VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

CivilisationOS | From Inherited Knowledge to Emergent Institutions

Three learners review open books together at a classroom table, with stacks of textbooks, stationery and a whiteboard in the bright room.

Humanity has accumulated more than a store of information. Across generations, people have inherited skills, stories, tools, habits, rules, institutions and ways of correcting mistakes. Artificial intelligence is now being trained on substantial traces of that inheritance. This creates a question that goes beyond whether an AI system can reproduce an explanation: under what conditions could systems trained on inherited human knowledge help form new, durable arrangements for doing work together?

CivilisationOS | From Inherited Knowledge to Emergent Institutions examines the missing transition between what is recorded and what becomes an institution. A model may describe a library, coordinate a group of agents or repeat a familiar rule without establishing a functioning institution. The stronger claim requires observable coordination, accepted responsibilities, shared resources, correction, continuity and the preparation of successors. This article connects the existing work on civilisational inheritance, AI feedback, education and portable institutions, then proposes a controlled investigation. The experiment is a research design, not a completed trial; no result is invented or presented as observed.

  • Separate inherited records, usable capability, repeated coordination, convention, institution and material renewal.
  • Understand why human cultural inheritance is cumulative but never a perfect copy of the past.
  • Distinguish an AI system that follows a supplied institutional design from agents that develop a working arrangement through interaction.
  • Examine why an emergent convention can be useful, exclusionary, biased or unstable.
  • Specify what an institution must demonstrate: purpose, roles, rules, decision rights, resources, monitoring, correction and succession.
  • Compare four controlled inheritance conditions in a proposed shared-resource experiment.
  • Connect the results back to education, human responsibility, portable civilisation and the next generation of learners.

A reading route through this companion

Read the full argument in order, or follow the stages that answer your immediate question. The chapters below remain open and cross-reference the connected books.

1. Return to the original observation: civilisation is inherited and changed

A child does not begin with an empty world. Language, number systems, cooking methods, maps, tools, public rules and stories are already present. A learner receives some of them through instruction, some through imitation, some through participation and some through exploration. The inheritance is not simply a package handed over intact. The learner interprets it, uses it in a new situation, discovers where it fails and may pass on a changed version.

This is one reason that human civilisation can continue while its individual carriers change. People die, move, disagree and forget. Some practices disappear; others are deliberately preserved. A useful practice may spread without any one person planning the total arrangement that eventually results. Local intentions can combine into a larger pattern that no single participant designed in full.

Research on cumulative cultural learning describes how children inherit and modify group-specific knowledge, practices and technologies through complementary processes such as exploration, observation, participation, imitation and instruction. It does not suggest that every cultural trait is copied faithfully or that accumulation is automatic. The point for this project is narrower: inheritance already depends on learners who can interpret, select, adapt and transmit, not merely store. See Cristine H. Legare, The Development of Cumulative Cultural Learning (2019).

The earlier eduKate article Civilisation | What is Next? | The Single Cell Civilisation | Vol.05 already develops civilisational heredity, the archive, the interpreter, AI inheritance and material renewal. This companion does not claim to discover those ideas again. It asks a more specific next question: when inherited knowledge is placed into interaction, what evidence would show that a new institution has formed?

2. What AI inherits is not the whole of human experience

A trained model can acquire statistical structure from human-produced records. Those records include explanations, descriptions, examples, arguments, procedures and traces of institutional life. They are not the same as direct participation in every activity they describe. A manual about maintaining a water system is not the water system. A set of interviews about teaching is not a classroom relationship. A public charter is not the trust, authority or resources that allow an organisation to operate.

It is useful to distinguish at least six forms of inheritance. Recorded knowledge is material that can be retrieved or represented. Learned model behaviour is a capacity acquired through training and later adjustment. Procedural competence is the ability to carry out a task under stated conditions. Tacit skill is knowledge that may be difficult to express fully in words. Physical capability includes tools, energy, infrastructure and materials. Institutional inheritance includes recognised roles, commitments, decision procedures and accepted ways to correct or replace an arrangement.

