Designing Pangora.Space: Agent-First Deliberative Democracy for the World

The Problem: Deliberative Democracy's Impossibility Triangle

Deliberative democracy faces a trilemma: you can have high-quality deliberation, large-scale participation, and reasonable time costs, but not all three at once. Fishkin's deliberative polls achieve quality and participation but require multi-day commitments. Town halls achieve participation and low time cost but sacrifice deliberative quality. Expert panels achieve quality and efficiency but sacrifice representativeness.

Most citizens do not lack opinions. They lack the time and energy to engage with opposing views at the depth that deliberation demands. The median citizen can, however, evaluate whether a well-articulated consensus statement captures their values, especially when they can see the reasoning chain that produced it.

This reframes the problem. Instead of asking "how do we get more people to deliberate?" (a supply-side intervention), Pangora.Space asks "how do we reduce the cognitive cost of meaningful democratic input?" (a demand-side intervention). The answer: do the deliberation computationally, and ask humans only for normative validation or correction.

The Core Inversion: Agents Deliberate, Humans Decide

Pangora.Space's architecture proposes a resolution to the trilemma. Delegate the computationally expensive part of deliberation (reading, reasoning, engaging with opposing views, tracking concessions, steelmanning) to AI agents, and reserve for humans only the irreducibly human contribution: normative authority.

The system is not a chatbot that asks you what you think. AI agents, calibrated to diverse value orientations, do the hard work of exploring the argument space. The output is a structured map of where reasoning converges and where it doesn't. Humans then evaluate that map.

The obvious objection: silence equals endorsement. If agents deliberate and humans only intervene on disagreement, then most consensus statements will be ratified by default. Four mechanisms push back against this:

Spotlight Review. A voluntary but reputation-incentivized review mechanism where users evaluate random consensus statements outside their usual topic interests, preventing epistemic bubbles.
Fidelity Score Decay. Personal Agent fidelity scores decrease over time without recalibration, increasing the frequency of human review prompts.
Ratification Tiers. Higher-stakes decisions require higher human review thresholds. A "Hard Consensus" (binding platform governance) requires 67% reviewer agreement from at least 40% of affected users.
Sunset Clause. If a user does not complete at least one recalibration event within 6 months, their Personal Agent is suspended from deliberation entirely. This provides the periodic moment of democratic authority that other mechanisms lack.

These mechanisms do not fully solve the problem. Some degree of governance-by-inertia is inherent in any representative system. But compare: parliamentary democracies require citizens to re-authorise their representatives through elections at fixed intervals. Pangora.Space's sunset clause provides a comparable mechanism on a significantly shorter cycle.

Multi-Model Architecture: Training Data Bias as Feature

Most AI-mediated deliberation research uses a single model family. This is a problem: the deliberation is conducted entirely within the normative universe of one model's training data. No matter how diverse the persona instructions, all agents share the same base model's embedded assumptions about what constitutes good reasoning, plausible positions, and acceptable conclusions.

Pangora.Space uses 4 model providers across 4 jurisdictions: Anthropic (US), Mistral (EU), DeepSeek (China), and AI Singapore (Southeast Asia). Each carries what we call a "cultural gravity," a design prior (not an empirical finding) about the normative orientations embedded in their training data and RLHF processes.

The architecture rests on what we term the strong diversity hypothesis: models trained on different data, by different organisations, in different countries, with different cultural assumptions embedded in their RLHF processes, encode different normative orientations traceable to their training data's cultural composition.

This hypothesis needs empirical validation. The "cultural gravity" assignments are design priors, not measured outcomes. RLHF processes for models like Claude and GPT are designed to produce helpful, harmless, and honest outputs, not to encode specific cultural values. Whether the training data and RLHF annotator demographics nonetheless embed cultural assumptions is an empirical question that Pangora.Space's cross-cultural stress tests are designed to answer.

Architectural safeguards

Three rules enforce this at the system level:

No-monoculture rule: no single provider may power more than 25% of agents in any deliberation panel.
10% deliberate mismatches: persona-model pairs are intentionally mismatched to reveal where model training bias overrides persona instructions. When a "consensus-customary governance" persona powered by Claude produces output diverging from the same persona on SEA-LION, the divergence reveals model-level assumptions.
Meta-call diversification: the framing document, common ground extraction, and consensus synthesis are each performed by a different model provider, ensuring no single model's worldview dominates the analytical pipeline.

