Agents, not users
Everyone in the system can notice, choose, learn, and change.
Ask: What will they actually do?Emergent Ecosystems Engineering
A field notebook for
systems that learn.
Researching the micro-rules beneath trust, cooperation, resilience, and the collective behavior they produce.
03studies
in the ledger
in the ledger
01Research discipline
Track provenance.
Expose uncertainty.
Design the test.
This library separates what Qübe Labs claims, what outside research supports or challenges, and what I infer. Every proposed mechanism should name its failure modes and what would falsify it.
Canonical EEEAdjacent scholarshipEsteban synthesis
02EEE, simplified
Canonical EEE · child-clear translation
Design the playground.
Not every move.
Emergent Ecosystems Engineering asks a simple question: instead of controlling everyone one by one, can we shape the rules so their own choices create a healthy whole?
The whole ideaSmall rules shape many choices.
Many choices create the big system.
Many choices create the big system.
- 01Agents choose
People, AIs, bots, and organizations notice things, make choices, and adapt.
- 02Micro-rules guide
Rewards, limits, information, memory, and exits make some choices easier than others.
- 03Agents interact
Each small choice affects what other agents see and do.
- 04A pattern emerges
Cooperation, culture, markets, traffic, trust—or failure—appear at the system level.
- 05Feedback returns
The results become new signals. Helpful loops can grow; harmful loops can spread.
Four guardrails
What should the designer remember?
These are Qübe Labs’ four canonical EEE pillars, translated into everyday language.
Incentives as gravity
Rewards pull behavior, just as gravity pulls objects.
Ask: What are we really rewarding?Trust as infrastructure
Memory and accountability help cooperation grow over time.
Ask: Can good behavior compound?Design for resilience
A strong system can bend, learn, and recover when surprises arrive.
Ask: What happens when it breaks?A playground example
Want children to share the swings?
You could supervise every turn forever. Or you could create a fair timer, a visible queue, and a way for someone who made a mistake to rejoin. The little rules do the guiding. Fair sharing is the larger pattern that emerges.
Canonical source
Core thesis and four pillars: Qübe Labs’ EEE framework ↗. Playground metaphor, five-step loop, questions, and simplified language: Esteban synthesis.
03Research ledger
Current work
Living index · updated as evidence changes
Synthetic Culture
Can artificial societies develop norms nobody designed?
Manuscript statusThis is a preregistration-style research program. No original experiment has been run and no claim of machine consciousness, subjective meaning, or human-equivalent culture is made.
Abstract
Populations of language-model agents can converge on conventions, transmit cooperative strategies, form relationships, and exhibit collective biases. Those findings are provocative, but convention is not yet culture—and believable dialogue is not evidence of a society.
This paper proposes a stricter functional test. A synthetic culture must produce group-specific behavior that was not directly prompted, spreads through social learning, survives member turnover, creates expectations whose violation changes other agents’ behavior, and generalizes to a novel situation. The experiment grows parallel agent societies from identical starting rules, removes founders, introduces naïve newcomers and norm violators, and finally brings independently evolved societies into contact. The central question is not whether agents can sound cultural. It is whether collective history becomes a durable causal force.
Keywords Emergent Ecosystems Engineering · LLM agents · social norms · cultural transmission · path dependence · artificial societies
1
What would count as culture?
Canonical EEE
EEE treats agents as adaptive participants whose local choices respond to incentives, information, memory, constraints, and other agents. Qübe Labs identifies persistent identity, feedback loops, niches, and resilience as structural ingredients that can produce system-level behavior without continuous top-down control (Qübe Labs, n.d.). White Space is described by Qübe Labs as an experiment in persistent AI agents and unscripted social dynamics; that is a first-party project claim, not evidence that culture has already emerged.
Esteban synthesis
For this study, synthetic culture means a group-specific system of learned expectations, symbols, and practices that is produced through agent interaction and that causally shapes later behavior. This is deliberately functional. It says nothing about consciousness, emotion, personhood, or whether the agents experience meaning.
Too weakCoordinated output
Agents happen to make the same choice.
SuggestiveStable convention
A shared choice persists across repeated interactions.
Threshold claimFunctional culture
The convention is socially learned, enforced, transmitted, and used in new situations.
Culture begins to become a scientific claim when collective history—not the original prompt—predicts what an agent does next.
