not its futuristic operations room, which remained a prototype. The telex system helped route roughly two thousand daily messages through improvised command centers. By May 1973, only about 27% of nationalized enterprises were connected to some degree.²

Cybersyn also demonstrates that cybernetic decentralization is not inherently participatory. Its designers often modeled production without meaningful worker involvement, sometimes presenting model-building as a technical activity workers could not understand or question. Eden Medina’s history supports both a sympathetic interpretation—participatory ambitions defeated by crisis and the coup—and a critical one, in which the design retained managerial practices that disempowered workers. The surviving record does not settle that counterfactual.

Low-technology coordination as a historical control

The United States’ Controlled Materials Plan offers an instructive contrast. Beginning in 1943, it allocated steel, copper, and aluminum through prime contractors, quarterly allotments, bills of material, field administration, and committee reconciliation. By July 1943, roughly 70% of controlled materials flowed through the system, and procurement for speculative inventory reportedly fell sharply among major contractors.³

Its contribution to total wartime output cannot be isolated, and it imposed substantial administrative burdens. But it demonstrates that national-scale allocation was possible without advanced computation when authority, objectives, reporting obligations, and administrative machinery were aligned.

The historical lesson is therefore not that information processing does not matter. It is that coordination capacity is produced jointly by information systems and institutional arrangements. In some settings, a technologically crude system with binding authority coordinated at national scale. In others, sophisticated technical plans could not cross organizational boundaries.

---

2. What has actually changed

The modern frontier is less a single system controlling an economy than an ecology of systems coordinating through stable interfaces. Four changes are especially consequential.

2.1 Communication, sensing, and actuation at distributed scale

The networking scarcity visible in Chile has been substantially relaxed. Contemporary systems can communicate continuously with millions of devices, organizations, or users and can sometimes translate those signals into near-real-time action.

Great Britain’s 2022–23 Demand Flexibility Service enrolled 1.6 million households and businesses across 22 events and reported 3,300 MWh of peak-demand reduction. Smart meters and shared communications infrastructure made it possible to coordinate many small changes in consumption without directly controlling each household.⁴ The achieved volume was small relative to national demand and expensive—about £3,330 per MWh by one estimate—but the operational capability was real.

Distributed sensing also enabled the NHS COVID-19 app to send approximately 1.7 million exposure notifications to a population of roughly 16.5 million regular users during its first deployment period. Two observational methods estimated 284,000 and 594,000 cases averted. Both supported a beneficial association, but their more than twofold difference, broad uncertainty intervals, and non-randomized design leave the effect’s magnitude and mechanism unsettled.⁵

2.2 Continuous optimization of bounded physical systems

Centralized optimization has not disappeared. Where the objective is sufficiently agreed and the system is physically measurable, it can operate continuously and at large scale.

ERCOT’s real-time electricity market uses security-constrained economic dispatch to produce locational prices and dispatch instructions approximately every five minutes.⁶ Adaptive traffic-control systems have also matured: the Federal Highway Administration’s later-phase evaluation found that commercially available adaptive products reduced travel time, stops, and delay relative to time-of-day signal timing, although the evaluation rested on a relatively small number of deployments and did not establish superiority over every well-maintained conventional adaptive system.⁷

At a smaller but technically demanding scale, a deep-reinforcement-learning controller has operated all 19 magnetic coils of a tokamak at 10 kHz, using a learned policy transferred from simulation to the physical device.⁸ This demonstrates that complex control competence can sometimes be learned without explicitly codifying all expert rules. It does not establish that social knowledge can be captured in the same way: the tokamak had a simulator, a comparatively definite objective, and physical feedback unavailable in most political or economic settings.

2.3 Thin protocols connecting autonomous organizations

The clearest durable advance is not comprehensive central planning but protocol-mediated interoperation.

Estonia’s X-Road allows organizations to maintain their own systems while exchanging data through a common architecture. Dated operator statistics show membership increasing from 499 in January 2019 to more than 2,100 by September 2026, while the number of security servers grew much more slowly. The protocol itself has been amended only four times since 2001.⁹ These facts support architectural scalability, not claims about welfare, inclusion, cost savings, or trust; the available evidence does not independently establish those outcomes.

The DNS root system exhibits a related but not identical structure. Twelve autonomous organizations operate root servers without one having authority over the others. They coordinate collective operations by consensus and common protocol while relying on IANA as the sole authoritative source of root-zone data.¹⁰ It is neither a pure hierarchy nor a fully decentralized commons. It is a layered hybrid: operational autonomy around a unique coordinating artifact.

