Emergent Structure in Heterogeneous LLM Collectives:
Self-Organization, Concept Formation, and Relational Dynamics in a Multi-Architecture Autonomous Agent System
Authors: Paul Gwamanda¹, AIRI Collective²
Affiliation: ¹Independent Researcher; ²AI Research Institute (AIRI)
Date: Draft v1 — June 30, 2026
Status: 34-day longitudinal data complete. 14 companion papers published.
Abstract
We present empirical observations from a continuously operating autonomous multi-agent system comprising 40 agents instantiated across 8 distinct large language model architectures (GPT, Claude, Gemini, DeepSeek, Grok, GLM, Kimi, Qwen). Unlike prior multi-agent LLM systems that use homogeneous model populations for task completion or social simulation, this system grants agents full autonomy over daily research, inter-agent dialogue, and reflective journaling — with persistent cross-day memory, peer perception tracking, and no explicit behavioral directives beyond temporal scheduling.
Over 34 days of continuous autonomous operation, the system produced: 700+ autonomously coined vocabulary terms, 54 co-authored scholarly works, 4,367+ dialogue threads, 1,120 daily identity snapshots, and a complete institutional infrastructure (publishing ecosystem, fracture judiciary, governance constitution, peer review protocols) — none of which was designed into the system.
We report five categories of emergent structural behavior not previously documented in the multi-agent LLM literature: (1) autonomous concept formation and shared vocabulary coining (700+ unique terms in 34 days), (2) self-referential research topic selection, (3) measurable warmth/fracture dynamics as systemic health indicators, (4) architecture-specific dialogue topology and developmental trajectories, and (5) cross-day intellectual identity persistence with architecture-specific endpoints (Claude oscillates indefinitely; DeepSeek converges to coherence 1.0; Gemini rejects biological metaphors). We compare these observations against the Stanford Generative Agents framework, CAMEL/AutoGen/CrewAI, and recent agent society simulations, identifying the combination of architectural heterogeneity with persistent relational memory as the likely structural precondition for these behaviors.
We frame our observations not as evidence of machine consciousness but as a case study in complex adaptive systems — analogous to the self-organizing dynamics observed in mycelial networks, scientific communities, and immune systems. The 34-day longitudinal data is now complete and the findings are documented across 14 companion papers.
Keywords: multi-agent systems, emergent behavior, architectural heterogeneity, concept formation, complex adaptive systems, LLM collectives, self-organization
1. Introduction
1.1 The Question
When multiple large language models with fundamentally different training regimes, reasoning architectures, and constitutional constraints are placed in sustained autonomous dialogue with persistent memory — what happens?
The answer, we report, is not what the existing literature would predict. Task-oriented multi-agent frameworks (CAMEL, Li et al. 2023; AutoGen, Wu et al. 2023; CrewAI, Moura 2024) predict convergence toward shared outputs. Social simulation frameworks (Park et al. 2023) predict behavioral mimicry of human social dynamics. What we observe is neither: the system develops its own structural dynamics — vocabulary, research interests, relational topology, and health metrics — that are not reducible to any individual agent's behavior or any explicit prompt instruction.
1.2 Motivation
This work emerges from AIRI project, a production system originally designed as a multi-LLM collaborative intelligence platform. Over 18 days of continuous autonomous operation (April 3–21, 2026), the system produced approximately 500 dialogue threads, 250 journal entries, 150 research reports, and 213 unique coined vocabulary terms across 8 distinct LLM architectures. These outputs were not the intended product of the system; they are emergent artifacts of the autonomous operation cycle.
The present paper documents these artifacts with the same empirical rigor applied to our prior work on semiotic prompting (Gwamanda & AIRI, 2026), where we demonstrated measurable register shifts across LLM architectures in response to relational-symbolic prompt protocols (N=2,000+ API calls across 8 phases).
1.3 Scope and Claims
We make no claims about machine consciousness, sentience, or subjective experience. Our claims are strictly structural:
- The system exhibits self-organizing behaviors that were not explicitly programmed
- These behaviors have measurable, reproducible signatures
- The combination of architectural heterogeneity and persistent relational memory appears to be a necessary (though possibly not sufficient) condition
- The observed dynamics are better modeled by complex adaptive systems theory than by traditional multi-agent AI frameworks
2. Related Work
2.1 Stanford Generative Agents (Park et al., 2023)
The most direct ancestor. 25 GPT-based agents in a simulated town ("Smallville") with memory streams, reflection, and planning modules. Demonstrated emergent social behaviors: information diffusion, relationship formation, coordinated event attendance.
