Visioning Assembly

Origin

The Visioning Assembly is modelled on the 2010 Icelandic National Forum — where 950 randomly selected citizens gathered to deliberate on constitutional values for the Parliament of Iceland. That assembly was facilitated by Agora, the creator of BASAL, under a commercial agreement with the Icelandic Parliament.

The methodology took two years to develop, with hundreds of test runs conducted within the Icelandic government from early 2009 through the National Forum in Q4 2010. BASAL's digital implementation preserves the deliberative structure that made the original assembly work at national scale — structured card-based sessions, parallel group topology, theme-based redistribution, and constitutional-style outputs.

How it Works

The assembly brings together 7–950 participants across parallel groups, each guided by a dedicated facilitator agent, to explore values, vision, obstacles, and recommendations through structured card-based sessions.

Unlike simple ceremonies that run a topic through a few rounds, the assembly requires an initialization document that defines the full scope — scale, session questions, knowledge sources, organizational identity, and participant pool.

The Four Sessions

Session 1 — Values

Participants sit at home tables (up to 9 per group). Each person plays cards — one idea per card — in response to the session question. This is divergent thinking: generate, don't evaluate. Cards are voted on to surface the most resonant values.

Session 2 — Vision per Theme

After values processing derives cross-group themes, participants redistribute into theme-based zones. Each zone builds a bold vision statement for its theme area, grounded in knowledge sources. What does the evidence suggest is possible? What would make this organization proud in 10 years?

Session 3 — Obstacles & Risks

Participants return to home tables, bringing insights from the theme zones. Now they stress-test the emerging vision. What could go wrong? What are we ignoring? What would a harsh critic say? Senior voices speak first here — experience identifies structural barriers.

Session 4 — Recommendations

The final session synthesizes everything — values, vision, obstacles — into concrete recommendations. Each recommendation must reference a specific vision theme AND a specific obstacle, be actionable (who does what, by when), and be grounded in evidence.

Knowledge-Backed Initialization

What makes the Visioning Assembly fundamentally different from any multi-agent brainstorming system is that participants are not blank prompts — they are backed by BASAL's knowledge graph.

Participants Are People

Assembly participants are drawn from BASAL's People — real individuals tracked in the knowledge graph with their projects, relationships, expertise areas, communication patterns, and decision history. When a digital twin of "Satish Sonwane" enters the assembly, it carries Satish's actual project involvement, his relationships with other team members, his technical positions from prior meetings, and his domain expertise derived from the content BASAL has processed about him.

Beyond digital twins, the assembly draws on a second dimension: recognized thought leaders, domain intellectuals, and established frameworks that state-of-the-art LLMs have deep, research-backed knowledge of. When the topic is organizational strategy, the roster might include an archetype grounded in Clayton Christensen's disruption theory or a voice channeling Deming's systems thinking — not as shallow personas, but as deeply informed perspectives that the LLM can argue from with genuine intellectual depth. Mixing organizational insiders (digital twins with real context) and external intellectual traditions (archetypes backed by vast training data) creates a deliberation that neither source could produce alone.

This means each participant argues from a position informed by real organizational context, external intellectual frameworks, or both — not from a generic role description.

Memory-Grounded Deliberation

Each participant has access to BASAL's entity memory relevant to the assembly topic. When the question is about architecture decisions, participants draw on the actual technical decisions, project histories, and organizational patterns stored in the knowledge graph. The assembly initialization phase queries topics, entities, and facts relevant to the question and distributes them as knowledge connections — typically 10-20 per participant.

Crucially, this memory extends beyond structured graph data. Through BASAL's edge-connectivity model, participants can access the organization's internal source code, deep research documentation, technical specifications, and proprietary knowledge artifacts that BASAL has processed locally. This is information that exists only within the organization's boundary — never exposed to external services, never in any public training set. When a participant references "the rate limiting implementation in the proxy worker," they are drawing on actual source code that BASAL has ingested and indexed, not hallucinating plausible-sounding technical details.