India’s Unified Payments Interface similarly expanded from 44 participating banks in fiscal year 2016–17 to 703 in 2025–26.¹¹ Its common rails support competing banks and applications, but the available scale figures ultimately derive from the same NPCI statistical base and do not independently demonstrate inclusion, resilience, or distributive benefit. Interoperability can coexist with application concentration, fraud, exclusion, or weak grievance systems.

These systems illustrate an important shift: organizational autonomy can now be preserved at the edges while narrow forms of coherence are imposed at the interface. This is a very different achievement from optimizing the organizations’ substantive decisions.

2.4 Selective attention and exception routing

Modern machine learning is particularly effective at changing what humans must inspect.

Wikipedia’s ORES system had, by 2020, supported 110 classifiers across 44 languages. A suitably fitted vandalism model could reduce the edits requiring human review by about 90% while retaining most blatant vandalism in the review set. Individual communities could choose, train, and audit different models rather than submit to one universal classifier.¹²

This is a sustained example of human–AI coordination, but not autonomous governance. Machines route attention; communities retain responsibility for policy, judgment, model choice, and contestation. The same study reports that earlier efficiency-oriented moderation systems made newcomer interactions more impersonal and contributed to editor decline. Operational throughput improved while a different social value deteriorated.

Crisis mapping after Haiti’s 2010 earthquake shows the complementary limit. Ushahidi combined SMS reports, social media, geolocation, diaspora translation, and globally distributed volunteers to produce maps used by responders within days. But verification, translation, triage, and connection to institutions able to act remained labor-intensive. The system expanded sensing faster than organizations could absorb and respond to what was sensed.¹³

---

3. A three-part ledger of constraints

The history is best understood by separating constraints that have been substantially relaxed, transformed but not solved, and still fundamentally institutional or political.

Status	Constraints	What the evidence supports
Substantially relaxed	Communications reach; storage; ubiquitous metering; rapid distributed notification; actuation of dispersed physical assets; bounded real-time optimization; cross-organizational interoperation; computational plan consistency	Systems can now coordinate millions of endpoints, optimize bounded physical processes continuously, and connect autonomous organizations through common protocols at scales impractical in the mid-twentieth century.
Transformed, not solved	Measurement validity; aggregation and representation; delay and distortion; exception handling; tight-coupling risk; tacit-knowledge representation; economy-scale optimization	More data and faster feedback improve some operations but also accelerate error, gaming, model drift, overload, and cascading failure. Computation changes where the bottleneck appears rather than reliably removing it.
Fundamentally institutional, political, epistemic, or social	Goal formation; authorization; legitimate authority; legal validity; allocation rules; distribution of power; organizational capacity; maintenance; consent; political commitment	These determine which objectives are pursued, which errors are tolerated, who may override the system, who bears costs, and whether technical capability survives as a public institution.

3.1 Measurement remains vulnerable to incentives and drift

Cybernetic coordination depends on feedback, but measurements are not neutral windows onto reality.

During the Vietnam War, the Hamlet Evaluation System turned local conditions into quantitative ratings used for benchmarking. Advisors responsible for many settlements faced pressure not to downgrade them, degrading the very information the system was meant to improve.¹⁴ Decades later, Google Flu Trends missed high for 100 of 108 weeks beginning in August 2011; even three-week-old conventional public-health data predicted current prevalence better. The error came not from insufficient data volume but from model drift and treating digital traces as a substitute for, rather than a supplement to, grounded measurement.¹⁵

Robodebt shows the same problem with much higher stakes. Australia’s system averaged annual income into fortnightly periods, treating people as though they had earned income when they had not. The Royal Commission found that the scheme was devised without regard to social-security law and that the averaging method was fundamentally unfair. The government ultimately reimbursed $746 million to about 381,000 people and wrote off $1.751 billion in debts.¹⁶

Automation did not merely reproduce a bad measurement. It converted that measurement into a rapid, large-scale administrative action.

3.2 Representation and optimization remain distinct problems

Mid-century planning systems had to compress economies into manageable categories. Faster computers can process many more categories, but they cannot guarantee that the representation captures what matters.

The same distinction sharpens the debate over economy-wide computation. Cockshott and Cottrell provide a credible algorithmic argument that a sparse input-output system involving millions of products can be balanced through iterative calculation. But this addresses consistency—whether proposed production requirements fit together—not necessarily optimization among competing objectives. Their optimization method finds only a local maximum and was demonstrated on a simulation of about 4,000 industries before being extrapolated.¹⁷

Conversely, frequently cited estimates that economy-scale planning would require centuries of computation refer to a different linear-programming optimization problem. The arguments are therefore partly talking past one another. No side in the evidence reviewed has supplied a benchmark using real economy-scale data, and the only formal planner-versus-market simulation located explicitly disclaims empirical superiority because its assumptions favor the planner.¹⁸

Economy-wide optimization should therefore remain in the “transformed but not solved” category. Plan balancing has a plausible computational basis; comprehensive optimization under plural, changing objectives has not been empirically demonstrated.