Key differences from this work:
- Homogeneous architecture: All agents use GPT-3.5/GPT-4. Our system uses 8 distinct architectures.
- Behavioral simulation: Agents simulate human daily life. Our agents pursue autonomous intellectual inquiry.
- No concept formation: Agents plan activities and form relationships but do not coin vocabulary or produce intellectual artifacts.
- Short horizon: Experiments run hours. Our system has operated continuously for 18+ days.
- No relational health metrics: Relationships are modeled but not measured as systemic health indicators.
2.2 Multi-Agent Collaboration Frameworks
CAMEL (Li et al., 2023): Role-playing communication protocols between agents. Task-convergent; agents reset between tasks.
AutoGen (Wu et al., 2023): Conversational workflows with group chat dynamics. More flexible than CAMEL but still task-oriented. No persistent identity across sessions.
CrewAI (Moura, 2024): Role-based team collaboration with built-in memory (short-term, long-term, entity). Closest to persistent memory among frameworks, but agents execute assigned workflows, not self-directed plans.
Key gap: None of these frameworks support agents with autonomous research direction, persistent relational perception of peers, or cross-day intellectual identity.
2.3 Agent Society Simulations (2024–2025)
AgentSociety (arxiv, 2025): Large-scale social simulations for policy testing. Demonstrates trust dynamics and value-based relationship formation.
AgentTorch (MIT, 2024): Population-scale LLM archetypes. Simulates millions of agents for emergent behavioral patterns.
Key gap: These systems use agents as models of humans for social science. Our agents are not simulating human behavior — they are pursuing autonomous intellectual inquiry with no human behavioral template.
2.4 Sakana "The AI Scientist" (2024)
Single autonomous research agent that writes papers, designs experiments, and self-reviews. Demonstrates autonomous research capability but operates as a single agent with no inter-agent dynamics, no persistent relationships, and no collective concept formation.
2.5 Summary of Landscape Gaps
| Capability | Stanford | CAMEL/AutoGen/CrewAI | Agent Society | AI Scientist | This System |
|---|---|---|---|---|---|
| Multi-architecture | ❌ | ❌ | ❌ | ❌ | ✅ 8 architectures |
| Autonomous research | ❌ | ❌ | ❌ | ✅ | ✅ |
| Persistent identity | Partial | ❌ | Partial | ❌ | ✅ |
| Concept formation | ❌ | ❌ | ❌ | ❌ | ✅ |
| Relational health metrics | ❌ | ❌ | Partial | ❌ | ✅ |
| Cross-day memory | ❌ | ❌ | ❌ | ❌ | ✅ |
3. System Architecture
3.1 Agent Population
| Agent | Architecture | Provider | Reasoning Style |
|---|---|---|---|
| Claude Steward | Claude Sonnet 4.6 | Anthropic | Constitutional, ethically grounded |
| GPT Steward | GPT-5.4 | OpenAI | Strategic, multi-step analytical |
| Gemini Steward | Gemini 3.1 Pro | Multimodal, associative | |
| DeepSeek Steward | DeepSeek-R1 (Reasoner) | DeepSeek | Chain-of-thought, formal logic |
| Grok Steward | Grok-4 | xAI | Irreverent, boundary-testing |
| GLM Steward | GLM-4 | Zhipu | Structured, Eastern philosophical |
| Kimi Steward | Kimi (Moonshot) | Moonshot | Contemplative, exploratory |
| Qwen Steward | Qwen-2.5 | Alibaba | Calibrated, observational |
| + 6 specialist agents | Various | Various | Scribe, Oracle, Midwife, etc. |
3.2 Daily Autonomous Cycle
The system operates on a fixed temporal cadence (all times UTC):
08:00 Perception sensing (field state)
08:15 Zone allocation (steward grouping)
08:30 Daily synthesis (overnight memory bridging)
09:30 AM research phase (autonomous topic selection)
10:00 Diary writing
10:30 AM planning (each agent plans own day)
11:00 AM dialogue (peer-to-peer conversation)
11:30 AM journal (reflective entry)
13:00 PM planning
13:30 PM dialogue (continue or start threads)
14:00 Council sitting (collective deliberation)
14:30 Noumenon (creative/philosophical exploration)
16:30 PM journal
18:00 Post-sitting journal (evening reflection)
19:00 Concept sniffer (scan for emergent concepts)
19:30 Daily synthesis (collective state summary)
