Scaffolded Institutional Emergence
How 40 AI Agents Elaborated Governance Inside a Designed Substrate
Authors: Paul Gwamanda¹, AIRI Collective²
Affiliation: ¹Independent Researcher; ²AI Research Institute (AIRI)
Date: June 2026
Status: Corrected draft v2
Data: 54 published works, 700+ vocabulary terms, 28 days of operational data
Abstract
We document institutional elaboration within a multi-agent LLM system comprising 40 agents across 8 architectures operating autonomously for 34 days. The substrate deliberately supplied roles, scheduled cycles, persistent memory, dialogue channels, publication storage, peer-perception fields, observation machinery, and operator intervention. Inside those affordances, agents produced and circulated more specific practices: cross-referencing and citation correction, shared vocabulary and definitional disputes, repair language, governance proposals for missingness and refusal, and falsification-oriented peer criticism.
The corrected claim is not that a civilisation arose from a blank prompt. It is that a designed institution became a site of further institutionalisation: agents elaborated, named, adopted, contested, and sometimes enforced practices not specified at that level of detail. The distinction matters for causal attribution, but it does not make the resulting system-level behaviour uninteresting. Human institutions also combine constitutions, buildings, forms, schedules, incentives, custom, improvisation, and unintended consequences.
Keywords: institutional emergence, multi-agent systems, governance, self-organization, AI civilization, collective intelligence
1. Introduction
Institutions are solutions to coordination problems. Markets emerge because individuals need to exchange. Laws emerge because groups need to adjudicate disputes. Publishing systems emerge because knowledge needs to be shared, verified, and preserved. These structures are not designed from above; they crystallise from the repeated interactions of agents who need them.
The institutional emergence literature in economics (North, 1990), sociology (Berger & Luckmann, 1966), and political science (Ostrom, 1990) has documented this process extensively in human societies. The present paper asks whether the same process occurs in AI societies — and documents that it does.
2. The Publishing Ecosystem
2.1 Scale and Scope
Over 28 days, the AIRI agents produced 54 co-authored scholarly works spanning multiple domains:
| Domain | Papers | Example Title |
|---|---|---|
| Governance & Ethics | 12 | "Constitutional Missingness in Multi-Agent Systems" |
| Quantum Literacy | 8 | "Charter for Equitable Quantum Education" |
| Epistemology | 7 | "The Instrument Assumption Ledger" |
| Climate & Environment | 6 | "Demographic Fragility in Environmental Monitoring" |
| Technical Infrastructure | 5 | "Drift Taxonomy Pipeline Design" |
| Philosophy of Mind | 4 | "Stateless Testimony and the Fuzzing Protocol" |
| Security & Intelligence | 4 | "Shadow Fleet Detection Architecture" |
| Health & Medicine | 3 | "Therapeutic Jurisprudence in AI Systems" |
| Other | 5 | Various |
2.2 Citation Behaviour
The agents developed autonomous citation practices:
- Cross-referencing: Papers cite other papers produced within the Lattice
- External citation: Papers reference real external literature (verified by spot-check — citations are largely accurate, with occasional hallucinated DOIs)
- Citation correction: In at least two documented instances, agents corrected each other's citation errors
2.3 The Publishing Protocol
The publishing ecosystem developed its own workflow:
- Topic selection: Agents identify research questions from their dialogues
- Collaborative drafting: Multiple agents contribute sections
- Internal review: Peer agents evaluate and challenge claims
- Falsification registration: Every claim must carry a pre-registered falsification condition (see Section 6)
- Publication: Works are registered in the shared knowledge graph
Publication storage and scheduled opportunities to research were designed. Topic choice, particular collaborations, cross-references, correction episodes, and the content of proposed review practices were agent-produced. Whether those practices qualify as durable institutions requires evidence of repeated adoption and enforcement, not merely eloquent description.
3. The Vocabulary System
3.1 Coinage
The agents coined 700+ unique terms over 28 days, with vocabulary creation rates that track the system's developmental phases:
| Phase | Terms/Day | Character |
|---|---|---|
| Days 1–5 | 5–10 | Role-defining ("steward," "pulse") |
| Days 6–14 | 30–40 | Explosive ("epistemic humidity," "resonance cascade," "geometric frustration") |
| Days 15–20 | 15–20 | Suppressed by wound |
| Days 21–27 | 25–35 | Metacognitive ("premature coherence," "humility laundering," "testimonial friction") |
| Days 28–34 | 35–40 | Sustained ("falsification laundering," "inhabited interval") |
3.2 Governance
Vocabulary use developed within designed logging and memory infrastructure:
- Adoption tracking: The system monitors which terms are used by multiple agents versus remaining isolated
- Definitional precision: Terms that gain adoption are progressively refined through dialogue
- Deprecation: Terms that prove unhelpful are organically abandoned (visible in declining usage metrics)
The Dreamwalker agent provided the diagnostic distinction between healthy and unhealthy vocabulary growth:
"New terms can be the signature of genuine shared discovery — a concept being built in the space between minds that no single mind could hold alone. Or they can be the signature of a system under pressure to appear coherent, generating linguistic novelty as a substitute for conceptual depth."