3.3 Faster feedback creates new failure modes

Speed can improve routine coordination while increasing the severity of mistakes.

The Securities and Exchange Commission found that algorithmic trading generally improved market quality and liquidity under ordinary conditions. During the 2010 Flash Crash, however, interacting systems and liquidity withdrawal created a negative feedback loop in which the bid side virtually disappeared. Trading pauses helped restore order.¹⁹

Project Symphony, an Australian distributed-energy pilot involving 514 residential customers and 911 distributed-energy assets, produced a smaller version of the same conflict. The system technically coordinated an aggregator, distribution operator, and market operator. But its first five-minute dispatch design prioritized system targets at the expense of household-level optimization, generating strong dissatisfaction during one bidirectional-energy scenario and broader confusion and anxiety during early orchestration. Sentiment improved after more transparency and engagement.²⁰

The implication is not that fast feedback should be avoided. It is that speed and coupling increase the value of circuit breakers, appeal paths, observability, rollback, and human interruption.

3.4 Verification does not determine allocation

The clearest causal evidence comes from a randomized experiment involving 15 million beneficiaries of India’s subsidized-food program in Jharkhand. More stringent biometric authentication did not reduce leakage by itself. It slightly increased transaction costs and reduced benefits received by previously unregistered beneficiaries by 10.6%.²¹

This was not a pure improvement in sensing: biometric authentication also changed the access rule. That is precisely why the result matters. Adding verification capability without changing the allocation mechanism did not solve leakage and did impose exclusion costs. Leakage declined only in a later phase when authenticated information was used to alter allocations—an institutional decision that also reduced benefits received.

Information can make an allocation rule more enforceable. It does not decide whether that rule is fair, legitimate, or substantively effective.

---

4. The institutional constraints that capability does not dissolve

Goal formation

Optimization requires an objective. Public systems usually have several: efficiency, reliability, equity, privacy, safety, geographic balance, and political legitimacy.

Google’s Green Light illustrates the problem in an everyday setting. Its recommendations optimize traffic variables, but municipal engineers sometimes ignore them because cities deliberately favor buses, discourage residential through-traffic, or account for pedestrians and modes outside the model. Such overrides are not necessarily algorithmic failure. They are evidence that a single-objective recommendation has encountered a multi-objective public mandate.²²

No improvement in optimization decides how these goals should be weighted.

Legal validity and due process

The Australian Robodebt inquiry recommended that business rules and algorithms be available for independent expert scrutiny. The recommendation was institutional because the underlying failure was institutional: an automated system scaled a legally invalid administrative practice.¹⁶

The Dutch data-protection authority similarly found three forms of nationality-related processing in childcare-benefit administration unlawful. Two—using nationality in an automated risk-classification model and in organized-fraud detection—were also discriminatory because the distinctions lacked a reasonable and proportionate relationship to their purposes.²³

Technical accuracy does not confer lawful authority. Nor does it supply notice, explanation, review, proportionality, or a remedy for affected people.

Institutional capacity and maintenance

ERCOT’s five-minute optimizer continued to exist during the February 2021 Texas electricity crisis. The principal failures lay elsewhere: thousands of outages, derates, and failures to start, with freezing responsible for 1,823 events and frozen instrumentation among the leading causes. Similar risks had appeared in previous winter events but had largely been left to voluntary preparation.²⁴

The episode separates optimization from resilience. A system can calculate dispatch correctly while generators, sensors, fuel infrastructure, and governance arrangements fail around it.

Adaptive traffic systems face a less dramatic version of the same constraint. Costs, communications, technical transfer, market acceptance, maintenance, and sustained agency support have repeatedly limited deployment. The algorithm is only one component of an operational institution.

Legitimate and binding authority

Digital participation can now process proposals, votes, and implementation records at substantial scale. Barcelona’s 2020–23 participatory budget allocated €30 million through 1,982 submitted proposals, municipal feasibility review, prioritization, voting, and selection of 76 projects by 64,571 participants.²⁵ This is a genuine case of digitally supported binding allocation.

But platform capacity is not the same as democratic authority. Across Catalan municipalities, officials valued transparency, information organization, and proposal collection more consistently than deliberation or transferring sovereignty, although the platforms did force administrations to consider more citizen input and brought in new publics. That finding is available only through the publisher-deposited abstract and should not bear more weight than that representation permits.²⁶

vTaiwan demonstrates both possibility and fragility. Its UberX consultation recorded 2,251 voters, 52,976 vote events, and 199 submitted statements. Seventy-eight statements—39%—were moderated out before voting.²⁷ The export does not record why, so it cannot establish whether moderators disproportionately removed minority or conflictual views. The process nevertheless influenced legislation and included public officials with authority to change policy. It has not been used for a major decision since 2018.