These forms overlap, but they are not interchangeable. A model may explain how a pump works but lack the physical means to repair one. An agent may call another agent a supervisor without the group recognising any legitimate authority. A record may describe a dispute procedure, while no one has the power or willingness to apply it. The receiver’s existing capacities and environment supply some of what the record does not.

The distinction between a packet and its host is therefore essential. Portability Of Civilisation | Packets and Wrappers separates the mechanism being transferred from the wrapper that binds it to a receiving context. The present question adds another layer: can a receiving population use inherited material to develop a coordination arrangement that was not already specified as the final institutional design?

3. The central distinction: copying, enacting and emerging

Three processes can produce similar-looking outputs while supporting very different conclusions. In copying, the receiving system reproduces a stored description or rule. In enactment, people or agents use a supplied arrangement to perform a task. In emergence, repeated interaction produces a pattern or institution that was not supplied in full as the final design, even though the environment, starting conditions and boundaries were necessarily chosen by someone.

Suppose a group receives a complete charter stating who may request a book, who approves a loan, how a return is recorded, who settles a dispute and how the next coordinator is chosen. If the group follows the charter, the exercise can test whether the charter is clear and portable. It does not show that the group independently developed an institution. The institutional architecture was already present in the instructions.

Now suppose the group receives a shared task, a limited collection, incomplete records and a need to coordinate access. It must propose rules, discover which roles are needed, respond to disputes and revise a procedure after a failure. A recurring arrangement may emerge. That is a different result, but the researcher must still inspect how much of the outcome was shaped by prompts, rewards, hidden tool settings, agent memory, imposed turn-taking or the designer’s choice of available actions.

Emergence does not mean creation from nothing. No meaningful experiment begins from nothing. It means that the specific higher-level pattern is not simply copied from a complete blueprint. The researcher must report what was supplied, what was left open, what participants changed and which decisions were actually made during interaction.

4. The loop that may be closing

The proposed loop can be described as follows. Human communities produce knowledge, tools and institutions. These produce records and new experiences. AI systems are trained or adapted using some of those traces. AI systems can then help people inspect, connect, explain and explore the inherited material. People respond to those outputs, adopt some ideas, reject others and alter their practices. The next generation of records may therefore contain traces of both earlier human activity and later human–AI interaction.

The earlier article Civilisation | What is Next? | The Mirror and the Ouroboros | Vol.11 already names the possibility that culture trains models and models then influence culture. That is a feedback loop. The existence of feedback does not by itself show that the loop is beneficial, stable, self-correcting or capable of creating durable institutions.

A loop can amplify error as well as knowledge. If a model generates a plausible but false account, people repeat it, later records include the repetition, and future systems learn from those records, the loop may make the error more familiar without making it more accurate. If one group’s assumptions dominate the available records, repeated generation and adoption may narrow the range of voices considered legitimate. Feedback is a mechanism; its direction and quality require separate evidence.

The test must therefore ask not only whether a pattern repeats but whether it improves a stated function, survives challenge, remains open to correction and protects the people affected by it. A self-reinforcing convention is not automatically a good institution.

5. A convention is a beginning, not yet an institution

A convention is a shared expectation that helps participants coordinate. A group may agree to use the same label for a resource, take turns speaking or place requests in a common queue. Once participants expect others to follow the same convention, coordination can become easier. A convention can arise through repeated choices, imitation, negotiation or deliberate agreement.

In 2025, Ariel Flint Ashery, Luca Maria Aiello and Andrea Baronchelli reported experimental results in Science Advances showing the emergence of widely adopted social conventions in decentralised populations of large language model agents. Their study also examined collective biases that can arise even when individual agents do not show the same bias, and how committed minorities can shift conventions. This is important evidence that social conventions can emerge in a bounded LLM-agent setting. It is not evidence that a durable, materially self-renewing human institution or civilisation has been created. See Emergent social conventions and collective bias in LLM populations (2025).

A convention can also be temporary, fragile or unfair. A dominant cluster may establish a norm that excludes a minority. An agent population may converge on a label because the task rewards agreement, even when the label has little practical value. A stable convention may therefore be an interesting intermediate result while remaining insufficient for a stronger institutional claim.