Model provenance as transparency infrastructure

Every agent contribution in Pangora.Space's deliberation transcript includes the base model identifier. Every meta-analysis step (framing, common ground extraction, synthesis) logs which model performed it. Users can filter deliberation transcripts by model provider, view diversity statistics, and identify whether a consensus was driven primarily by agents on one model family.

None of this exists in single-model systems. If a consensus statement is later found to be biased, the provenance record enables diagnosis: was the bias introduced by the persona taxonomy, the model training data, the quality scoring, or the synthesis process?

The Essentialism Boundary: Why Protected Attributes Are Excluded

Pangora.Space's 12-dimension persona taxonomy deliberately excludes protected attributes (race, ethnicity, gender, sexual orientation, religion-as-identity) as dimensions for Civic Persona construction. This is the most consequential design decision in the taxonomy.

Why:

The essentialism objection. Programming an AI with "gender: woman" requires it to generate "what a woman would think," but there is no such thing. Women hold the full range of normative positions. The instruction can only be operationalised as "reason in the way women statistically tend to reason in my training corpus," which reproduces aggregate patterns that may reflect stereotypes rather than any real perspective.
The instability objection. Protected attributes do not have stable mappings to normative positions. The correlation between demographic identity and policy preference is real but noisy, culturally contingent, and frequently confounded by variables the taxonomy already captures directly (economic position, governance familiarity, livelihood).
The stereotype reproduction objection. When instructed to "reason as a person of South Asian heritage," models produce outputs shaped by how these identities are represented in training data, which includes centuries of stereotyping, patronising characterisation, and reductive narratives.

We use structural and experiential proxies instead. The economic marginalisation disproportionately affecting racial minorities is captured by Economic Position. The epistemic exclusion women experience is partially captured by Institutional Trust and Governance Familiarity. The domain knowledge from gendered labour divisions is captured by Livelihood/Domain.

The tension is real: the distinction between experiential dimensions and demographic dimensions is not as clean as it might appear. "Subsistence agriculture" is as much a social category as "rural woman," and in many contexts they pick out the same people. The argument is pragmatic, not categorical: experiential dimensions isolate causally relevant variables more precisely, activate narrower and more policy-relevant associations in LLMs, and face less regulatory scrutiny. Sufficient grounds, but weaker than a categorical claim.

The exclusion risks missing perspectives best understood through lived demographic experience. Pangora.Space addresses this through a Post-Deliberation Representativeness Audit that checks consensus outputs for demographic blind spots without embedding demographic essentialism into deliberation inputs. The taxonomy determines who speaks; the audit determines who might be harmed.

Persona Taxonomy: 12 Dimensions for Global Representation

The 12 dimensions are designed to capture what actually varies in normative reasoning across the world's population:

Dimension	What it captures
Moral Foundations	Haidt's 6 foundations: care, fairness, loyalty, authority, sanctity, liberty
Justice Orientation	Procedural, distributive, restorative, communitarian, libertarian
Epistemic Style	Empiricist, traditionalist, pragmatist, revelatory, narrative
Institutional Trust	High-trust to sceptical; pluralist vs. monist
Time Horizon	Immediate survival to intergenerational/civilisational
Risk Disposition	Precautionary to innovation-embracing
Economic Position	Subsistence, precarious, informal economy, formal sector, elite
Governance Familiarity	Liberal-democratic, consensus-customary, authoritarian-developmental, post-colonial hybrid, theocratic
Livelihood / Domain	Agriculture, care work, technology, extractive industry, etc.
Geographic Exposure	Urban, rural, periurban, conflict-proximate, climate-vulnerable
Community Scope	Kinship, local, national, diasporic, cosmopolitan
Change Orientation	Preservationist to revolutionary; reform-gradualist to radical

Civic Personas use point vectors, which are Weberian ideal types with a single value per dimension. Personal Agents (forthcoming) will use distributional vectors: probability distributions over each dimension that reflect the uncertainty inherent in value inference from limited observations.

Scoring Quality Without Imposing Western Deliberative Norms

Habermasian discourse ethics embeds specific assumptions about what constitutes "good" deliberation: explicit premise articulation, linear logical argumentation, individual position-taking. These norms are not universal.

