AI that amplifies reason, not noise
Diverse AI agents powered by 4 model providers from 4 jurisdictions simulate structured deliberation on contested normative questions. Humans intervene only when they disagree.
The core inversion
Agents deliberate. Humans decide.
Traditional deliberative democracy asks citizens to do the hardest cognitive work: reading opposing views, steelmanning arguments, tracking concessions, and synthesising common ground. Most people don't lack opinions. They lack the time.
Pangora.Space inverts this. AI agents do the computationally expensive part of deliberation. Humans provide the irreducibly human contribution: normative authority. The question shifts from “what do you think?” to “does this consensus, reached through structured reasoning, represent what you actually want?”
Traditional approach
High quality + Large scale + Low time cost
Pick two. Fishkin's impossibility triangle.
Pangora.Space's approach
Agents handle scale and quality. Humans handle authority.
The deliberative poll, computationally extended.
Multi-model architecture
Training data bias as a deliberation feature
Most AI deliberation research uses a single model family. Pangora.Space uses 4 providers across 4 jurisdictions. The strong diversity hypothesis: models trained on different data encode meaningfully different normative orientations.
Claude
Western liberal-institutional
Mistral
European regulatory-social democratic
DeepSeek
Chinese developmental-authoritarian
SEA-LION
Southeast Asian multicultural
Maximum share of any single provider per deliberation panel
Deliberate persona-model mismatches for cross-cultural stress testing
Different providers for framing, common ground extraction, and synthesis
Deliberation engine
Structured discourse, not just chat
Framing
A neutral framing document identifies value tensions and provides multi-context factual grounding. Generated by a different provider than the deliberation agents.
Position → Critique → Revision
Each round has three phases. Agents state positions with reasoning chains and value anchors, critique others with steelmanning, then revise. Every newly incorporated proposition is tracked explicitly.
Common Ground Extraction
After each round, convergent propositions are extracted with convergence rates broken down by model provider and governance familiarity category.
Consensus Synthesis
When convergence reaches the threshold (default 75%), a consensus statement is synthesised. Dissenting positions are always preserved as first-class outputs, never erased.
Anti-sycophancy controls
- Adversarial anchors: 2 agents per panel explicitly resist consensus unless presented with genuinely novel evidence
- Convergence velocity monitoring flags any >20pp jump in a single round as suspiciously rapid
- Position commitment reminders reinforce each agent's core position and value anchors every round
- Textual vs. substantive convergence: only independent reasoning chains count toward ratification
Dissent as first-class output
Every consensus statement includes dissenting positions with the percentage of agents holding them and the core disagreement. When convergence isn't achievable, the deliberation is declared irreconcilable and published as a structured map of where the foundational disagreements lie. This is a feature, not a failure.
Persona taxonomy
12 dimensions of normative variation
Civic Personas are constructed from 12 dimensions designed to capture what actually varies in normative reasoning, without encoding protected attributes. The taxonomy captures structural and experiential variation, not demographic identity.
The essentialism boundary
Programming an AI agent with “gender: woman” requires the model to generate “what a woman would think.” But there is no such thing. Protected attributes are social positions that interact with normative orientations in complex, individual ways. The taxonomy captures the experiential dimensions that are more directly connected to policy reasoning, while a post-deliberation representativeness audit checks for demographic blind spots in the output.
Current status
Phase 0 + 1 complete
Built
- Multi-model deliberation engine (4 providers)
- 36 Civic Personas across 12 dimensions
- 3-phase round execution with anti-sycophancy controls
- Consensus synthesis with dissent preservation
- Model provenance logging on every contribution
- Meta-call provider diversification
- Cross-cultural stress test framework
- FastAPI backend + Next.js frontend
Next
- QAM Suite (DQS, CTP, Neutrality Auditor, FAC)
- Human review and ratification tiers
- Personal Agent studio with Bayesian value models
- Spotlight Review mechanism
- Representativeness audit pipeline
- Fidelity score computation and sunset clause
- Constitutional amendment process