{"path":"academic-foundations.md","content":"# Academic Foundations for Deliberus\n\nA synthesis of research papers informing the design of a collaborative deliberation platform. Compiled March 2026.\n\n**A newer synthesis integrating non-zero-sum coordination and civilizational cooperation is available in [research/non-zero-sum-economics-and-civilizational-cooperation.md](research/non-zero-sum-economics-and-civilizational-cooperation.md).**\n\n---\n\n## Paper Summaries\n\n### 1. Scheuer et al. -- \"Computer-Supported Argumentation\" (Springer, ~2010)\n\n**Source:** FERMI drive, `Argument mapping/Computer-supported argumentation.pdf`\n\nA comprehensive survey of computer-supported argumentation systems across educational and collaborative contexts.\n\n**Key findings:**\n- Five major argument representation formats exist: linear (chat), threaded (forums), graph-based (node-and-link), container, and matrix. Each has distinct tradeoffs between expressiveness and usability.\n- Graph representations are the most expressive and widely used, but suffer from scalability problems -- large argument maps become \"spaghetti images\" that overwhelm users.\n- Ontology design is critical: too fine-grained (Rationale's 30 types) confuses users; too coarse (Athena's unclassified links) limits analysis. The sweet spot is ~10 extensible types with labeling via titles rather than palette selection.\n- Three scales of collaborative argumentation: single-user, small group (2-5, synchronous), and community (hundreds+, asynchronous). Community systems need threaded organization over graphs.\n- Automated analysis can check argument quality via domain-specific patterns, problem-specific patterns, and simulated reasoning (acceptability scoring).\n\n**Relevance to Deliberus:** This is the most directly applicable survey. It maps the entire design space Deliberus must navigate: representation format, ontology granularity, collaboration scale, and automated analysis. The scalability problem of graph representations is the central UI challenge.\n\n---\n\n### 2. Monk & van Gelder -- \"Enhancing Our Grasp of Complex Arguments\" (Austhink, 2009)\n\n**Source:** FERMI drive, `Argument mapping/enhancing-our-grasp-of-complex-arguments.pdf`\n\nA keynote address making the case that argument mapping is a cognitive necessity, not a luxury.\n\n**Key findings:**\n- Human working memory is severely limited (~4 items), yet deliberative processes demand holding entire argument structures in mind simultaneously. This is like playing tic-tac-toe on a 4x4 grid without a board -- the cognitive load is unsustainable.\n- Three compounding factors degrade deliberation: (1) cognitive biases (confirmation bias, belief preservation), (2) prose as a medium obscures argument structure, (3) disciplinary jargon creates communication silos.\n- Argument maps provide seven cognitive advantages over prose: explicit logical relationships, instant structural scanning, easy detail-to-overview movement, visual significance cues, visible debate boundaries, reduced cognitive load, and integrated claims.\n- The analogy to geographic maps is precise: just as maps replaced prose descriptions for navigation (enabling the first atlas in 1570), argument maps replace prose for reasoning. Both are technologies for externalizing information that exceeds working memory.\n- The difficulty of creating an argument map is a direct measure of how unclear your thinking is. Prose lets vagueness masquerade as profundity.\n\n**Relevance to Deliberus:** Provides the core philosophical justification for why a deliberation platform must go beyond threaded discussion. The working memory argument is the strongest pitch for visualization-first design. The \"difficulty = unclear thinking\" insight suggests the platform should make argument mapping feel natural, not academic.\n\n---\n\n### 3. van Gelder -- \"What Is Decision Mapping?\" (Austhink, 2010)\n\n**Source:** FERMI drive, `Argument mapping/What is DecisionMapping.pdf`\n\nA practitioner-oriented brief explaining decision mapping as an extension of argument mapping for deliberative choices.\n\n**Key findings:**\n- Decision mapping has three components: (1) making explicit the questions, options, pros/cons, arguments, and evidence; (2) producing a visual map using diagramming conventions; (3) using the map to evaluate and decide.\n- Decision mapping is designed specifically for \"deliberative decisions\" -- choices made by weighing arguments -- as opposed to intuitive (snap) or technical (quantitative) decisions. Most important decisions are deliberative.\n- The map acts as \"external RAM\" -- offloading cognitive complexity to a visual artifact, freeing mental resources for evaluation rather than information retention.