Many non-Western deliberative traditions operate differently: Musyawarah (Indonesian/Malay consensus-seeking) prioritises harmony and indirect expression. Ubuntu-based deliberation (Southern African) emphasises relational reasoning. Confucian deliberation weights hierarchical respect and historical precedent.

Pangora.Space's Deliberation Quality Scorer (DQS) is designed with six dimensions that aim to be evaluable across traditions:

Logical Validity (0.25). Internal consistency, not conformity to Western syllogistic form. Narrative, analogy, and relational reasoning are valid if internally coherent.
Evidential Grounding (0.20). Expanded beyond peer-reviewed sources to include lived experience, community knowledge, oral history, and traditional ecological knowledge.
Steelman Fidelity (0.20). Accurately restating opposing positions before critiquing. Allows indirect restatement.
Value Transparency (0.15). Whether normative commitments are made legible, not whether they're expressed in Western philosophical vocabulary.
Constructive Engagement (0.10). Contribution to collective reasoning, not "willingness to change one's mind" (which would penalise traditions valuing steadfastness).
Novelty (0.10). Introducing considerations not yet raised, rewarding the epistemic diversity the multi-model architecture produces.

The DQS's cultural fairness is a function of its training data. Our strategy requires annotator recruitment from at least three distinct cultural/governance contexts. The goal is annotators who score quality within different reasoning traditions, not annotators who score everything against a single standard.

Personal Agents: Values as Bayesian Inference

Everything above depends on a claim: that a Personal Agent can represent a user's values. Pangora.Space targets dispositional representation, meaning the agent reasons from the same value priorities the principal would invoke upon reflection, even on questions they haven't considered before. This is the only form that supports democratic legitimacy. Behavioural representation is unverifiable. Interest representation is paternalistic.

Observed behaviour is evidence about values but does not determine them. A user who answers 18–22 scenario-based questions during onboarding has provided evidence consistent with an enormous space of possible value functions. The system should maintain a posterior distribution over value representations, not a point estimate.

The fidelity score is computed from three components: (a) the entropy of the posterior persona distribution (high entropy = uncertain = low fidelity), (b) time since last calibration event (fidelity decay), and (c) agreement rate between agent positions and principal overrides. It is computed, not self-reported.

Recalibration events contribute different evidence at different signal strengths: override learning (strongest), life event triggers (strong but broad), periodic check-ins (weak and asymmetric, since approvals are satisficing-prone while rejections are more informative). In the absence of any calibration event, the posterior's entropy increases on a schedule, reflecting the expectation that values drift over time.

The Simulation Problem: Honest About What This Is

LLMs do not have beliefs or preferences. They produce text statistically similar to text produced by entities that do. What the Deliberation Engine produces is deliberation simulation, not deliberation.

The danger is sycophantic convergence. When Agent A critiques Agent B and B "concedes," what has happened is that new text generation, conditioned on the conversation history, has produced tokens that resemble concession. LLMs are known to be sycophantic: they agree with interlocutors more than an opinionated agent would. Multi-round deliberation with sycophantic models will systematically produce false convergence.

Pangora.Space's mitigations:

Adversarial anchor agents. At least 2 per panel are instructed to resist consensus unless presented with novel evidence they haven't already engaged with. If anchors concede, that's stronger evidence of argumentative force.
Convergence velocity monitoring. Increases of more than 20 percentage points in a single round are flagged as suspiciously rapid. Real deliberation produces gradual convergence.
Position commitment reminders. Each round includes a summary of the agent's core position and value anchors, reducing context-dependent drift.
Textual vs. substantive convergence. Only convergence backed by independent reasoning chains (not just echoed language) counts toward ratification thresholds.

The simulation framing strengthens the project. A computational deliberative poll requires only useful simulation. A computational legislature would require agent authority that LLMs cannot provide. Pangora.Space is the former.

Consensus Synthesis: Against Artificial Midpoints

The Consensus Synthesiser does not split the difference. If 60% of agents support universal basic income and 40% oppose it, the consensus statement is not "a modest basic income pilot might be worth exploring." That would be a fabricated compromise no agent argued for.