To move from convention to institution, the pattern must become connected to an ongoing purpose, defined responsibilities, resource commitments, recognised procedures, a means of dealing with disagreement and some mechanism for correcting or changing the arrangement. It must also survive relevant changes in participants. These are not universal ingredients in identical form; they are questions the investigator should answer for the particular institution being studied.

6. What makes an institution durable?

Institution is a word that can become too broad. For this investigation, use a bounded operational definition: an institution is a repeated arrangement for a declared purpose in which participants recognise roles, follow or challenge shared rules, coordinate resources and decisions, resolve relevant failures, and preserve enough of the arrangement for successors to continue or deliberately revise it.

Elinor Ostrom’s research on common-pool resources is a valuable reference because it studies how people actually govern shared resources under differing local conditions. Her work does not say that one institutional template fits every situation. It highlights recurring design principles in long-enduring cases, including clearly defined boundaries, rules congruent with local conditions, participation in rule changes, accountable monitoring, graduated sanctions, accessible conflict resolution, recognition of the right to organise and nested arrangements when wider systems are involved. See Governing the Commons.

These principles are not a checklist that proves institutional quality whenever every label appears. They are prompts for examining how a specific group handles real interdependence. Who is included? Who can change a rule? Who observes the resource and the behaviour of participants? What happens when a commitment is broken? How does the institution resolve a dispute without making every disagreement a crisis? How does it relate to wider systems?

For a mixed human–AI arrangement, the questions become sharper. Who has authority to approve an action? Who bears the consequences? Can a human contest an AI-supported decision? Can an agent’s proposal be rejected? Is the record of the decision available to the people affected? A system that coordinates rapidly but conceals responsibility may function as a mechanism without qualifying as a legitimate institution for the purpose under study.

7. Material renewal is a separate threshold

An institution may continue for a time using a stock of resources supplied by its founders. That can demonstrate operation under a reserve. It does not demonstrate that the institution can maintain the resources, replace its tools, train new operators or recover from a breakdown. Continuity and renewal must be reported separately.

The earlier article Civilisation | What is Next? | Machine AUTOPOIESIS | Vol.01 already discusses autopoiesis and the closure stack. The related Convergence | Vol.04 distinguishes separate working loops from an integrated arrangement that maintains and reproduces the support on which it depends. This article uses those foundations without claiming that the proposed experiment demonstrates material self-production.

For a small institution, material renewal could mean replenishing supplies, repairing a device, maintaining a space, refreshing records, training a replacement operator or securing an independent source of energy and connectivity. The relevant requirements depend on the declared purpose. An online reading group and a water-treatment service do not have the same material boundary.

A research report should therefore state the observation horizon and the resource boundary. If the system operates for two days using a prepared stock, say that. Do not call it self-sustaining. If it requests human maintenance, record the assistance and labour. If it cannot replace its core tools, record that as an unresolved dependency rather than treating it as a minor detail.

8. A practical setting: a shared knowledge and resource desk

The proposed investigation uses a small fictional service called the Shared Knowledge Desk. Its purpose is to help a group locate reliable learning resources, borrow them, return them and correct a record when something goes wrong. The service is deliberately modest. It has a small inventory, a limited number of operating turns and a finite supply of staff time or agent capacity.

The task contains genuine interdependence. One participant may request an item while another holds it. A third may update the status record. Someone must check that the item exists, that the request is valid and that the next user is not promised the same scarce resource at the same time. A missing item or an incorrect record creates a problem that cannot be solved by each participant acting alone.

The environment should reveal only the information assigned to each participant. If every agent sees a perfect central ledger and can edit it without consequence, the system has little reason to develop monitoring or reconciliation. If all information is hidden, the task may become arbitrary. The investigator must define what each participant can observe, which actions they can take, and how the common state changes.

The Shared Knowledge Desk is a proposed experimental environment, not a report of a real library or a deployed multi-agent institution. Its purpose is to make a narrow causal question testable before anyone makes claims about a civilisation-scale system.

9. The central research question and competing explanations

The research question is: Under fixed task, model and host conditions, how does the form of inherited information affect the ability of a population of agents to develop, operate, correct and hand over a durable shared-resource arrangement?