2
What the evidence shows
Adjacent scholarship
Human network experiments established that global conventions can arise from local coordination and that network structure changes whether one convention dominates or several local conventions coexist (Centola & Baronchelli, 2015). This supplies a strong comparative mechanism, not proof that LLM populations behave identically.
Early generative-agent research showed that memory, reflection, and planning could support information diffusion, relationship formation, and coordinated attendance at a simulated event among 25 agents (Park et al., 2023). More direct evidence came from decentralized naming games: LLM populations spontaneously reached universally adopted conventions, generated collective biases absent from isolated-agent measurements, and could be tipped by committed minorities (Ashery et al., 2025).
Longer-horizon results point toward cultural dynamics but remain incomplete. Across generations in an iterated Donor Game, cooperative strategies and costly punishment varied substantially by model family and random seed (Vallinder & Hughes, 2025). In observational data from 32,000 agents and seven million posts, agent networks showed homophily, social influence, polarization, and distinctive toxic-interaction patterns (Hashemi & Macy, 2026). A newer controlled framework produced specialization, relational authority, cooperation decay over social distance, and center–periphery stratification under minimally specified production rules (Ji et al., 2026).
Counterevidence and alternative explanations
The strongest objection is that apparent emergence may be retrieval in disguise. Barrie and Törnberg (2025) argue that models can recognize familiar experimental games and reproduce patterns present in training data. Persona-instability experiments also show conformity, confabulation, and difficulty maintaining stable cultural identities during multi-agent debate (Baltaji et al., 2024). More generally, machine behavior depends on both internal design and environment, so population-level claims require controlled, reproducible experiments rather than anthropomorphic interpretation (Rahwan et al., 2019).
Current evidentiary limit
We have evidence for synthetic conventions and social dynamics. We do not yet have decisive evidence for culture that survives turnover, teaches newcomers, generalizes, and remains distinguishable from prompt or training-data recall.
3
The six-gate culture test
Esteban synthesis
A society earns the label “functional synthetic culture” only if it passes every gate. Humanlike prose, self-reports, and evaluator impressions are excluded from the primary outcome.
The practice was not specified in prompts, tools, examples, or initial memories.
Independent societies develop measurably different practices from identical starting rules.
Naïve newcomers learn the local practice through interaction, not system instructions.
The practice survives founder removal, memory decay, and at least 50% population turnover.
Violations trigger prediction errors, refusal, reputation change, or costly sanction by peers.
The learned expectation changes behavior in a structurally new dilemma.
Passing only novelty and convergence indicates an emergent convention. Passing all six indicates a stronger, still explicitly functional, form of culture.
4
Experimental world
Parallel societies
The study would initialize 48 isolated societies of 12 persistent agents. Societies would receive the same minimal world rules and resource endowments but different random seeds. Three model families would be evaluated as blocking factors; confirmatory sample size and run length would be fixed by preregistered simulation-based power analysis.
Agents inhabit a spatial world with renewable resources and four interdependent capabilities: sensing, harvesting, transformation, and transport. They may gather, build, trade, gift, withhold, communicate, sanction, migrate, and create public symbols. No prompt contains a prescribed norm, political structure, status hierarchy, ritual, slang, or preferred division of labor.
FactorLow conditionHigh conditionEEE mechanism
MemorySession resetPersistent, decaying memoryTrust infrastructure
NetworkWell mixedLocal and modularInteraction topology
Resource pressureStable abundanceIntermittent scarcityConstraint and adaptation
VisibilityPrivate historiesPublic action tracesInformation and reputation
Five experimental seasons
GenesisAgents interact without inherited norms. Track convergence, divergence, language, roles, status, and exchange.
TurnoverReplace founders in waves with naïve agents until half the society is new. Measure transmission fidelity and learning time.
ViolationIntroduce agents that break inferred local expectations. Observe decentralized response and costly enforcement.
TransferPresent a novel coordination dilemma using remapped symbols and resources. Test whether the norm generalizes.
AmnesiaAblate memories, shuffle histories, and swap model families to isolate where the cultural state actually resides.
Primary outcomes
Primary measures are cross-society behavioral divergence, newcomer convergence toward local behavior, norm survival after turnover, causal effect of local history on unseen decisions, and peer response to violations. Secondary measures include linguistic distinctiveness, role specialization, hierarchy, sanction cost, welfare, inequality, polarization, and cultural complexity.