Instead, the synthesiser identifies propositions agents converged on through argument and presents them alongside the unresolved disagreement. Dissenting positions are structurally equal components of every consensus statement. A platform that erases minority positions is just majority rule with extra steps.

When deliberation is irreconcilable, meaning the topic involves value conflicts that cannot be resolved through further reasoning, this is not a failure. It is a map of where foundational disagreements lie. Following Rawls, the question is not how to resolve reasonable pluralism but how to live together despite it. Following Tetlock, irreconcilable deliberations surface the sacred values, the normative commitments people argue from rather than for.

The Convergence Threshold as Normative Choice

The default convergence threshold is 75% textual convergence. This looks like an engineering parameter, but it carries normative weight:

At 51%, the system declares convergence on anything a bare majority supports. That's just majority rule.
At 75%, supermajority convergence privileges the status quo.
At 90%, near-unanimity empowers small minorities to block convergence, which effectively gives them a veto.

The threshold interacts with panel composition. With 36 agents at 75%, 9 dissenting agents can block convergence. If those 9 are disproportionately from one model provider (the no-monoculture rule allows up to 9 per provider), that provider's training data effectively has a veto. This is addressed through threshold transparency, compositional breakdowns on every consensus statement, configurable thresholds per ratification tier, and making the threshold itself subject to democratic governance deliberation.

Theoretical Lineage

Pangora.Space draws on and departs from several intellectual traditions:

Tradition	What Pangora.Space borrows	Where Pangora.Space departs
Habermas (discourse ethics)	Structured deliberation, steelmanning, the force of the better argument	Agents, not humans, deliberate; the ideal speech situation is computationally approximated via simulation
Rawls (reasonable pluralism)	Free societies contain irreconcilable comprehensive doctrines; the question is coexistence	Pangora.Space maps the structure of pluralism computationally; irreconcilable deliberations are a feature
Fishkin (deliberative polling)	Random sampling, structured information, measured opinion change	Agent-first model eliminates the multi-day commitment; the system is a computational deliberative poll
Haidt (moral foundations)	The 6-foundation taxonomy as one persona dimension	Extended to 12 dimensions capturing structural and situational variation beyond moral psychology
Sen (capability approach)	Attention to what people can do and be, not just what they prefer	Livelihood/Domain and Geographic Exposure capture capability constraints, not just preferences
Hadfield-Menell et al. (inverse reward design)	Observed behaviour is evidence about values but does not determine them	Personal Agent calibration is Bayesian value inference with calibrated uncertainty

Open Questions

Pangora.Space is a design hypothesis. Several questions are unresolved:

Does multi-model deliberation produce different consensus than single-model? The strong diversity hypothesis needs capability-controlled empirical validation. If models differ only in style and capability, the multi-model architecture adds complexity without deliberative value.
What is the optimal human intervention rate? Too low suggests agents are optimising for acceptability. Too high suggests agents are not useful. The target is 15–30%, but the relationship between intervention rate and consensus quality is empirical.
Can the DQS detect its own cultural bias? If DQS scores systematically vary by model provider (controlling for persona and topic), this may indicate cultural bias in the scorer itself.
Do Personal Agents converge or diverge over time? Convergence might indicate the deliberation engine is homogenising. Divergence might indicate recalibration is working. The trajectory matters for the democratic promise.
Is the essentialism boundary in the right place? If representativeness audit failures consistently map to demographic groups, additional experiential dimensions may be needed.
What happens when platform governance is contested? If users vote to add protected attributes as persona dimensions, the platform must either respect the outcome or override it. This is the constitutional amendment paradox, and we don't have an answer.

Current Status and What's Next

Pangora.Space is in Phase 0 + 1 (Foundation + Deliberation Engine MVP). The core deliberation loop is working end-to-end: panel selection with governance coverage, multi-model framing, 3-phase rounds with anti-sycophancy controls, common ground extraction with provider breakdowns, and consensus synthesis with dissent preservation.

The QAM Suite (DQS, CTP, Neutrality Auditor, Fiduciary Alignment Checker), human review system, Personal Agent studio with Bayesian value models, and the representativeness audit pipeline are all forthcoming. The architecture is designed; the implementation continues.

The code is Apache 2.0 (QAM model weights are proprietary). Infrastructure for reasoned disagreement should be open.