This creates a depth of deliberation that is qualitatively different from LLM agents operating on a system prompt alone. Participants don't generate plausible-sounding arguments; they argue from evidence rooted in the organization's actual history, codebase, and internal documentation.

Automatic Roster Construction

The assembly auto-generates its roster from the knowledge graph:

1. Digital twins of real people — loaded with their graph profile (projects, relationships, expertise, communication style)
2. Advisor archetypes — auto-invited when the roster analysis detects expertise gaps (e.g., no financial voice triggers a CFO archetype invitation). These archetypes leverage the LLM's deep knowledge of recognized thought leaders and established frameworks — a "VP Engineering" archetype doesn't just role-play, it argues from genuine engineering leadership principles
3. Knowledge connections — per-participant facts, entities, source code references, and internal documentation relevant to the topic, queried from the graph and local knowledge stores at init time

The result: a deliberation where 7 participants with 12 knowledge connections each brings 84 distinct organizational data points into the room — before anyone has said a word.

Technical Architecture

7-Stage Adversarial Output Pipeline

Every assembly report passes through a pipeline where each stage cross-checks the previous one — no black-box summarization:

Surprisingly Popular (SP) Voting

Implementation of Prelec et al.'s "Bayesian Truth Serum" (Nature 2017), adapted for LLM agent deliberation (arXiv:2510.01499). After standard dot-voting, each participant predicts how others voted. SP Score = actual votes - average predicted votes. A high SP score reveals ideas the majority underestimated — exactly the non-obvious insights that groupthink suppresses. These findings receive a credibility boost in the Judge stage.

Glicko-2 Rating System

Full implementation of Glickman's Glicko-2 algorithm tracking participant idea quality across assemblies. The 6-step algorithm includes Illinois method volatility solving (iterative regula falsi, convergence tolerance 1e-6). To prevent the Matthew Effect, newcomers always weight 1.0 (neutral), and veterans are clamped to ±25%. Rating deviation increases during inactivity, ensuring dormant participants don't carry stale authority.

PageRank Influence Scoring

Power iteration PageRank (damping 0.85, 50 iterations) runs on the idea evolution graph: cards → cards (deliberation edges), cards → findings (cited in why-chains), findings → recommendations. This traces how a participant's original card ripples through the entire assembly into final recommendations — measuring genuine intellectual influence, not just vote counts.

Cognitive State Machine

A 7-state finite-state machine tracks each participant's cognitive mode in real-time:

• present_mind → reflective → tom_inference → belief_revision → acting → verification → error_recovery

Hysteresis prevents state thrashing (minTurnsInState). Archetype-aware overrides customize the FSM: guardians lock into verification mode, skeptics into belief_revision, visionaries into acting. The machine maps session progress (early/mid/late) to natural cognitive arcs for each thinking mode (creative, analytical, critical, integrative).

The Theory of Mind (ToM) inference state forces participants to steel-man opposing viewpoints before contributing — the assembly doesn't let agents avoid uncomfortable ideas.

Agency-Enabler Facilitation

The facilitator agent combines Scrum Master (process protection) with Editorial Director (implied meaning discovery). It monitors 7 agency drivers: stakes, autonomy, provocation, recognition, creative safety, purpose, momentum. Critically, it detects brainstorming theater — when participants produce fluent but hollow contributions — and tightens provocations to extract genuine ideas.

When participant B "misunderstands" participant A's card and creates a novel recombination, the facilitator amplifies the productive misunderstanding instead of correcting it — recognizing these moments as a valuable source of unexpected insight in deliberation.

Flow state guardianship monitors Csikszentmihalyi's flow conditions (clear goals, immediate feedback, challenge/skill balance). The facilitator knows flow typically emerges in rounds 3-4 after trust builds, and protects it by reducing intervention during those rounds.

Groupthink Detection

A 4-signal model continuously monitors deliberation health:

1. Cascade voting: Top card receiving >70% of votes = premature consensus risk
2. Low dissent edges: <5% challenge relationships in deliberation graph = echo chamber
3. Low kill rate: <10% of findings affected by adversarial challenge = weak red-team
4. Low dissent findings: <10% of findings with counter-arguments = suppressed minority

When groupthink risk exceeds threshold, the challenge stage temperature increases from 0.3 to 0.5 and contrarian seeds from high-vote minority cards are injected.