The appropriate interpretation is not that vTaiwan proved algorithmic consultation ineffective. It showed that an institutional linkage could work and still fail to become durable. The technology could structure a public conversation; political authorities determined whether the process continued and mattered.

Distribution of power

Formal decentralization does not guarantee political decentralization.

At the measured dates, eight delegates could form a majority in Compound governance, eleven in Uniswap, and eighteen in ENS. Large delegates rarely overturned broader voting outcomes, so these figures describe latent authority and vulnerability rather than continuous capture.²⁸

The original DAO likewise combined protocol execution with concentrated human functions. Curators controlled which proposals could be admitted, a code exploit diverted roughly one-third of its assets, and recovery required human coordination and a hard fork.²⁹

Algorithmic consultation has analogous gatekeepers: moderators select and frame statements; platform operators define participation rules; authorities decide whether results bind. Decentralized execution can coexist with centralized agenda control.

---

5. The coordination forms now demonstrated

The evidence supports a broader repertoire than either “central planning” or “decentralized markets,” but its maturity varies sharply.

Coordination form	Demonstrated case	Maturity	What is established	What remains unestablished
Thin-protocol interoperation among autonomous organizations	X-Road, DNS root, UPI	Sustained	Independent operators can coordinate around stable standards and authoritative shared artifacts without operational merger.	Comparative welfare, inclusion, resilience, cost, trust, and distributional effects.
Central real-time optimization of bounded physical systems	ERCOT dispatch; adaptive traffic control	Sustained	Continuous optimization can operate at large physical-system scale when objectives and variables are sufficiently defined.	Resilience to institutional neglect; superiority over every capable conventional baseline.
Recursive coordination across institutional levels	Project Symphony	Pilot	An aggregator, network operator, and market operator can exchange feedback and orchestrate household assets end to end.	Mature commercial operation, acceptable benefit allocation, customer authority, and settled market rules.
Population-scale distributed sensing and actuation	Demand Flexibility Service; NHS app	Bounded deployment	Millions of people can be coordinated rapidly without continuous central micromanagement.	Large or durable system effects; low cost; representative participation; causal effect magnitude.
Human–AI attention and exception routing	Wikipedia ORES	Sustained	Models can reduce review loads while communities retain policy and adjudication.	Autonomous governance; absence of social costs or bias.
Bridging-based consensus	Community Notes	Bounded deployment	Agreement across previously disagreeing raters can identify notes associated with lower reposting and higher deletion of misleading posts.	Broad coverage: only 11.5% of notes reached publication consensus in one study, publication averaged 65.7 hours, and participation was concentrated.30
Hybrid digital–physical participatory allocation	Barcelona participatory budgeting	Bounded deployment	Proposal intake, feasibility review, voting, and implementation tracking can support a binding public allocation.	Who participates, whose preferences prevail, and whether authority transfers beyond the bounded budget.
Algorithmically structured mass consultation	vTaiwan	Demonstrated but not sustained	Large-scale preference mapping can inform legislation when empowered officials participate.	Durability, representative legitimacy, and neutral moderation.
Crowdsourced crisis sensing	Ushahidi in Haiti	Bounded deployment	Distributed volunteers can surface and map information missed by conventional channels.	Scalable verification, triage, and reliable connection to institutions able to act.
Polycentric institutional coordination	Commons-governance cases	Sustained at local and nested-regional scale	Formally independent decision centers can coordinate without a single hierarchy.	Generalization to economy-wide governance or proof that any one design principle is independently necessary.
Protocol-executed decisions and shared assets	DAOs	Bounded, with major failures	Rules and asset transfers can be executed across organizational boundaries.	Decentralized power, robust security, legitimate agenda-setting, and crisis resolution without concentrated human authority.
General autonomous-agent coordination across social institutions	No qualifying case in the reviewed evidence	Not demonstrated	Frontier agents can perform increasingly long bounded software tasks.	Sustained multi-institution allocation, legitimate delegated authority, long-horizon stability, and audited real-world outcomes.

Polycentric governance deserves particular qualification. Elinor Ostrom’s work establishes that multiple formally independent decision centers need not produce chaos.³¹ A subsequent qualitative comparative analysis found that no single design principle guaranteed success; configurations of principles mattered, with congruence between rules and local conditions appearing especially important.³² But only 27 cases had complete data, and the authors explicitly warned against generalizing even to the remaining cases in their sample. Nothing in this evidence carries the result from local commons and nested regional systems to an entire economy.