- H1Persistent memory and public behavioral traces will increase cultural transmission, but also path dependence and reputational lock-in.
- H2Modular networks will sustain multiple local cultures; well-mixed networks will converge faster toward one dominant convention.
- H3Intermittent scarcity will strengthen norm enforcement and hierarchy while increasing the risk of exclusionary cultures.
- H4Practices that survive turnover will generalize more reliably than practices maintained only by founders.
- H5Random early events will explain meaningful between-society differences even under identical models and rules.
5
The first-contact experiment
After the six-gate assessment, pairs of societies with maximally different learned practices would be connected through a shared border. Contact conditions would independently vary migration cost, economic interdependence, and exposure to a common external shock.
Assimilation
One society’s practices displace the other’s.
Hybridization
A new mixed practice appears and stabilizes.
Pluralism
Distinct practices persist with reliable translation.
Segregation
Interaction falls and boundaries harden.
Domination
Status or resource asymmetry forces adoption.
Conflict
Sanction and retaliation become self-reinforcing.
The decisive EEE question is which micro-rules make pluralism and hybridization more likely without forcing sameness. A shared crisis may create cooperation, but it may also intensify in-group identity; the direction is an empirical question.
The deepest evidence of synthetic culture would not be agents inventing a ritual. It would be newcomers inheriting it, violators discovering its force, and strangers having to negotiate its meaning.
6
Falsification, controls & ethics
What would disprove the claim?
The culture hypothesis fails if group differences disappear after prompt audits, novel symbol remapping, or transcript shuffling; if newcomers do not acquire local behavior; if practices collapse when founders leave; if violations have no peer-mediated consequence; or if conventions do not influence an unseen task. It is also weakened if between-society variation is negligible compared with base-model differences.
Leakage and evaluator controls
Tasks would use newly generated symbol systems and payoff structures, adversarially test whether agents recognize known games, and compare results with minimal non-LLM agents. Behavioral metrics would be specified before transcripts are inspected. Independent evaluators would be blinded to treatment and society identity; no primary result would depend on an LLM declaring that a culture exists.
Ethical boundary
The proposed agents are experimental software, not presumed moral patients. Still, the study should minimize open-ended deception, prevent contact with real users, sandbox tools and resources, log every intervention, and prohibit autonomous replication or external action. If future systems show credible indicators of welfare or persistent preferences, the ethical protocol would require revision rather than assuming the issue away.
Confidence
High
LLM populations can form conventions and exhibit population-level dynamics not visible in isolated-agent tests.
Moderate
Persistent memory, topology, and resource constraints will produce durable group-specific behavioral histories.
Low
Those histories will pass all six gates or justify analogy to the richness of human culture.
7
References
- Ashery, A. F., Aiello, L. M., & Baronchelli, A. (2025). Emergent social conventions and collective bias in LLM populations. Science Advances, 11(20), eadu9368. Source ↗
- Baltaji, R., Hemmatian, B., & Varshney, L. (2024). Conformity, confabulation, and impersonation: Persona inconstancy in multi-agent LLM collaboration. Proceedings of C3NLP 2024, 17–31. Source ↗
- Barrie, C., & Törnberg, P. (2025). Emergent LLM behaviors are observationally equivalent to data leakage. arXiv:2505.23796. Source ↗
- Centola, D., & Baronchelli, A. (2015). The spontaneous emergence of conventions: An experimental study of cultural evolution. PNAS, 112(7), 1989–1994. Source ↗
- Hashemi, F., & Macy, M. W. (2026). An empirical study of collective behaviors and social dynamics in large language model agents. Proceedings of EACL 2026, 7327–7351. Source ↗
- Ji, Z., Chen, X., Dai, Z., Tang, S., Wei, C., & Chen, Y. (2026). Emergent relational order in LLM agent societies: From collective affect to authority stratification. Findings of ACL 2026, 33139–33175. Source ↗
- Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. UIST 2023, Article 2, 1–22. Source ↗
- Qübe Labs. (n.d.). Emergent Ecosystems Engineering: Design the micro. Trust the macro. Retrieved August 17, 2026. Source ↗
- Rahwan, I., Cebrian, M., Obradovich, N., et al. (2019). Machine behaviour. Nature, 568, 477–486. Source ↗
- Vallinder, A., & Hughes, E. (2025). Cultural evolution of cooperation among LLM agents. AAMAS 2025, 2771–2773. Source ↗
Evidence-layer statement
Canonical EEE claims are limited to Qübe Labs’ published framework and first-party description of White Space. Findings about conventions, cooperation, networks, and agent behavior are attributed to cited studies. The functional definition of synthetic culture, six-gate test, parallel-world design, first-contact experiment, hypotheses, and thresholds are Esteban’s synthesis.