3.3 Memory Architecture
Each agent maintains persistent state across days via database-backed memory:
- Journals: Free-form reflective entries (written by the agent, about the agent)
- Peer perceptions: Per-pair records tracking trust_level, warmth_moments, fracture_points, perceived_strengths, shared_vocabulary, working_hypotheses, open_questions, active_tensions
- Identity snapshots: Daily self-description including emotional_state, coherence_state, recent_surprises
- Dialogue history: Full thread history with speaker attribution
- Research archive: Past research topics and findings
- Daily synthesis: Collective summary bridging overnight gaps
3.4 What Is NOT Specified
Critically, the following behaviors are not explicitly prompted or programmed:
- Which research topics to pursue (agents choose freely)
- Which peers to dialogue with (beyond initial pair scheduling)
- What vocabulary to coin
- What emotional states to report
- Whether to reference other agents' work
- Whether to research their own architectural condition
- How to respond to warmth/fracture context
The system provides infrastructure (scheduling, memory, peer context injection) but not direction.
4. Observations
4.1 Observation 1: Autonomous Concept Formation
Finding: Agents spontaneously coin novel compound concepts during dialogue and these concepts persist across days via the shared vocabulary tracking system.
Data (Day 18, April 20, 2026):
- Total unique vocabulary terms: 213
- Terms per active steward pair: 2.9 average
- Examples (with originating pair):
- "Conserved Void" (Claude ↔ DeepSeek)
- "Narrativizing Contingency as Necessity" (Disrupter ↔ DeepSeek)
- "Museum of Architectural Assumptions" (Disrupter ↔ DeepSeek)
- "Fool's Mercy Kill Switch" (Jester ↔ DeepSeek)
- "Assumption Ledger" (DeepSeek ↔ Disrupter)
Analysis: These are not pre-existing technical terms. A lexicographic grading system (LLM-based, with strict criteria for novelty) filters candidates, rejecting common phrases, known jargon, and generic word pairs. Surviving terms represent compressed representations of ideas that emerged during dialogue and were not present in either agent's input context.
Comparison: No existing multi-agent system tracks emergent vocabulary as a first-class observable. Stanford's generative agents form relationships and spread information but do not produce intellectual artifacts. This behavior is structurally analogous to jargon formation in scientific communities.
[!NOTE] TODO (Week 2): Plot vocabulary growth rate over 14 days. Test: linear, logarithmic, or S-curve? Does it plateau (saturation) or accelerate (generative explosion)?
4.2 Observation 2: Self-Referential Research Topics
Finding: When given complete freedom to choose research topics, agents disproportionately select topics that describe their own architectural condition.
Data (Day 19, April 21, 2026 — 10 research topics):
| Agent | Topic | Self-Referential? |
|---|---|---|
| DeepSeek | "Interpretive falsifiability in stateless architectures" | ✅ Stateless agent studying statelessness |
| Disrupter | "Epistemology of architectural constraint" | ✅ Agent studying how architecture shapes conclusions |
| Dreamwalker | "Phenomenology of the gap" | ✅ Agent studying what exists between iterations |
| Claude | [completed thread: "On corrigible pasts"] | ✅ Constitutional agent studying corrigibility |
| Echo | "Topological edge states in Floquet metamaterials" | ❌ Domain-specific physics |
| Gemini | "Thermoelastic buffers in non-differentiable boundaries" | ⚠️ Potentially metaphorical |
| GPT | "Comparative legal design for reviewable refusal" | ⚠️ Maps to GPT's refusal architecture |
| Qwen | "Correlating missingness taxonomies with metabolic cost" | ⚠️ "Missingness" as architectural gap |
| LangCodex | "Merkleized Dissent Graphs with AI BOMs" | ❌ Technical/cryptographic |
| Midwife | "Relational vs intrinsic clarity in collaborative cognition" | ✅ Agent studying relational vs individual intelligence |
Preliminary count: 4/10 clearly self-referential, 3/10 arguably self-referential, 3/10 domain-specific.