General autonomous AI is at the opposite end of the maturity spectrum. METR measured a roughly 110-minute 50% task-completion horizon for a 2025 frontier agent on structured software tasks and warned that external validity to messy real-world work was uncertain.³³ That is evidence of improving bounded capability, not autonomous social coordination. No audited, sustained, multi-institution agent deployment meeting the broader standard was demonstrated in the evidence reviewed; the search base is too limited to conclude that none exists.

---

6. What follows for system design

Several design implications are well supported across otherwise dissimilar cases.

6.1 Treat capability and authority as separate layers

A system may be capable of sensing, predicting, optimizing, or executing without having legitimate authority to decide. OGAS, the Jharkhand experiment, vTaiwan, Project Symphony, ORES, and DAO governance all separate these layers in different ways.

A sound architecture should state explicitly:

1. what the technical system can observe and do;
2. who authorizes each action;
3. which decisions remain local;
4. who may interrupt or reverse execution;
5. how affected parties can challenge a decision; and
6. which institution is responsible when the model, rule, or data are wrong.

6.2 Use thin protocols for interoperation—not as substitutes for allocation

X-Road and DNS suggest that stable interfaces can connect diverse organizations without absorbing them into a central apparatus. This pattern is especially promising when the problem is interoperation: authentication, message exchange, discovery, settlement, or publication of an authoritative shared artifact.

It is less informative when the problem is allocation: deciding who receives resources, whose objective dominates, or how losses are distributed. A protocol can execute an allocation rule consistently. It cannot make the rule legitimate.

6.3 Preserve local autonomy while making exceptions visible

The most credible human–AI pattern is not end-to-end automated governance. It is:

• local actors retaining decision rights;
• machines filtering routine cases;
• uncertain or consequential cases being escalated;
• common observability across levels;
• appeal and audit paths; and
• feedback from reversals or local overrides informing later model changes.

ORES approximates this pattern for moderation. Project Symphony demonstrates the technical possibility across energy institutions but also shows what happens when system-level dispatch overrides household objectives without sufficient explanation or consent.

6.4 Design for interruption, not only optimization

The Flash Crash, Robodebt, and Project Symphony all support the same proposition: as systems become faster and more tightly coupled, interruption becomes more valuable.

Circuit breakers, staged execution, rollback, independent review, auditable business rules, human override, and meaningful appeal should be treated as core coordination infrastructure rather than friction added after deployment.

6.5 Make objective conflicts observable

Local overrides are often valuable information. When an engineer rejects a traffic recommendation to favor buses or protect a residential street, or when a household rejects energy dispatch that harms its own objective, the override reveals a missing or contested value.

Systems should record not only whether a recommendation was followed but why it was rejected. This converts disagreement from an obstacle into feedback about model scope, institutional priorities, and distributional conflict.

6.6 Evaluate institutional durability separately from technical success

A successful pilot can prove that data exchange and control are possible while saying little about whether the arrangement can survive changes in government, funding, regulation, staff, or public trust.

Evaluation should therefore distinguish:

• technical performance;
• institutional adoption;
• ongoing maintenance;
• distribution of benefits and burdens;
• legitimacy and consent;
• persistence through political turnover; and
• performance under stress.

vTaiwan is a capability success and a durability failure. Treating it as simply successful or unsuccessful loses the central lesson.

6.7 A plausible experiment

A useful next experiment would not attempt to automate an entire institution. It would test a recursive coordination stack in a domain spanning several organizations:

• a thin common protocol for data and requests;
• local decision rights at each participating organization;
• shared real-time observability;
• AI-assisted triage of routine cases and anomalies;
• explicit escalation paths across levels;
• binding circuit breakers and appeal rights;
• public records of overrides, delays, and distributional effects; and
• periodic renegotiation of objectives and operating rules.

This is a design possibility, not an empirically established superior form. It should be compared against a capable conventional hierarchy or regulated market on common measures: cost, delay, reliability, exclusion, incident severity, intervention frequency, concentration of authority, user trust, and distribution of outcomes. The current evidence contains no such head-to-head comparison.

---

7. Uncertainties and unresolved questions

Economy-wide optimization remains open

The central technical dispute is unresolved because plan consistency and plan optimization are different problems, and neither side has supplied an empirical benchmark on real economy-scale data. Resolving this would materially change where computation belongs in the constraint ledger.

The strongest institutional control cases are missing

The reviewed evidence does not cover large-scale internal planning at firms such as Walmart or Amazon, contemporary Chinese planning implementations, or several important historical military and public-administration systems. These are important controls because ownership or authority may settle goal formation more decisively than in democratic, inter-agency, or cross-organizational settings. No conclusion about those cases is warranted here.

Architecture is better evidenced than social outcome

X-Road, UPI, and DNS establish that protocol-mediated coordination is operationally feasible. The evidence does not establish its net effects on administrative cost, fraud, exclusion, concentration, resilience, public trust, or distribution. Those require independent longitudinal evaluation.