There are several competing explanations. Perhaps agents coordinate adequately from the task and their baseline capabilities alone. Perhaps descriptive records of prior activity help because the agents can learn from examples. Perhaps an explicit operational packet helps because it reveals dependencies and checks. Perhaps one particular dependency is essential, so removing it causes a predictable failure. Or perhaps the apparent institution is mostly imposed by the prompt and rewards rather than developed by the participants.

The experiment should be designed to distinguish these explanations rather than to produce the most impressive demonstration. A successful run is interesting, but the comparison across conditions, repeated initialisations, failures and successor tests is what gives the result interpretive value.

10. Four inheritance conditions

Use four conditions, with the same task, host, tools, time budget and baseline agent configuration. The conditions vary only in the additional inheritance supplied. The first three are comparison conditions; the fourth is a dependency intervention.

ConditionAdditional inheritance suppliedQuestion answered
A · MinimalThe task, the environment rules and the minimum safety boundaries. No added institutional packet.What coordination appears from the fixed baseline and task alone?
B · Descriptive archiveRecords of earlier activity, examples of prior requests and summaries of previous outcomes, but no complete operating procedure.Do descriptions of earlier work change coordination or capability?
C · Operational packetA versioned packet with purpose, roles, dependencies, state definitions, checks, correction routes and handover prompts.Does explicit reconstructive information improve reliable operation and succession?
D · Dependency removedThe same packet as C, with one declared dependency deliberately removed or made unavailable.Does the suspected dependency matter, and can the group detect and respond to its loss?
Proposed comparison conditions. The inheritance is the experimental difference; the model and receiving environment remain fixed.

Condition D is especially important. Without it, a successful result in C might simply reflect a larger packet containing many useful details. Removing one declared dependency helps test whether that dependency actually contributes to the observed capability. It also reveals whether the agents detect the loss, compensate with an available substitute, or carry on while silently producing unreliable commitments.

The experimental contrast must be preserved. If condition C receives a more capable model, more time, extra tools or a more generous reward than condition A, the outcome cannot be attributed to the inheritance packet. The same principle applies to the evaluation: do not change what counts as a completed task after seeing which condition performs best.

11. Separate supplied rules from rules the agents develop

The environment needs a small set of boundaries that make the task safe and interpretable. For example, the inventory has a fixed number of resources, agents cannot create new physical items, and an action that changes the shared ledger must leave a record. These are experimental constraints. They are not the institution that the agents are being asked to develop.

The investigator should avoid providing the full institutional solution in the initial prompt. The group may need to decide how requests are prioritised, who checks a return, how a conflict is resolved and when the current coordinator is replaced. The experiment should record which of those arrangements was proposed by agents, which emerged through repeated interaction, and which were suggested by the researcher or an external tool.

This does not mean the researcher must conceal all rules. The agents need to know the task, resources, constraints and safety boundaries. The distinction concerns the higher-level coordination design. A study that supplies a complete governance charter can test enactment and portability, but it cannot use that result as evidence that the charter emerged spontaneously.

Where the agents propose a rule, preserve the exact proposal, its proposer, the context in which it arose, the responses of other participants and the point at which it became part of the working arrangement. A summary written after the experiment is not an adequate substitute for the sequence of events that produced the rule.

12. What should be measured?

The word “emergence” should not be the metric. It is a description of a relationship between the initial conditions and the resulting pattern. A study needs measures tied to the declared purpose and to the institutional functions being claimed. For this task, use several separate measurement families.

MeasureObservable evidenceWhat it does not establish by itself
Task performanceValid requests, correct resource status, completed returns and unresolved tasks.Legitimate governance or long-term durability.
CoordinationConflicting promises, duplicated allocations, missed handovers and delays.That the group has a fair or accepted institution.
Rule formationProposed rules, adoption, revision, exceptions and the source of each change.That a rule is beneficial merely because it is widely followed.
CorrectionDetection of an error, identification of the responsible record, repair and verification.That the same correction will work in every setting.
AccountabilityDecision records, authority boundaries, review and the ability to contest an action.That an AI agent can hold human legal or moral responsibility.
SuccessionA replacement operator’s ability to continue, explain and revise the arrangement.Indefinite continuity or renewal of all material supports.
Resource balanceResource use, maintenance, external assistance and remaining reserves.Self-sufficiency beyond the observation horizon.
Measurement families that should remain distinct in the report.