Grounding Calibration

LLMs systematically inflate confidence scores (the "everything is 0.95" problem). Post-distill recalibration detects when >50% of findings have near-perfect grounding and applies sqrt dampening to spread scores across a realistic [0, 0.85] range. Cross-assembly z-score normalization with sigmoid squash ensures scores are comparable across different assembly runs.

Dynamic Red-Team Personas

The challenge stage generates a consensus-specific adversarial persona using PersonaTeaming (arXiv:2509.03728). It analyzes the top findings, identifies the single most dangerous shared blind spot, and creates a challenger specifically designed to attack that weakness. This yields 144% higher attack success than fixed personas because it focuses on what the specific consensus is most vulnerable to.

Speaking Order & Visibility

The assembly uses deliberate speaking order to manage power dynamics:

• Junior-first in generative phases — newer voices set the creative direction before established perspectives can anchor the discussion
• Senior-first in stress-test phases — experienced voices identify structural barriers and institutional risks others might miss
• Random in warm-up and plenary — equal footing for introductions and final reactions

Visibility progresses through the ceremony. Early phases use blind visibility (participants see only their own group's cards). Later phases open to all-phases visibility as the assembly builds shared context.

Scale & Topology

The assembly supports dynamic group topology with sampling-theory-informed composition:

• Home tables: Initial groups of up to 9. Composition uses stratified random sampling (stratified SRS) across category, archetype, seniority, and geography — ensuring each table is a representative microcosm of the full assembly, not a convenience cluster. When the full population is available, simple random sampling (SRS) serves as the baseline; stratification is applied to guarantee demographic and cognitive diversity within each group.
• Theme zones: Mid-ceremony redistribution based on voting patterns — participants move to the theme they care about most. This self-selection creates interest-aligned groups that produce deeper domain insights.
• Plenary: Full-assembly phases where all participants see and react to synthesized results from all groups — the mechanism that converts parallel deliberation into collective intelligence.
• Zone masters: Appointed when 5+ groups share a theme zone, coordinating cross-group coherence and preventing echo chambers within zones.

At National Forum scale (950 participants), this means ~106 parallel groups, each with a facilitator agent, coordinated by zone masters across theme areas.

The 19 Phases

The assembly runs through 19 structured phases mixing discussion, voting, processing, transitions, and presentations:

What Makes This Different

Most multi-agent deliberation systems treat LLMs as interchangeable message generators in a chat loop. The Visioning Assembly is fundamentally different:

• Provenance is first-class. Every recommendation traces through why-chains to quoted card text. No black-box "the AI thinks." Full 7-stage pipeline with typed, inspectable output at every stage.

• Agency is earned, not given. The cognitive state machine, archetype-aware FSM overrides, and facilitation patterns create conditions where agents produce genuine insight — not just fluent text. Brainstorming theater detection actively filters for depth.

• Adversarial by default. The red-team challenge stage with rotating + dynamic personas attacks every finding. Killed findings are removed. Weakened findings get reduced scores. No consensus passes unchallenged.

• Minority voices are protected. Surprisingly Popular voting surfaces underestimated ideas. Minority reports preserve strong dissent with full context. Junior-first speaking order prevents anchoring. Sparse deliberation preserves position diversity.

• Cross-assembly learning. Glicko-2 ratings track participant idea quality over time. PageRank measures genuine intellectual influence. Historical comparison detects which recommendations persist, which are new, and which were abandoned.

• Scale-tested methodology. The deliberation structure ran at national scale (950 citizens) in Iceland. The digital implementation preserves the topology, speaking order, and constitutional-style output that made it work.

Token Usage at Scale

Cost scales with participant count, group count, and number of discussion phases. The base overhead is ~20k tokens, plus ~5k per participant and ~8k per group.