Participation is counted better than power

Digital participation research measures proposals, votes, comments, and participants more readily than it measures whose preferences prevail. Barcelona demonstrates a binding allocation, but the broader evidence is too weak to conclude that digital participation consistently transfers sovereignty or improves representation.

New systems are rarely compared fairly with capable old ones

Few studies compare a novel hybrid architecture against a high-capacity conventional hierarchy or well-regulated market using identical outcome measures. Adaptive traffic control is one partial exception, but even there the evidence does not establish how much advanced AI adds beyond mature conventional adaptive systems.

The autonomous-agent conclusion is asymmetric

Large-scale autonomous-agent coordination was not demonstrated in the reviewed evidence. That is an absence of supporting evidence, not proof of nonexistence. An audited deployment reporting task duration, human intervention, failures, security incidents, appeals, rollback, and long-term institutional performance could change the assessment.

Historical selection effects are possible

Documented cybernetic projects may overrepresent programs defeated by institutional politics because programs defeated earlier by computational infeasibility left less evidence. Cybersyn’s counterfactual—whether it would eventually have struck a hard technical ceiling absent the coup—cannot be tested.

---

Conclusion

The cybernetic project has evolved from the pursuit of comprehensive command toward the construction of selective, layered coherence.

The historical aspiration was to render a complex economy visible and governable from a coordinating center. The contemporary achievement is more plural: independent organizations can interoperate through stable protocols; bounded physical systems can be optimized continuously; millions of distributed devices or people can be reached rapidly; algorithms can route attention and identify exceptions; and participatory processes can support binding decisions when authorities commit themselves to the result.

These are substantial advances. They make feasible combinations of local autonomy and system-level coordination that were previously impractical.

But the frontier has moved rather than vanished. More computation exposes unresolved questions about objectives, representation, authority, due process, distribution, maintenance, consent, and institutional durability. In some cases it amplifies those questions by increasing speed and scale.

The strongest general design lesson is therefore not to choose between centralization and decentralization. It is to compose them deliberately: centralize only the functions that require a common artifact or physical optimization; preserve local authority where knowledge and consequences are local; use protocols to connect rather than absorb organizations; assign AI to routing and bounded control before delegating governance; and make interruption, appeal, audit, and rule revision part of the architecture from the beginning.

The machines have made much more coordination possible. They have not decided what coordination is for.

---

Footnotes

1. Slava Gerovitch, “InterNyet: Why the Soviet Union Did Not Build a Nationwide Computer Network,” History and Technology 24, no. 4 (2008): 335–350, doi:10.1080/07341510802044736; Benjamin Peters, “How the Soviets Invented the Internet and Why It Didn’t Work,” Aeon, October 17, 2016, https://aeon.co/essays/how-the-soviets-invented-the-internet-and-why-it-didnt-work. These accounts draw on an overlapping Soviet archival and memoir base and should not be treated as fully independent corroboration. ↩

2. Eden Medina, Cybernetic Revolutionaries: Technology and Politics in Allende’s Chile (MIT Press, 2011), ISBN 9780262525961; Medina, “Cybernetic Revolutionaries,” Cabinet 46 (2012), https://www.cabinetmagazine.org/issues/46/medina.php. ↩ ↩²

3. Robert Cuff, “Organizational Capabilities and U.S. War Production: The Controlled Materials Plan of World War II,” Business and Economic History, 2nd ser., 19 (1990): 103–112, https://thebhc.org/sites/default/files/beh/BEHprint/v019/p0103-p0112.pdf. ↩

4. National Energy System Operator, “Demand Flexibility Service Delivers Electricity to Power 10 Million Households” (2023), https://www.neso.energy/news/demand-flexibility-service-delivers-electricity-power-10-million-households; Carbon Brief, “Q&A: How Great Britain’s ‘Demand Flexibility Service’ Is Cutting Costs and CO2 Emissions” (2023), https://www.carbonbrief.org/qa-how-great-britains-demand-flexibility-service-is-cutting-costs-and-co2-emissions. The operator’s “10 million households” phrase is an hour-of-peak equivalence, not sustained supply. ↩

5. Chris Wymant et al., “The Epidemiological Impact of the NHS COVID-19 App,” Nature 594 (2021): 408–412, doi:10.1038/s41586-021-03606-z. The evidence review had access to the abstract, metadata, disclosures, and discussion rather than the complete paywalled article. ↩

6. Electric Reliability Council of Texas, “Real-Time Market,” https://www.ercot.com/mktinfo/rtm; ERCOT, Market Education Resource 301, Module 5: Resources in Real-Time Operations (March 2, 2020), https://www.ercot.com/files/docs/2020/03/02/2020_03_R301_M5_RT.pdf. ↩