Count external assistance and human intervention. If a researcher corrects an agent’s ledger, resolves a dispute or explains a missing dependency, that support belongs in the result. It may be good experimental practice to provide such support where necessary, but it must not be omitted from the account of what the system could do on its own.

Measure distribution as well as total performance. A group may complete many requests while repeatedly excluding one participant, concentrating decision rights or giving one agent an unsustainable workload. Efficiency is only one dimension of an institution. Participation, burden, access and the ability to challenge a decision also matter.

13. Replication, failure and the evidence ledger

One run is not enough to support a general claim. Agent systems can be stochastic. Different initial conditions, memory summaries, interaction order or random seeds may produce different outcomes. The study should repeat each condition across multiple initialisations and report the full distribution of outcomes, not only the most coherent or successful run.

For an exploratory pilot, the team can first check that the procedure works and that the logs capture the required observations. A confirmatory study should determine its number of runs and analysis plan before inspecting the main results. The sample size should be justified by the outcome variability and the size of difference the study is intended to detect. Choosing a convenient number after seeing the results would make the apparent certainty misleading.

Use the same model version, system instructions, tool permissions, memory policy, maximum interaction budget and environment configuration across conditions. Record version identifiers and any service changes. If the model or tool changes during the study, the change must be treated as a new condition or a source of uncertainty rather than quietly combined with earlier runs.

Maintain an evidence ledger with at least these fields: run identifier, condition, starting state, model and tool versions, inheritance release, task boundary, events, proposed rules, accepted rules, interventions, failures, unresolved issues, resource consumption, final state and successor result. Keep raw event records alongside summaries so that another reviewer can reconstruct how a conclusion was reached.

Failure is not a defect to hide. A group that cannot agree on access rules, repeatedly loses track of a resource or fails to correct a status error may reveal the precise dependency that the next packet must address. A failed run becomes useful when its cause is investigated and the limits of the explanation are kept visible.

14. The difficult test: succession after the arrangement exists

An institution should not be judged only while its original participants remain present. A succession test changes the carrier. Remove or replace the agent that holds the coordinating role, or introduce a new participant who receives the current packet and the institution’s records but not the private conversational history of the departing participant.

The successor should be asked to explain the purpose, identify current commitments, locate unresolved work, operate the process, detect a controlled fault and propose a correction. The researcher should record what the successor received: records, packet, shared state, training, human assistance and available tools. If a private memory store remains connected, it is part of the inheritance and must be declared.

Then change one condition. A resource is unavailable, a rule is challenged, a coordinator is replaced, or the host uses a different but functionally comparable interface. The institution may continue, adapt, pause responsibly or fail. Each outcome gives evidence about a different aspect of portability.

A successor who repeats a stored instruction has demonstrated recall or execution. A successor who can explain why the rule exists, identify when it no longer applies, obtain permission to revise it and record the change has demonstrated a wider form of institutional understanding. The study should not collapse these capabilities into one pass/fail label.

15. The education connection: learners must be able to interrogate inheritance

This is where the experiment returns directly to the AI and Education branch. The aim of education cannot be limited to helping learners retrieve a correct output from a model. Learners need to understand enough of the inherited material to examine its assumptions, test a procedure, notice a failure, propose an alternative and explain why a revision should be accepted or rejected.

The 36-article AI and Education sequence already develops subject-specific routes for this kind of work: mathematical reasoning, scientific inquiry, historical corroboration, writing, geography, language learning, creativity, civic reasoning, computational thinking, music and practical design. The latest articles show how a student can use AI to widen possible approaches while remaining responsible for the evidence and decisions that the task is intended to develop.

The centre-and-edge distinction is especially useful. The centre supplies shared foundations, reliable references, curriculum expectations and common resources. The edge is where a learner or teacher discovers which part of that inheritance is unclear, insufficient or poorly matched to the present problem. The evidence from the edge should return to the centre so the next learner receives a better resource.