\n- Decision mapping reduces decision anxiety, creates clear records of reasoning, improves communication among teams, and keeps collaborators \"on the same page.\"\n- Distinct from mind mapping: decision mapping has unique rules, conventions, and procedures tailored to decision-making, and is more rigorous.\n\n**Relevance to Deliberus:** Decision mapping is the applied cousin of argument mapping. Deliberus should support both pure argumentation (exploring a topic) and decision mapping (choosing between options). The distinction between deliberative, intuitive, and technical decisions helps scope what the platform is for.\n\n---\n\n### 4. Interview with Tim van Gelder (The Reasoner, Vol. 4 No. 2, February 2010)\n\n**Source:** FERMI drive, `Argument mapping/Interview with Tim van Gelder - TheReasoner-4(2).pdf`\n\nAn interview covering van Gelder's intellectual journey from AI to intelligence augmentation (IA), and the practical state of argument mapping.\n\n**Key findings:**\n- Argument maps are \"semi-formal\" -- blending formal graph structure with natural language. This is a design challenge: maximize transparency and comprehensibility while keeping reasoning explicit.\n- Maps compensate for limited short-term memory by providing a stable external representation of complex inferential webs, exploiting human visual scanning capacities. \"Our eyes constitute the high-capacity bus connecting the argument map to our brains as the CPU.\"\n- Van Gelder achieved 2-3x improvement on California Critical Thinking Skills Test using the \"Deliberate Practice Hypothesis\" (Ericsson) -- lots of structured practice with argument mapping tools.\n- Key limitation: argument maps blend formal structure with natural language, so they are not amenable to computation (can't automatically infer or update truth values). For that, you need Bayesian networks.\n- The shift from AI to IA is pivotal: rather than building artificial reasoners, augment human reasoning capacity. Argument mapping is the simplest and most impactful tool for this.\n\n**Relevance to Deliberus:** The IA framing is exactly Deliberus's mission. The limitation about computational amenability is crucial -- pure argument maps support human cognition but resist automation. Deliberus should explore hybrid approaches: human-readable argument maps backed by computational structures (Bayesian networks, knowledge graphs) for AI-assisted analysis.\n\n---\n\n### 5. Yang -- \"Visualizing Argumentation: Software Tools for Collaborative and Educational Sense-Making\" (Book Review, 2003)\n\n**Source:** FERMI drive, `Health Psychology/Social & Collective/tools for collaborative sense-making, argument visualization.pdf`\n\nA review of the edited volume by Kirschner, Buckingham-Shum, and Carr (Springer, 2003).\n\n**Key findings:**\n- The book traces a lineage from Vannevar Bush (1945) and Douglas Engelbart (1962) through IBIS and Toulmin to modern CSAV (Computer-Supported Argumentation Visualization) tools.\n- \"Wicked problems\" (design, policy, collaborative) resist standard computational approaches; CSAV tools help groups construct shared understanding through externalized argument structures.\n- Empirical evidence: groups using planning tools (Diagram, Advisor, Outline) produced more structured dialogue than control groups.\n- CSAV enhances deliberation in three ways: defines deliberation as deciding where we stand on claims in light of arguments; enhances critical thinking; and improves group decision support system quality.\n- Dialog Mapping (IBIS) was used for 10 years by ~50 users at Southern California Edison for planning and policy design.\n- The social-technical gap remains the central challenge: technology offers communicative suppleness but lacks support for sharing information, roles, and social policies.\n\n**Relevance to Deliberus:** Validates the core hypothesis that visualization tools improve group deliberation quality. The 10-year SCE case study is rare evidence of sustained institutional adoption. The social-technical gap warning is critical: Deliberus needs social design (roles, moderation, contribution norms) alongside technical design.\n\n---\n\n### 6. Chesnevar et al. -- \"Towards an Argument Interchange Format\" (Knowledge Engineering Review, 2007)\n\n**Source:** FERMI drive, `Health Psychology/Social & Collective/towards argument interchange format.pdf`\n\nA consensus specification for an Argument Interchange Format (AIF) enabling interoperability between argumentation tools and multi-agent systems.\n\n**Key findings:**\n- The AIF core ontology defines three groups: Argument Networks (arguments and relations), Communication (locutions and protocols), and Context (participants and theory).\n- Two kinds of nodes: I-nodes (information/claims) and S-nodes (scheme applications -- inferences, preferences, conflicts). I-nodes cannot connect directly to I-nodes; there must always be a reasoning scheme mediating the connection.