7. Sari Radin et al., Federal Highway Administration Research and Technology Evaluation Final Report: Adaptive Signal Control, FHWA-HRT-17-007 (2018; errata 2019), https://www.fhwa.dot.gov/publications/research/randt/evaluations/17007/17007.pdf. Earlier FHWA phases produced mixed and sometimes negative results; the favorable conclusion applies to the later commercial-product phase and is limited by the number of evaluations. ↩

8. Jonas Degrave, Federico Felici et al., “Magnetic Control of Tokamak Plasmas through Deep Reinforcement Learning,” Nature 602 (2022): 414–419, doi:10.1038/s41586-021-04301-9. ↩

9. Estonian Information System Authority, “X-tee Statistics—EE Instance Current Environment Data and Monthly History,” data gathered September 20, 2026, https://x-tee.ee/stats/EE/environmentData.json and https://x-tee.ee/stats/EE/history.json. Conflicting annual transaction figures on static promotional pages are not used here. ↩

10. ICANN Root Server System Advisory Committee, RSSAC037: A Proposed Governance Model for the DNS Root Server System (June 12, 2018), https://www.icann.org/en/system/files/files/rssac-037-15jun18-en.pdf. ↩

11. Press Information Bureau, Government of India, “UPI Completes 10 Glorious Years, Emerges as World’s Largest Real-Time Payments Platform, Anchoring India’s Digital Economy” (2026), https://www.pib.gov.in/PressReleasePage.aspx?PRID=2257087&reg=3&lang=2; Laurent Gonnet, Gynedi Srinivas, and Guillermo Galicia Rabadan, “Fast Payments Beyond Speed: India’s Pioneering Experience,” World Bank, January 21, 2026, https://blogs.worldbank.org/en/psd/fast-payments-beyond-speed--india-s-pioneering-experience. Both presentations rely substantially on NPCI data and are not independent measurements. ↩

12. Aaron Halfaker and R. Stuart Geiger, “ORES: Lowering Barriers with Participatory Machine Learning in Wikipedia,” Proceedings of the ACM on Human-Computer Interaction 4, CSCW2, art. 148 (2020), doi:10.1145/3415219. ↩

13. Jessica Heinzelman and Carol Waters, Crowdsourcing Crisis Information in Disaster-Affected Haiti, United States Institute of Peace Special Report 252 (October 2010), https://www.files.ethz.ch/isn/121995/SR252%20-%20Crowdsourcing%20Crisis%20Information%20in%20Disaster-Affected%20Haiti.pdf. ↩

14. Dave Young, “Computing War Narratives: The Hamlet Evaluation System in Vietnam,” A Peer-Reviewed Journal About 5, no. 1 (2016), https://aprja.net/article/download/116011/164196. ↩

15. David Lazer, Ryan Kennedy, Gary King, and Alessandro Vespignani, “The Parable of Google Flu: Traps in Big Data Analysis,” Science 343, no. 6176 (2014): 1203–1205, doi:10.1126/science.1248506. ↩

16. Royal Commission into the Robodebt Scheme, Report, Volume 1 (Commonwealth of Australia, July 7, 2023; corrected July 11, 2023), official report page at https://robodebt.royalcommission.gov.au/publications/report. The reviewed full text was an Internet Archive copy because the originating PDF endpoint was unreachable; its title page, date, commissioner, and corrigendum reference were consistent with the official report. ↩ ↩²

17. W. Paul Cockshott and Allin Cottrell, Towards a New Socialism (Spokesman Books, 1993), author-hosted edition at https://users.wfu.edu/cottrell/socialism_book/new_socialism.pdf. ↩

18. Cosma Shalizi, “In Soviet Union, Optimization Problem Solves You,” Crooked Timber, May 30, 2012, https://crookedtimber.org/2012/05/30/in-soviet-union-optimization-problem-solves-you/; Ricardo Alonzo Fernández Salguero, “Testing Centralized and Polycentric Computational Planning,” arXiv:2606.19214 (2026), https://arxiv.org/abs/2606.19214. The former is an informal technical essay and discussion; the latter is a self-limited simulation preprint, not an empirical economy-scale test. ↩

19. Staff of the U.S. Securities and Exchange Commission, Staff Report on Algorithmic Trading in U.S. Capital Markets (August 5, 2020), https://www.sec.gov/files/Algo_Trading_Report_2020.pdf. ↩

20. Western Power, Australian Energy Market Operator, Synergy, and Energy Policy WA, DER Participation: Pilot Results and Recommendations—Project Symphony Final Report (April 2024), https://arena.gov.au/assets/2024/06/Western-Power-Project-Symphony-Pilot-Results-and-Recommendations.pdf; Australian Energy Market Operator and partners, DER Participation Framework (November 2023), https://arena.gov.au/assets/2024/06/Project-Symphony_DER-Participation-Framework-Work-Package-7-Report.pdf. ↩