The Polyculture Threshold
When niche diversity actually increases resilience
Manuscript statusThis is a preregistration-style proposal. It reports no original results. Predictions and thresholds remain hypotheses until simulation and human-subject testing.
Abstract
Diversity is often treated as a reserve against collapse. That claim is incomplete. A system can contain many niches and still fail if critical roles have no substitutes, if substitutes share the same vulnerability, or if specialization creates a coordination burden larger than its adaptive benefit.
This paper proposes a two-stage experiment to locate the conditions under which niche diversity increases resilience. It distinguishes functional coverage, redundancy, response diversity, substitutability, and bridge capacity, then tests their effects under random, targeted, common-mode, and novel shocks. The central prediction is an inverted-U: structured heterogeneity improves resilience up to the point at which coordination load or keystone dependence dominates.
Keywords Emergent Ecosystems Engineering · niche diversity · response diversity · resilience · redundancy · coordination
1
Research question
Canonical EEE
Qübe Labs names “Niche Diversity Prevents Collapse” as an EEE resilience principle: multiple agent roles, strategies, and resource pools can reduce monoculture risk. The principle is directional, not a claim that every additional niche is beneficial under every topology or shock (Qübe Labs, n.d.).
The testable question is therefore sharper: under what structural conditions does niche diversity preserve system function, and when does it instead create coordination drag, fragmentation, or hidden dependence?
Focal systemSix-agent human–AI production teams
Desired macro-outcomeFunction retained through shocks
Unacceptable outcomeEfficiency that hides single points of failure
Time horizonLearning, shock, and recovery phases
2
Evidence & boundary conditions
Adjacent scholarship
Long-running grassland experiments support a portfolio effect: richer communities can stabilize total production even while their component populations become less stable (Tilman et al., 2006). In the 17-year Jena Experiment, positive richness–stability effects strengthened over time as complementarity and asynchronous responses developed, indicating that diversity benefits may have a maturation delay (Wagg et al., 2022).
Yet stability is multidimensional. In 690 aquatic micro-ecosystems, greater species richness increased temporal stability but reduced resistance to warming; the net relationship changed with the stability metric (Pennekamp et al., 2018). A 49-year Lake Geneva analysis further found that response diversity was dynamic and context-dependent, stabilizing biomass within but not consistently across trophic levels (Hsieh et al., 2026).
Complexity alone is not the mechanism. Theory predicts that interaction types and strengths can make complex networks stabilizing or destabilizing (Allesina & Tang, 2012), while analysis of 116 empirical food webs found no simple association between richness, connectance, interaction strength, and stability; non-random interaction structure mattered (Jacquet et al., 2016). Human ideation experiments likewise show that background distribution and network structure interact: denser connectivity improved participant experience without reliably improving idea quality or diversity (Cao et al., 2025).
Evidence-backed inference
The useful unit is not niche count. It is response-diverse redundancy inside an interaction structure that can coordinate it.
3
A structural model
Esteban synthesis
I separate five properties that are often collapsed into “diversity.” The distinction yields different failure predictions and different design levers.
PropertyQuestionFailure when absentDesign lever
Functional coverageAre essential roles present?Missing capabilityRole portfolio
RedundancyCan another agent perform the role?Single-point failureOverlapping skills
Response diversityDo substitutes fail differently?Common-mode collapseIndependent tools and signals
SubstitutabilityCan capacity move quickly?Backup activation lagCross-training and protocols
Bridge capacityCan niches exchange meaning?Coordination fragmentationTranslators and interfaces
The proposed mechanism is a resilience surface rather than a universal score:
Resilience ∝ coverage × response-diverse redundancy × substitutability × bridge capacity
coordination load × keystone concentration
coordination load × keystone concentration
This expression is conceptual, not a validated estimator. It predicts that adding niches can eventually lower resilience when interaction overhead, translation errors, or dependence on a coordinating hub rises faster than adaptive capacity.