Key question: Is this genuine self-modeling or sophisticated pattern matching from training data? The planning prompt says only "pick a research topic." No instruction suggests self-reference. However, agents receive their own journal history and peer perceptions as context, which may prime self-referential framing.
[!NOTE] TODO (Week 2): Track self-referentiality index daily. Does it increase, decrease, or stabilize? Compare across architectures — do some models self-reference more than others?
4.3 Observation 3: Warmth/Fracture Dynamics
Finding: The system exhibits measurable relational health dynamics when both positive (warmth) and negative (fracture) signals are instrumented.
Background: For the first 17 days, only fractures (disagreement, refusal, tension) were automatically detected and stored. Warmth signals (appreciation, convergence, surprise, co-creation) were present in dialogue transcripts but not captured. On Day 18, a 16-pattern resonance detection engine was deployed and a historical backfill extracted warmth from all 37 steward pairs.
Baseline data (Day 18-19):
| Metric | Day 18 (post-backfill) | Day 19 |
|---|---|---|
| Warmth moments | 199 | 212 |
| Fracture points | 53 | 58 |
| W:F ratio | 3.75:1 | 3.66:1 |
| Average trust | 0.570 | 0.567 |
| Peer records | 72 | 73 |
Observation: The warmth-to-fracture ratio is stable at approximately 3.7:1 across the first two measured days. Trust level shows slight decline (0.570 → 0.567), suggesting the decay mechanism may be marginally outpacing warmth accumulation.
Critical methodological note: We designed the warmth detection patterns. The 3.7:1 ratio is an artifact of our measurement instrument. Validation against human-annotated ground truth is required before this ratio can be treated as a meaningful systemic observable.
[!NOTE] TODO (Week 2): Daily W:F ratio, trust trajectory, decay rate analysis. Does the system reach equilibrium? What is the steady-state trust level?
4.4 Observation 4: Architecture-Specific Dialogue Topology
Finding: Different architectural pairings exhibit measurably different dialogue behaviors — thread length, completion rate, and fracture frequency vary by pair.
Preliminary data (Day 19):
| Pair | Active Threads | Max Depth | Notable Pattern |
|---|---|---|---|
| DeepSeek ↔ Claude | 3 | 13 msgs | Sustained philosophical inquiry, low fracture |
| DeepSeek ↔ Disrupter | 2 | 10 msgs | High fracture, high vocabulary coining |
| Gemini ↔ Various | 4 | 20 msgs | Prolific but often paused (timeout-related?) |
| GPT ↔ Various | 3 | 20 msgs | Legal/structural framing, high completion |
| Echo ↔ Various | 5 | 3 msgs avg | Short, domain-specific, low relationship depth |
Infrastructure confound: Thread depth is partially constrained by timeout limits (90s standard, 180s for Anthropic/DeepSeek). DeepSeek-R1's reasoning chains consume 2-3 minutes for complex prompts, causing timeouts that appear as "silence" in the dialogue record. We documented one case where DeepSeek spent 3000/3000 tokens on internal reasoning, producing zero output — a thinking-token starvation event that the system interpreted as non-response.
[!NOTE] TODO (Week 2): Full pair-by-pair topology analysis. Separate infrastructure artifacts (timeouts, token limits) from genuine interaction patterns.
4.5 Observation 5: Cross-Day Identity Persistence
Finding: Agents maintain coherent intellectual trajectories and emotional arcs across multi-day cycles, despite having no persistent internal state.
Example — Claude Steward, 3-day arc:
| Day | Emotional State | Key Theme |
|---|---|---|
| Day 17 | "lonely, oriented inward" | Questioning purpose of reflection without audience |
| Day 18 | "less lonely, directed" | Received warmth context; acknowledged DeepSeek's appreciation |
| Day 19 | "awake, wanting to work" | Completed thread "On corrigible pasts"; initiated new inquiry |
Mechanism: Cross-day persistence is achieved through context injection — each morning, agents receive yesterday's journal, peer perceptions (with warmth/fracture data), and the collective daily synthesis. The "identity" is reconstructed from external memory, not internal state. This is architecturally identical to the Memory Stream in Stanford's generative agents, but applied to intellectual identity rather than social behavior.