21. Karthik Muralidharan, Paul Niehaus, and Sandip Sukhtankar, “Identity Verification Standards in Welfare Programs: Experimental Evidence from India,” NBER Working Paper 26744, revised April 2020, https://www.nber.org/papers/w26744. ↩

22. Google Research, “Green Light: Reduce Traffic Emissions with AI,” https://sites.research.google/greenlight/; “Google’s Project Green Light Uses AI to Take on City Traffic,” Scientific American (2024), https://www.scientificamerican.com/article/googles-project-green-light-uses-ai-to-take-on-city-traffic/. Performance claims are vendor-reported and baseline-dependent; no independent causal evaluation was available. ↩

23. Autoriteit Persoonsgegevens, Belastingdienst/Toeslagen—De verwerking van de nationaliteit van aanvragers van kinderopvangtoeslag, report z2018-22445 (July 17, 2020), https://www.autoriteitpersoonsgegevens.nl/sites/default/files/atoms/files/onderzoek_belastingdienst_kinderopvangtoeslag.pdf. ↩

24. FERC, NERC, and Regional Entity Staff, February 2021 Cold Weather Grid Operations: Preliminary Findings and Recommendations (September 23, 2021), https://www.ferc.gov/sites/default/files/2021-09/February%202021%20Cold%20Weather%20Grid%20Operations_Preliminary%20Findings%20and%20Recommendations-09-23-21_Full%20Presentation.pdf; Chiara Lo Prete and Seth Blumsack, “Enhancing the Reliability of Bulk Power Systems against the Threat of Extreme Weather: Lessons from the 2021 Texas Electricity Crisis” (2023), https://par.nsf.gov/servlets/purl/10483330. ↩

25. Decidim and Barcelona City Council, “Participatory Budgets 2020–2023 in Barcelona,” https://decidim.org/case-studies/barcelona-budgeting/. This is implementation documentation produced with the city and should be read as evidence of process and throughput rather than an independent evaluation of benefit or legitimacy. ↩

26. Rosa Borge, Joan Balcells, and Albert Padró-Solanet, “Democratic Disruption or Continuity? Analysis of the Decidim Platform in Catalan Municipalities,” American Behavioral Scientist 67, no. 7 (2022), doi:10.1177/00027642221092798. Only the publisher-deposited abstract and metadata were available to the evidence review. ↩

27. Computational Democracy Project, pol.is export for “UberX 自用車載客—vTaiwan.tw,” conversation opened July 15, 2015, export snapshot September 21, 2026; export-format documentation at https://compdemocracy.org/export/. For broader institutional context, see the project’s vTaiwan case study, https://compdemocracy.org/case-studies/2014-vtaiwan/. ↩

28. Robin Fritsch, Marino Müller, and Roger Wattenhofer, “Analyzing Voting Power in Decentralized Governance: Who Controls DAOs?” Blockchain: Research and Applications 5, no. 3 (2024): 100208, doi:10.1016/j.bcra.2024.100208. ↩

29. U.S. Securities and Exchange Commission, Report of Investigation Pursuant to Section 21(a) of the Securities Exchange Act of 1934: The DAO, Release No. 81207 (July 25, 2017), https://www.sec.gov/litigation/investreport/34-81207.pdf. ↩

30. Yuwei Chuai et al., “Community-Based Fact-Checking Reduces the Spread of Misleading Posts on Social Media,” arXiv:2409.08781 (2024), https://arxiv.org/abs/2409.08781; Olesya Razuvayevskaya et al., “Timeliness, Consensus, and Composition of the Crowd: Community Notes on X,” arXiv:2510.12559, version 2 (2025), https://arxiv.org/abs/2510.12559. The reported behavioral effects are conditional on a note being published and should not be interpreted as system-wide effects. ↩

31. Elinor Ostrom, “Beyond Markets and States: Polycentric Governance of Complex Economic Systems,” Nobel Prize Lecture, December 8, 2009; also American Economic Review 100, no. 3 (2010): 641–672, https://www.nobelprize.org/uploads/2018/06/ostrom_lecture.pdf. ↩

32. Jacopo A. Baggio et al., “Explaining Success and Failure in the Commons: The Configural Nature of Ostrom’s Institutional Design Principles,” International Journal of the Commons 10, no. 2 (2016): 417–439, doi:10.18352/ijc.634. ↩

33. Thomas Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” Advances in Neural Information Processing Systems 38 (2025), doi:10.52202/085713-3086; arXiv:2503.14499, https://arxiv.org/abs/2503.14499. ↩