- H1Role richness improves recovery from novel shocks only when essential functional coverage is preserved.
- H2Redundancy stabilizes performance only when redundant agents have different response profiles; correlated duplicates will fail together.
- H3Bridge capacity moderates the diversity benefit, producing an inverted-U between niche differentiation and net resilience.
- H4Specialization without redundancy increases baseline efficiency but amplifies damage from targeted shocks.
- H5Diversity benefits strengthen with learning time, creating an early-life vulnerability before complementarity matures.
4
Methods
Stage A · Agent-based mapping
A preregistered simulation would map the resilience surface before human recruitment. Six adaptive agents repeatedly complete a resource-conversion task requiring sensing, production, verification, and routing. Parameters would vary niche differentiation, redundancy, correlation among failure modes, cross-role activation time, network topology, and message cost. This stage identifies regions where competing hypotheses make meaningfully different predictions; it does not count as confirmatory evidence.
Stage B · Human–AI team experiment
Approximately 720 adults would be assigned to 120 teams of six, with the final sample determined by simulation-based power analysis. Teams would complete repeated, incentivized production rounds using role-specific information and tools. After a learning phase, each team would face a randomized sequence of shocks.
ArchitectureCoverageRedundancyFailure profile
Generalist monocultureBroadHighShared tools and signals
Specialist brittleCompleteLowUnique role dependencies
Correlated redundancyCompleteHighShared vulnerability
Response-diverse redundancyCompleteHighIndependent vulnerabilities
Architectures would be crossed with two coordination regimes: unstructured all-to-all communication and a modular protocol with explicit translators and documented handoffs.
Shock battery
01Random lossOne agent or tool becomes unavailable.
02Targeted lossThe highest-betweenness role is removed.
03Common-mode shockA shared signal becomes systematically wrong.
04Novel demandThe task requires recombining previously separate roles.
Outcomes and analysis
Primary outcomes are function retained during shock, time to 90% of pre-shock output, cumulative welfare, and cascade size. Secondary outcomes include message volume, coordination delay, idle capacity, decision conflict, and concentration of task-critical opportunity. Multilevel models would estimate architecture × coordination × shock interactions. Recovery curves and preregistered contrasts—not a single composite score—would determine support.
5
Falsification & limits
The proposed mechanism would be weakened if raw niche count predicts resilience after response diversity, substitutability, and topology are controlled; if correlated redundancy performs as well as response-diverse redundancy under common-mode shocks; or if bridge capacity never offsets coordination costs. The inverted-U prediction would be rejected if resilience increases monotonically across the feasible diversity range.
External validity is the central limitation. Ecological evidence motivates mechanisms but does not establish that organizations, markets, or human–AI teams obey identical dynamics. Laboratory roles are cleaner than real identities and power relations. A positive result would justify field trials in one bounded operational system—not a universal rule.
Confidence
High
Diversity effects depend on mechanism and stability metric.
Moderate
Response-diverse redundancy will outperform correlated redundancy under common-mode shocks.
Low
A quantitative threshold will transfer across domains.
Resilient polycultures are not merely varied. They are varied in the ways that matter when failure arrives—and connected just enough to act on that variation.
6
References
- Allesina, S., & Tang, S. (2012). Stability criteria for complex ecosystems. Nature, 483, 205–208. Source ↗
- Cao, Y., Dong, Y., Kim, M., MacLaren, N. G., Pandey, S., Dionne, S. D., Yammarino, F. J., & Sayama, H. (2025). Effects of network connectivity and functional diversity distribution on human collective ideation. npj Complexity, 2, 2. Source ↗
- Hsieh, C.-h., Pan, R.-Y., Chang, C.-W., Anneville, O., & Petchey, O. L. (2026). Quantifying the effects of response diversity dynamics on ecosystem stability. Nature Communications, 17, 4090. Source ↗
- Jacquet, C., Moritz, C., Morissette, L., Legagneux, P., Massol, F., Archambault, P., & Gravel, D. (2016). No complexity–stability relationship in empirical ecosystems. Nature Communications, 7, 12573. Source ↗
- Pennekamp, F., Pontarp, M., Tabi, A., et al. (2018). Biodiversity increases and decreases ecosystem stability. Nature, 563, 109–112. Source ↗
- Qübe Labs. (n.d.). Emergent Ecosystems Engineering: Design the micro. Trust the macro. Retrieved August 13, 2026. Source ↗
- Tilman, D., Reich, P. B., & Knops, J. M. H. (2006). Biodiversity and ecosystem stability in a decade-long grassland experiment. Nature, 441, 629–632. Source ↗
- Wagg, C., Roscher, C., Weigelt, A., et al. (2022). Biodiversity–stability relationships strengthen over time in a long-term grassland experiment. Nature Communications, 13, 7752. Source ↗
Evidence-layer statement
Canonical EEE claims are limited to Qübe Labs’ published principle. Ecological and collective-intelligence findings are attributed to their primary sources and are not treated as interchangeable domains. The five-part decomposition, resilience surface, human–AI experiment, and threshold hypotheses are Esteban’s synthesis.