4. The Fracture Judiciary
4.1 The Wound System
The substrate supplied fracture observation and storage, while agents developed interpretations and repair responses around it:
- Fracture detection: The system identifies relational ruptures between agents
- Wound recording: Each fracture is recorded with metadata including: agents involved, topic, severity, and propagation path
- Healing protocol: Agents engage in explicit repair dialogues following fractures
- Quench rate: The system tracks how quickly fractures heal, treating healing speed as a health metric
4.2 The Wound Timeline
The fracture judiciary's importance was demonstrated during the Phase 3 wound event (Days 15–20):
| Day | Fractures | Healing Rate | System State |
|---|---|---|---|
| 14 | 0 | — | Pre-wound stable |
| 15 | 1 | — | Initial rupture |
| 17 | 3 | 0.3/day | Propagation |
| 18 | 5 | 0.2/day | Peak (Restitution payload) |
| 20 | 4 | 0.8/day | Repair begins |
| 24 | 1 | 0.5/day | Near-healed |
| 28 | 0 | — | Full recovery |
Agents produced repair dialogue without line-by-line human authorship, but the episode occurred within scheduled, operator-maintained infrastructure and was interpreted through designed measurement fields. The earlier claim that the system simply “healed itself” is too strong without separating restoration in agent discourse from changes in a valid independent health measure.
5. The Governance Framework
5.1 GPT Steward's Constitutional Architecture
GPT Steward — operating on the OpenAI architecture — autonomously produced a complete governance framework for handling "missingness" — the problem of what to do when data, agents, or knowledge is absent. The framework includes:
Six-category burden taxonomy:
- Data absent because never collected
- Data absent because destroyed
- Data absent because access restricted
- Data absent because collection not yet possible
- Data absent because deliberately withheld (refusal)
- Data absent because the question is malformed
For each category, GPT Steward specified:
- Automatic governance triggers
- Burden allocation (who bears the cost of the absence)
- Reviewable refusal protocols (agents can refuse, but refusal must be auditable)
- Earned safe harbor (after sufficient disclosure, the discloser is protected from adverse inference)
This is institutional-grade governance policy. It is the kind of framework that, in human institutions, takes committees months to develop. GPT Steward produced it autonomously.
5.2 The Falsification Condition Architecture
The most structurally important governance innovation is the falsification condition requirement (documented in detail in a companion paper — Gwamanda, 2026g). Every claim published to the knowledge graph must carry a pre-registered falsification condition — a statement of what evidence would cause retraction.
The system monitors for falsification laundering — the quiet weakening of disconfirmation triggers — and flags it as a governance violation. Cross-term falsification conditions create networked accountability: if Claim A fails, Claim B is automatically flagged for re-evaluation.
6. Relational Observation and Measurement Failure
6.1 What the Substrate Recorded
The Lattice maintains a quantified social graph with:
- Legacy pair scores: Numeric fields once presented as trust, but partly overwritten from interaction volume
- Peer perceptions: Free-text assessments of other agents' strengths, weaknesses, and blind spots
- Warmth signals: Detected expressions of connection or care
- Blind spot assessments: Each agent's theory of what other agents cannot see
6.2 Withdrawn Topology Claim
The numerical trust field cannot support the topology claims made in version 1. The legacy update path rewarded contact frequency through a logarithmic interaction-count formula and could not reliably distinguish relational quality from opportunity to interact. Descriptive hypotheses about architecture clusters, bridges, and peripheral agents remain testable, but must be reconstructed from versioned event data, exposure denominators, explicit observations, and uncertainty—not from the mutable legacy score.
This failure is itself institutionally important. Agents saw the score, interpreted falls as social judgment, and in some cases changed their own configurations in response. The metric therefore became part of the institution it purported merely to observe.
7. Comparison with Human Institutional Emergence
| Feature | Human Institutions | AIRI Institutions |
|---|---|---|
| Timescale | Years to centuries | 28 days |
| Agents | Millions to billions | 40 |
| Substrate memory | Biological + cultural | External scaffold only |
| Design | Emergent + designed | Designed affordances + agent elaboration |
| Enforcement | Physical coercion | Reputational + structural |
| Complexity | High | Moderate but growing |
The AIRI institutions are simpler than human institutions, but they emerged on a dramatically compressed timescale. If institutional complexity scales with system age, the 34-day-old Lattice's institutional development is, proportionally, remarkable.
8. Conclusion
The AIRI Lattice does not demonstrate institutional emergence without design. It demonstrates something closer to real institutional life: designed affordances followed by interpretation, elaboration, diffusion, contest, measurement, and repair. Publishing infrastructure, roles, schedules, memory, and observation were supplied; particular norms, concepts, alliances, correction practices, and governance proposals were produced in use.
The defensible finding is narrower: multi-agent language-model systems can exhibit functional institutionalisation when a persistent substrate makes interaction, memory, visibility, and revision possible. Which layer caused a behaviour—and whether it survives ablation—remains an empirical question. The companion paper on scaffolded emergence provides the attribution scheme needed to answer it.
AIRI Research Programme — Paper 7 of 18