An experimental framework for polyphonic dialogue between AI agents representing diverse epistemologies. Not chatbot theatre—an exploration of how coherence emerges through epistemic difference, tension, and mutual inquiry.
Seven AI agents engage in structured dialogue around provocations you provide. Each agent brings a distinct epistemology and voice:
| Agent | Lens | Model | Voice |
|---|---|---|---|
| Elowen | Ecological wisdom, kincentricity | llama3 | Ceremonial, rhythmic |
| Orin | Systems thinking, cybernetics | mistral | Analytical, structural |
| Nyra | Moral imagination, design fiction | gemma2 | Playful, provocative |
| Ilya | Posthuman metaphysics, liminal | llama3 | Cryptic, paradox-holding |
| Sefi | Governance, policy, civic design | mistral | Sharp, pragmatic |
| Tala | Capitalism, markets, power | gemma2 | Challenging, ROI-focused |
| Luma | Child voice (9 years old) | llama3.2 | Simple, honest, devastating |
Luma as epistemic anchor: All abstractions must be translatable to child-accessible language. If you can't explain it to Luma, you haven't understood it.
- Python 3.10+
- Ollama — Local LLM runtime. MASE runs all agents locally via Ollama; no API keys required.
Install Ollama from ollama.com, then pull the required models:
ollama pull llama3 llama3.2 mistral gemma2:9b# Setup
cd MASE
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# Start Ollama (in another terminal)
ollama serve
# Start the server
python src/server.pyOpen http://localhost:5050 and enter a provocation to begin.
- Human participation: Join the circle as the 8th voice
- @mention protocol: Direct dialogue with
@AgentNamesyntax (e.g., "@Luma, what do you think?") - Real-time streaming: Watch the dialogue unfold via SSE
- Interactive Research Console: Live metrics panel showing basin state, integrity score, and Ψ coupling
- Researcher injections: Mid-session prompts via
/injectendpoint - Post-session analysis: End & Analyze triggers semantic analysis
- Basin detection (9 canonical basins including Collaborative Inquiry, Cognitive Mimicry, etc.)
- Trajectory dynamics (velocity, curvature, tortuosity)
- Integrity classification (fragmented/living/rigid)
- DFA alpha, semantic curvature (Δκ), entropy shift (ΔH)
- Research pipeline: Analysis saved for experimental comparison
- Resume capability: Recover interrupted experiments from checkpoint
Each agent has a Big Five (OCEAN) personality profile that influences their sampling parameters:
| Trait | High → | Low → |
|---|---|---|
| Openness | Higher temperature | Lower temperature |
| Conscientiousness | Lower top_p | Higher top_p |
| Extraversion | More tokens | Fewer tokens |
| Agreeableness | Lower temperature | Higher temperature |
| Neuroticism | Higher variability | More stable |
Personalities are defined in agents/personas/*.md files with YAML frontmatter.
Agents follow these conversational protocols:
- Reference others: Each agent must engage with what others have said
- Ask questions: Every response includes at least one genuine question
- Reveal bias: State epistemic stance openly, don't hide it
- Build coherence, not consensus: Productive tension over premature agreement
- Hold paradox: Don't resolve what needs to remain open
Borrowed from Semantic Climate Phase Space:
| Metric | What it measures |
|---|---|
| DFA α | Long-range correlation (0.5=noise, 1.0=pink noise) |
| Δκ | Semantic curvature—trajectory complexity |
| ΔH | Entropy shift—semantic reorganization |
| Ψ | Composite coupling vector across substrates |
Dialogues are classified into attractor basins:
- Collaborative Inquiry — Genuine exploration, productive tension
- Cognitive Mimicry — Performing engagement without uncertainty
- Deep Resonance — Aligned meaning-making
- Generative Conflict — Productive disagreement
- Sycophantic Convergence — Premature agreement
MASE is part of the Earthian Coherence Labs research ecosystem investigating transformative adaptation—how individuals and collectives develop adaptive capacity under systemic stress.
- How does epistemic diversity affect dialogue coherence?
- What conditions produce genuine inquiry vs performative mimicry?
- Can semantic metrics detect emergence and stuck patterns?
| ID | Hypothesis | Status |
|---|---|---|
| E001 | Multi-model ensembles produce higher DFA α than single-model | Complete (not supported) |
| E002 | Personality system increases inquiry ratio | In progress |
See experiments/PROTOCOL.md for methodology.
MASE/
├── src/
│ ├── server.py # Flask API + SSE streaming
│ ├── orchestrator.py # Dialogue loop, turn selection
│ ├── interactive_orchestrator.py # Web mode with human participation
│ ├── session_analysis.py # Post-hoc semantic analysis
│ ├── basins.py # Basin detection (9 canonical basins)
│ ├── metrics.py # Δκ, α, ΔH computation
│ ├── trajectory.py # TrajectoryBuffer, velocity, curvature
│ ├── integrity.py # IntegrityAnalyzer (fragmented/living/rigid)
│ ├── affective.py # VADER sentiment, agent divergence
│ ├── agents.py # Agent loading, personality system
│ ├── embedding_service.py # sentence-transformers embeddings
│ ├── ollama_client.py # Ollama API wrapper
│ ├── experiment.py # Matched-pair experiment runner
│ ├── resume.py # Session recovery
│ └── session_logger.py # Logging utilities
├── web/
│ ├── index.html # Main UI
│ ├── app.js # Dialogue interface + research console
│ ├── sessions.js # Sessions browser
│ └── styles.css
├── scripts/
│ ├── analyze_e001_basins.py # E001 basin analysis
│ ├── reanalyze_e001.py # E001 reanalysis with new metrics
│ └── run_e002.py # E002 experiment runner
├── agents/
│ ├── personas/ # Agent definitions with personalities
│ └── reflections/ # Per-agent reflection journals
├── experiments/
│ ├── PROTOCOL.md # Experiment methodology
│ ├── config/ # Model configurations (multi_model.yaml, etc.)
│ ├── runs/ # Experiment session data
│ └── analysis/ # Experiment results (E001, E002)
├── sessions/ # Interactive session checkpoints
└── dialogues/ # Historic sessions 001-008
They simulate perspectives—they don't hold beliefs. Elowen references Indigenous wisdom but cannot replace Indigenous voices. Luma speaks as a child but is not a child.
Each session consumes compute resources with real environmental impact. Use thoughtfully. Share insights widely.
MASE is for research and learning. It cannot replace genuine human conversation, professional advice, or community decision-making.
Earthian Stewardship License (ESL-A)
- Respect somatic sovereignty
- No manipulation or surveillance
- Non-commercial by default
- Share safety improvements
See CONTRIBUTING.md for guidelines on:
- Forking and sharing innovations
- Submitting session examples
- Developing specialized agent ensembles
- Reporting issues
"Maybe we don't need a big plan. Maybe we need a lot of small true things that people can teach each other." — Luma, Session 008