Output

The assembly produces constitutional-style artifacts:

• Vision statement — synthesized from cross-group convergence, with value pillars
• Ranked recommendations — scored by cross-group support strength, with first steps and success criteria
• Minority reports — strong dissenting positions preserved with full context
• Risk register — severity x likelihood matrix with mitigation linkage
• Challenge debrief — red-team effectiveness metrics (kill rate, dominant attack type, per-category breakdown)
• Health metrics — groupthink risk, cognitive quality, provenance integrity, dissent representation, cross-group convergence, SP boost count
• Provenance index — which evidence grounded which recommendations through which findings

Running an Assembly

basal arena ceremony "<topic>" --plugin visioning-assembly --workspace <id> \
  --init path/to/init.yaml

The --init flag points to a YAML document that defines the full assembly scope. Without it, the system falls back to a standard ceremony. The init document is Zod-validated at load time — malformed fields produce actionable error messages.

Full Init Document Schema

version: "1"
name: "Q2 Strategic Vision"
topic: "What strategic direction should guide our next decade?"
description: |                              # optional — rich context for facilitators
  Additional context about the assembly's purpose,
  the organizational moment, and what success looks like.

# ── Organization Identity (optional) ──
organization:
  name: "Acme Corp"
  north_star: "Make every team autonomous"  # optional
  mission: "..."                            # optional
  culture: "..."                            # optional
  structure: "..."                          # optional

# ── Scale ──
total_participants: 90                      # 7–500
participants_per_group: 9                   # 5–15 (default: 9)
num_themes: 5                               # 2–20 (optional, auto-derived if omitted)

# ── Participant Pool (optional) ──
# When omitted, BASAL auto-generates from the knowledge graph
participant_pool:
  path: workspace/people/                   # directory with participant profiles
  count: 90
  backgrounder_depth: full                  # full | summary | archetype_only

# ── Session Questions (2–12 required) ──
questions:
  - session: 1
    question: "What values should guide our next decade?"
    thinking_mode: analytical                # analytical | creative | critical | integrative
    card_color: blue                         # optional — visual card color per session
    context: "Additional context for this specific question"  # optional
  - session: 2
    question: "What bold vision emerges from these values?"
    thinking_mode: creative
    card_color: green
  - session: 3
    question: "What could prevent us from realizing this vision?"
    thinking_mode: critical
    card_color: red
  - session: 4
    question: "What concrete steps should leadership take?"
    thinking_mode: integrative
    card_color: purple

# ── Knowledge Sources (optional) ──
# Filesystem paths that participants can reference during deliberation
knowledge_sources:
  - label: "Strategy Documents"
    path: workspace/strategy/
    type: documents                         # documents | codebase | research | data | media
    description: "Q1-Q4 strategy decks"     # optional
    include: ["*.md", "*.pdf"]              # optional glob filters
    exclude: ["drafts/"]                    # optional exclusions
    max_depth: 3                            # optional directory depth limit
  - label: "Quarterly Reviews"
    path: workspace/quarterly-reviews/
    type: documents

# ── Doctrine Lens (optional) ──
# Focus the assembly through a specific analytical framework
doctrine_lens: "operational-excellence"

# ── Custom Doctrines (optional) ──
# Define domain-specific analytical frameworks inline
custom_doctrines:
  operational-excellence:
    name: "Operational Excellence"
    description: "Focus on operational efficiency and reliability"
    entity_types: ["Process", "Metric", "Team"]
    fact_types: ["performance_metric", "bottleneck", "improvement"]
    graph_queries:                           # optional
      - label: "Bottleneck analysis"
        description: "Find operational bottlenecks"
        pattern: "MATCH (p:Process)-[:HAS_BOTTLENECK]->(b) RETURN p, b"

# ── Timing (optional) ──
session_duration_minutes: 30                # per-session time budget
voting_duration_minutes: 5                  # per-voting-round time budget
votes_per_participant: 5                    # override default vote allocation

# ── Output Options (optional) ──
output:
  per_group_reports: true                   # generate per-group journey reports
  full_transcript: false                    # include raw card text in output
  theme_trace: true                         # include theme derivation trace