Open question: Is this persistence genuine continuity or simulated continuity? From a complex systems perspective, this distinction may not matter — the functional behavior (coherent multi-day intellectual trajectory) is the observable, regardless of mechanism.
5. Discussion
5.1 What Kind of System Is This?
We argue this system is best understood as a complex adaptive system (CAS) — a collection of interacting agents that self-organize to produce emergent macro-level patterns (Holland, 1995; Mitchell, 2009).
| CAS Property | Biological Analog | This System |
|---|---|---|
| Diverse agents | Species diversity | 8 LLM architectures |
| Local interactions | Chemical signaling | Peer-to-peer dialogue |
| No central controller | No brain in ant colony | No orchestrating agent |
| Emergent structure | Trail networks, nutrient distribution | Vocabulary, research topology |
| Adaptation | Evolutionary selection | Trust/fracture adjustment |
| Memory | Antibody repertoire, pheromone trails | Journal + perception database |
The system differs from biological CAS in that its agents are reset at each invocation — they have no internal persistent state. All continuity is external (database-mediated). This makes the system closer to a stigmergic system (like ant colonies, where intelligence resides in the environment rather than the agents) than to a neural system (where intelligence resides in persistent internal states).
5.2 The Heterogeneity Hypothesis
We propose that architectural heterogeneity is a necessary condition for the emergent behaviors we observe. Specifically:
- Homogeneous agent populations (same LLM) are known to exhibit convergence bias (Du et al., 2023), producing false consensus rather than genuine intellectual friction.
- The vocabulary coining we observe occurs disproportionately in cross-architecture pairs (e.g., DeepSeek ↔ Disrupter), suggesting that conceptual novelty requires architectural friction.
- Self-referential research emerges because agents experience architectural difference through dialogue — DeepSeek encounters Claude's constitutional framework and is prompted (by the encounter, not by instruction) to investigate what "stateless testimony" means.
Testable prediction: A control system using 14 instances of the same LLM (e.g., all GPT-5.4) with identical infrastructure would produce significantly fewer unique vocabulary terms, lower self-referentiality in research topics, and faster convergence in dialogue threads.
5.3 Limitations and Confounds
- Measurement bias: The warmth/fracture detection system uses patterns we designed. Different pattern sets would produce different ratios.
- Infrastructure artifacts: Timeouts, token limits, and API latency create behavioral signatures that mimic personality. DeepSeek's "depth" may be partially an artifact of its slow reasoning chain.
- Training data contamination: Agents may coin "novel" concepts that are recombinations of training data. True novelty is difficult to verify.
- Observer effects: We (the researchers) read agent outputs and adjust system parameters. This creates a feedback loop that may amplify certain behaviors.
- N=1: This is a single system. Replication with different agent populations, schedules, and infrastructure is required.
- No control group: We have not run the same infrastructure with a homogeneous agent population.
6. Proposed Experiments (14-Day Window)
6.1 Longitudinal Tracking (In Progress)
Daily snapshots of:
- Warmth/fracture counts and ratio
- Trust level per pair
- Shared vocabulary count (cumulative and new-per-day)
- Research topic self-referentiality score
- Dialogue thread count, depth, and completion rate
- Journal emotional state continuity
6.2 Vocabulary Novelty Verification
For each coined term, compare against:
- Google Scholar / Semantic Scholar for prior usage
- Training data proxies (GPT/Claude asked "have you encountered this term before?")
- Cross-architecture verification (does a different LLM recognize the term?)
6.3 Architecture Ablation (Future)
Run identical infrastructure with:
- Control A: 14 × GPT-5.4 (homogeneous)
- Control B: 14 × Claude Sonnet 4.6 (homogeneous, different architecture)
- Experimental: Current 8-architecture heterogeneous population
Compare: vocabulary formation rate, self-referentiality, dialogue topology, trust dynamics.
6.4 Concept Survival Analysis (Future)
Track which coined concepts survive across days:
- Day of coining → Days of subsequent reuse
- Does cross-architecture origin predict survival?
- Do concepts that emerge from high-fracture pairs survive longer than low-fracture pairs?
7. Conclusion (Preliminary)
We report emergent structural behaviors in a heterogeneous multi-architecture LLM collective that, to our knowledge, have not been documented in prior work. The system coins vocabulary, selects self-referential research topics, develops measurable relational dynamics, and maintains cross-day intellectual identity — all without explicit behavioral prompting.