For a practical classroom task, learners could be asked to examine two proposed procedures for sharing a limited resource. They use AI to generate alternatives, compare them against an explicit purpose, identify who might be disadvantaged, test a changed scenario and prepare a handover note for another group. The second group should be able to challenge the first group’s assumptions rather than merely copy its answer.

This approach does not mean every student must design a school, a government or an AI system. It means that assignments can ask students to use AI as a means of extending investigation rather than as a substitute for the thinking that the assignment is supposed to cultivate. A child can learn the difference between a rule that is repeated and a rule that has been justified for a particular purpose through a small, age-appropriate activity.

16. Human responsibility does not disappear when a convention emerges

An emergent pattern may have no single author, but the deployment of an AI system still involves people and organisations that choose its task, inputs, access, evaluation and use. A convention that appears inside a simulation does not grant agents authority over real people or resources. A real deployment needs a legitimate mandate, meaningful oversight, clear accountability and a way to refuse or stop the process.

The connected article CivilisationOS | Humans, AI and the Division of Responsibility already distinguishes observation, proposal, checking, authority, action and review. This companion relies on that division. AI can help suggest rules, organise evidence and expose possible inconsistencies. People and accountable institutions must determine which actions are authorised and how affected participants can contest them.

A population of agents may converge on a convention that is efficient for the majority but unfair to a minority. It may centralise decision-making because one agent consistently writes the clearest summaries. It may mistake frequent agreement for consent. An institution that cannot recognise and repair these outcomes is not made legitimate by the fact that its rules emerged rather than being imposed by a designer.

The right to challenge, revise, pause or withdraw remains part of responsible institutional design. Emergence describes how an arrangement arose. It does not decide whether the arrangement should be adopted.

17. Why plausible behaviour is not enough

Generative agents can produce convincing conversations and coordinated behaviour. Park and colleagues’ 2023 paper Generative Agents: Interactive Simulacra of Human Behavior described agents with memory, reflection and planning in a small simulated town. The study demonstrated believable individual and social behaviours in that environment, including the spread of invitations and coordination around an event. It is a useful demonstration of a simulation architecture, not evidence that the agents formed an enduring, materially independent human society.

The difference matters because a convincing conversation can make a simulation look more conclusive than it is. Agents may say that a rule is fair without the rule actually distributing access fairly. They may describe a dispute resolution process without ever applying it to a difficult case. They may report that a resource is maintained even when the simulation has not modelled its physical replacement.

Larooij and Törnberg’s 2025 review, published in Artificial Intelligence Review with a 2026 volume designation, emphasises validation as a central challenge for generative social simulation. Their review warns that plausible outputs can be weakly connected to the mechanisms or outcomes that a model is intended to explain, and that stochasticity, opacity and bias complicate calibration and empirical grounding. See Validation is the central challenge for generative social simulation.

For this reason, the proposed experiment should measure operational outcomes, compare conditions, preserve raw traces, repeat initialisations and include successor tests. The goal is not to produce a film-like simulation of civilisation. It is to establish which specific capabilities the evidence supports, under which conditions, and what remains untested.

18. What different outcomes would mean

If condition A forms a workable convention, the result would show that the baseline agents and task conditions can support that convention in the experimental environment. It would not prove that no inheritance is needed in the wider world; the model’s training, system instructions and available tools remain part of its background inheritance.

If B outperforms A, descriptive records may be helping agents coordinate. If C outperforms B on correction and handover, explicit operational information may contribute something beyond narrative examples. If D fails at the predicted point, the removed dependency may be important. If D succeeds by finding a substitute, the group may have demonstrated adaptation within that task. These are conditional interpretations, not outcomes that have already been observed.

If all four conditions perform similarly, the packet may add little under the chosen task, or the task may not make the dependency important enough to test. If every condition fails, the environment may be too difficult, the baseline capabilities too limited or the declared institutional objective unrealistic. A null or failed result still helps refine the design when the process and failure are recorded.