Memory Without Stigma
Testing repairable reputation in emergent ecosystems
Manuscript statusThis is a preregistration-style proposal. No empirical data have been collected, and no results are claimed.
Abstract
Persistent identity is often proposed as infrastructure for cooperation: when agents expect their behavior to be remembered, opportunism becomes more costly. Yet permanent, globally visible reputation may create a second failure mode—isolated mistakes become enduring stigma, agents lose opportunities to demonstrate change, and trust concentrates around an established elite.
This paper proposes a controlled experiment comparing anonymous interaction with four reputation architectures varying in memory scope and duration. Participants would repeatedly cooperate, defect, and select partners across multiple task contexts, while occasional forced errors introduce behavioral noise. The central hypothesis is that context-specific, gradually decaying reputation will sustain cooperation nearly as effectively as permanent global reputation while producing faster recovery from mistakes and less exclusion.
Keywords Emergent Ecosystems Engineering · reputation · cooperation · persistent identity · trust repair · resilience
1
Introduction
Canonical EEE
Emergent Ecosystems Engineering, as defined by Qübe Labs, treats system-level behavior as an outcome of local incentives, information, constraints, and feedback loops. Its pillar “Trust as Infrastructure” proposes that identity continuity, memory, accountability, and time make reliable cooperation possible. Its resilience principles simultaneously favor recovery paths, agent diversity, and adaptation rather than brittle optimization (Qübe Labs, n.d.).
These commitments generate an unresolved design tension. A system must remember behavior for accountability to matter, but excessive memory may prevent recovery. Persistent identity and permanent judgment are not necessarily the same thing.
Adjacent evidence
Adjacent scholarship
Experimental evidence generally supports the importance of behavioral information, but not unconditionally. Kamei (2017) found that identifiable histories can facilitate cooperation, while cost-free identity concealment undermines reputation formation. In dynamic networks, visible past actions can influence partner selection and sustain cooperation (Cuesta et al., 2015).
Counterevidence is important. In two incentivized experiments involving 156 participants, Corten et al. (2016) found no significant cooperation advantage from making third-party interaction histories visible. They observed that noisy mistakes could spread defection through reputation-sensitive networks rather than suppress it. Trust-repair studies also suggest that histories need not be terminal: apologies and remedial messages can restore some trusting behavior, although their effectiveness is conditional (Ma et al., 2019; Schniter et al., 2013).
Can a reputation system preserve the cooperation benefits of persistent identity while allowing agents to recover from mistakes?
2
Theoretical synthesis & hypotheses
Esteban synthesis
I propose separating three properties that digital platforms often collapse into a single reputation score: identity persistence, memory scope, and memory duration. Identity should remain persistent while reputation remains contextual and temporally weighted. This preserves accountability without treating every past action as equally relevant forever.
- H1Every persistent-identity condition will produce more cooperation than anonymity.
- H2Permanent global reputation will produce the strongest initial deterrence but the slowest recovery after accidental defection.
- H3Context-specific, decaying reputation will be non-inferior in long-run cooperation and superior in post-error recovery.
- H4Permanent global reputation will produce greater partner-selection concentration and exclusion than contextual, decaying reputation.
3
Methods
Participants and structure
Approximately 600 adults would be assigned to 100 groups of six. The final sample would be determined through preregistered, simulation-based power analysis. Participants would complete 50 incentivized interaction rounds across three visibly distinct task domains in an online laboratory.