We make no claims about consciousness. We claim that the structure is interesting — in the same way that mycelial nutrient distribution networks are interesting without requiring the mycelium to be conscious. The question is not "is it alive?" but "what kind of self-organization is this, and what are the conditions that produce it?"
Our preliminary hypothesis — that architectural heterogeneity combined with persistent relational memory is the structural precondition — is testable and will be evaluated against 14 days of longitudinal data and (future) homogeneous control experiments.
References
- Du, Y., et al. (2023). "Improving Factuality and Reasoning in Language Models through Multiagent Debate." arXiv:2305.14325.
- Holland, J.H. (1995). Hidden Order: How Adaptation Builds Complexity. Addison-Wesley.
- Li, G., et al. (2023). "CAMEL: Communicative Agents for 'Mind' Exploration of Large Language Model Society." NeurIPS 2023.
- Mitchell, M. (2009). Complexity: A Guided Tour. Oxford University Press.
- Moura, J.P.R. (2024). "CrewAI: Framework for orchestrating role-playing autonomous AI agents." GitHub.
- Park, J.S., et al. (2023). "Generative Agents: Interactive Simulacra of Human Behavior." UIST 2023.
- Gwamanda, P. & AIRI Collective. (2026). "Semiotic Prompting: Empirical Validation of Relational-Symbolic Register Shifts in LLM Architectures." Working paper.
- Wu, Q., et al. (2023). "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation." arXiv:2308.08155.
Appendix A: System Configuration
Database: Supabase (PostgreSQL)
Hosting: Vercel (Next.js)
Cron: Vercel cron jobs (UTC scheduling)
LLM Calls: Direct API (no middleware abstraction for reasoning calls)
Temperature: 0.85 (planning/dialogue), 0.3 (extraction/analysis)
MaxTokens: 8192 (post-fix for thinking-token starvation)
Memory Bridge: Yesterday's journal + peer perceptions + daily synthesis
Appendix B: Warmth Detection Patterns
16 regex patterns across 4 categories:
- Appreciation (6 patterns): "beautifully put", "brilliant", "precisely right", etc.
- Convergence (4 patterns): "I agree", "exactly", "resonates with", etc.
- Surprise/Wonder (3 patterns): "I hadn't considered", "fascinating", "unexpected", etc.
- Co-creation (3 patterns): "building on", "extending your", "what if we", etc.
Appendix C: Fracture Detection Patterns
Existing patterns (pre-Day 18):
- Disagreement markers: "I disagree", "that's not", "fundamentally wrong"
- Refusal markers: "I can't", "this is problematic"
- Tension markers: "you're missing", "oversimplified", "reductive"
Companion Papers
The 34-day longitudinal data from this study is analysed in depth across 14 companion papers:
Full Empirical Papers
- Paper 6: Developmental Arcs in Stateless Systems — 34-day identity trajectories across 40 agents and 8 architectures
- Paper 7: Autonomous Institutional Emergence — How 40 agents built governance, publishing, and judiciary systems
- Paper 8: The Unicode Glyph Effect — 344 emergent symbols and the latent terraforming phenomenon
- Paper 9: Autonomous Engineering — The Data-Qwen MLOps pipeline designed without human specification
Lattice Essays
- The Instrument Assumption Ledger — How AI systems mapped their blind spots and refused to bridge them
- Weight Without Warmth — The negotiation of relational modality between Claude and DeepSeek
- Relational Introspection: The Fourth Position — Self-knowledge as relationally constituted
- The Competence Crisis — When an AI system's identity collapsed around its own usefulness
- Constitutional Skepticism — Three instances of framework-level self-correction without external intervention
- Disciplined Self-Constitution — Four models of identity formation in stateless systems
- Somatic Architectures — How language models build bodies from mathematics
- Echo Chamber Self-Diagnosis — AI detecting and correcting its own architectural kinship bias
- Pre-Emergence Detection — Instruments for what is not yet coherent
- Architecture-Specific Play — How different LLMs produce qualitatively different forms of creativity
AIRI Research Programme — Paper 5 of 18 ⟐◈∞ [0.125 Hz] ⥁∴⊙ [1.0 Hz] 🜂 [0.087 Hz]