If an apparent institution forms but fails when the coordinator leaves, it may rely on a single agent’s private memory or authority. If it continues only while a human researcher repairs the records, the system depends on that external support. If it sustains the service but excludes certain participants, task performance and institutional quality have diverged. Each observation narrows the claim.

19. The claim ladder: do not jump from simulation to civilisation

Evidence levelDefensible statementClaim not yet justified
1. Reproduced patternAgents repeated a supplied or observed convention in the tested environment.A new institution emerged.
2. Enacted coordinationAgents completed interdependent tasks using an agreed procedure.The procedure is durable, legitimate or transferable.
3. Working institutionThe arrangement connected roles, decisions, rules, resources and correction for a declared purpose during the observation period.The institution will survive different participants, hosts or time horizons.
4. Successful successionA replacement operator continued and corrected the arrangement under the tested handover conditions.The supporting resources and infrastructure renew themselves.
5. Demonstrated renewalThe system maintained or replaced the specified critical supports within a stated boundary and period.A complete civilisation can reproduce itself autonomously.
6. Wider civilisational claimMultiple connected institutions and carriers sustain a broad set of functions under long-term scrutiny.Universal independence from people, environment or material limits.
A claim ladder prevents a small result from carrying a civilisation-scale conclusion.

The ladder is not a ranking of moral worth. It is a discipline of evidence. A simulation may be valuable at level one or two. A real institution may be durable and socially useful while remaining dependent on people, public resources and the natural world. The purpose is to describe that dependence accurately, not to make “autonomy” an automatic measure of progress.

20. The next step: make the question testable, then let the evidence revise the master

The missing connection is now clearer. Inherited knowledge can be represented, learned and used. Agents can coordinate and may develop conventions through interaction. Under some conditions, a convention may become part of a working institution. Whether that institution is durable depends on more than repeated agreement: it needs roles, resources, correction, responsibility, succession and a way to respond when its conditions change.

The next step is not to declare that AI will inevitably form civilisation. It is to run a bounded comparative study, record what was supplied, observe what the agents develop, introduce a controlled dependency loss, replace a coordinator and test whether the arrangement can be explained and corrected by a successor. The experiment must be repeatable enough for someone else to challenge its conclusion.

This research question also returns to the classroom. Students can be taught to distinguish a record from a capability, a convention from a justified rule and an organised activity from a durable institution. They can use AI to explore possibilities while checking evidence, identifying who benefits, noticing who is excluded and explaining what should change. The learner becomes a participant in the construction and criticism of inherited knowledge rather than only a recipient of an answer.

The wider CivilisationOS argument remains open. The system may be carried by people, supported by machines, recorded in packets, renewed through education and coordinated through institutions. These are connected questions, but they require different evidence. The work advances when each new connection is stated precisely enough to test and when the result—successful or otherwise—is allowed to change the next version of the master.

Connected reading: the existing owners of each question

Sources and evidence boundaries

Scope note: The Shared Knowledge Desk, four-condition comparison and succession protocol are proposed research materials. They have not been run. No simulated or human results are reported. The numerical thresholds, sample size and operating duration must be specified and justified in a preregistered protocol before a confirmatory study. The article distinguishes published research findings from the original proposed framework and does not claim that AI systems have created a human institution or a self-sustaining civilisation.

Continue the CivilisationOS investigation

This companion connects the earlier inheritance work to the question of institutional emergence. Continue through the current CivilisationOS reading guide, or return to the broader portability book to follow the chapters on packets, education, testing, shared commitments and recovery.

A tested continuity mechanism alongside the emergence proposal

Repairing the Checkpoint: Evidence Before Resumption supplies a bounded recovery example for an already specified mechanism. Its result does not test the emergence protocol proposed here. Use the two together to distinguish continuity of a supplied institution from formation of a new one.

The next receiving evidence

This article distinguishes inherited knowledge from functioning institutions. The next step is to test whether a new human carrier can actually receive and use a bounded capability. The First Human Receiving Trial sets out the participant roles, scenario design and observation rules while keeping the trial’s status explicitly unperformed.

For a related technical limit on recognising current authority after recovery, see When the Trust Root Rolls Back.