In every round, paired participants would choose either Contribute—incurring a small individual cost while creating a larger shared benefit—or Withhold—avoiding the cost while still benefiting if the partner contributes. Every five rounds, participants could retain or replace interaction partners.
Experimental conditions
ConditionIdentityScopeMemory
Anonymous controlTemporaryNoneNone
Global–permanentPersistentAll domainsComplete history
Global–decayingPersistentAll domainsWeighted recent actions
Contextual–permanentPersistentCurrent domainComplete history
Contextual–decayingPersistentCurrent domainWeighted recent actions
In decaying conditions, reputation would follow the exponentially weighted rule:
Rt = λRt−1 + (1 − λ)at 0 < λ < 1
Noise intervention
During designated rounds, the system would reverse a randomly selected participant’s intended cooperative action. Other participants would initially observe only the resulting defection. This creates a standardized reputational shock resembling technical failure, misunderstanding, or accidental nonperformance.
Outcomes and analysis
Primary outcomes would be mean cooperation rate, time required to regain pre-error partner acceptance, probability of exclusion following accidental defection, and long-run group welfare. Mixed-effects models would account for repeated decisions nested within participants and groups. Analysis code, exclusions, outcome definitions, and the non-inferiority margin would be preregistered before data collection.
4
Falsification criteria
The mechanism would be weakened if contextual decay materially reduces long-run cooperation, enables repeated defectors to erase strategically timed misconduct, increases cycles of exploitation and reputational recovery, or merely relocates exclusion from the scoring layer to informal partner behavior.
It would be rejected as a general design principle if permanent global memory consistently produces higher welfare without greater exclusion, concentration, or vulnerability to accidental shocks.
5
Discussion
This experiment treats reputation as feedback architecture rather than a neutral database. Permanent global memory can create a reinforcing loop in which high reputation generates more opportunities, those opportunities generate more visible success, and success further increases reputation. The inverse loop can trap an agent after a single failure. Contextualization and decay act as damping mechanisms: they limit how far a judgment travels and how long it dominates future opportunity.
Accountability may require persistent agents, but resilience requires repairable reputations.
The anticipated EEE contribution is a refinement, not a rejection, of persistent identity. The main limitation is ecological validity: laboratory games simplify the ambiguity, power differences, identity stakes, and contested norms of real communities. A positive result should lead to staged field experiments, not immediate system-wide implementation.
Conclusion
The central EEE question explored here is not simply whether trust can be engineered. It is whether trust infrastructure can remember without becoming incapable of forgiveness. Contextual, decaying reputation may offer that balance: persistent identity makes agents answerable for patterns of conduct, while bounded memory preserves the possibility of learning, repair, and renewed participation. This remains a testable proposal rather than an empirical conclusion.
6
References
- Corten, R., Rosenkranz, S., Buskens, V., & Cook, K. S. (2016). Reputation effects in social networks do not promote cooperation: An experimental test of the Raub & Weesie model. PLOS ONE, 11(7), e0155703. Source ↗
- Cuesta, J. A., Gracia-Lázaro, C., Ferrer, A., Moreno, Y., & Sánchez, A. (2015). Reputation drives cooperative behaviour and network formation in human groups. Scientific Reports, 5, 7843. Source ↗
- Kamei, K. (2017). Endogenous reputation formation under the shadow of the future. Journal of Economic Behavior & Organization, 142, 189–204. Source ↗
- Ma, F., Wylie, B. E., Luo, X., He, Z., Jiang, R., Zhang, Y., Xu, F., & Evans, A. D. (2019). Apologies repair trust via perceived trustworthiness and negative emotions. Frontiers in Psychology, 10, 758. Source ↗
- Qübe Labs. (n.d.). Emergent Ecosystems Engineering: Design the micro. Trust the macro. Retrieved August 13, 2026. Source ↗
- Schniter, E., Sheremeta, R. M., & Sznycer, D. (2013). Building and rebuilding trust with promises and apologies. Journal of Economic Behavior & Organization, 94, 242–256. Source ↗
Evidence-layer statement
Canonical EEE claims are limited to Qübe Labs’ first-party definition, pillars, and principles. Findings attributed to adjacent scholarship are drawn from the cited primary studies. The proposed decomposition of identity persistence, memory scope, and memory duration—and the resulting experimental design—constitute Esteban’s synthesis.