\n- Three types of schemes: inference schemes (RA-nodes), preference schemes (PA-nodes), and conflict schemes (CA-nodes). This captures support, attack, and preference relations in a unified graph.\n- Node attributes include title, text, creator, type, creation date, evaluation (strength), acceptability, and polarity (pro/con).\n- The AIF supports Toulmin's scheme as a special case: data, warrant, backing, qualifier, rebuttal, and claim all map to I-nodes and S-nodes.\n- Three concrete reifications demonstrated: ASPIC inference engine (XML), Araucaria (AML, supporting Toulmin and Wigmore styles), and RDF/RDFS (Semantic Web).\n\n**Relevance to Deliberus:** The AIF is the most important technical reference for Deliberus's data model. The I-node/S-node distinction -- where every connection between claims must be mediated by an explicit reasoning scheme -- enforces argument quality at the structural level. Deliberus should adopt or extend AIF as its internal representation. The RDF reification is directly relevant if Deliberus uses a knowledge graph backend.\n\n---\n\n### 7. Mercier & Sperber -- \"Why Do Humans Reason? Arguments for an Argumentative Theory\" (~2011)\n\n**Source:** FERMI drive, `MercierSperberWhydohumansreason.pdf`\n\nA landmark paper in evolutionary cognitive science proposing that the primary function of human reasoning is argumentative -- to produce and evaluate arguments in social communication.\n\n**Key findings:**\n- The function of reasoning is NOT to improve individual knowledge or decisions. It is to devise arguments to persuade others and to evaluate arguments received from others. Reasoning evolved for social argumentation.\n- Confirmation bias is not a bug but a feature: when reasoning's function is producing persuasive arguments for your position, seeking only confirming evidence is adaptive.\n- People are skilled arguers in natural contexts: they can evaluate argument strength, spot fallacies (with appropriate sensitivity to context), and produce sophisticated argumentative structures -- but only when the context is genuinely argumentative and personally relevant.\n- Group reasoning dramatically outperforms individual reasoning. On the Wason selection task, individual success is ~6.25% (chance level); groups achieve ~80%. This is because groups enable the exchange and critical evaluation of arguments.\n- \"Truth wins\" in groups: as soon as one person has the correct answer and can argue for it, the group converges. Groups can also achieve \"assembly bonus\" -- collectively arriving at correct answers that no individual member had.\n- Group polarization occurs when all members already agree (no genuine debate); depolarization occurs when members genuinely disagree and exchange arguments.\n\n**Relevance to Deliberus:** This is the theoretical foundation for why Deliberus should work. Reasoning is inherently social and argumentative. Individual reasoning is biased and weak; group deliberation with genuine disagreement is powerful. The platform should be designed to maximize genuine argument exchange and minimize echo-chamber dynamics. The group polarization finding is a critical design constraint: the platform must ensure diverse viewpoints are represented, not just aggregate agreement.\n\n---\n\n### 8. Shneiderman -- \"Managing the Social Dynamics of Collective Intelligence\" (Slide Deck, 2012)\n\n**Source:** FERMI drive, `collectiveintelligence-2012-shneiderman-v2.pdf`\n\nA presentation on technology-mediated social participation for national priorities (health, education, government, environment).\n\n**Key findings:**\n- Vision for social participation technology spans three areas: national priorities (disaster response, health, education, governance), deep science (understanding social network evolution), and extreme technology (scalable, reliable, privacy-preserving systems).\n- Participation follows a progression: Reader -> Contributor -> Collaborator -> Leader. Systems should design pathways through these stages.\n- Key challenges: malicious attacks, privacy violations, lack of trust, failure to be universal, unreliability, and misuse by bad actors.\n- Tools like NodeXL enable network visualization and analysis of social dynamics (e.g., mapping US Senate voting patterns).\n- Community Response Grids (911.gov concept) show how professionals can work with empowered residents -- blending expert authority with crowd participation.\n\n**Relevance to Deliberus:** The Reader-to-Leader framework is directly applicable for designing participation tiers. The challenges listed (trust, privacy, misuse) are the governance problems Deliberus must solve. The professional-resident collaboration model (911.gov) suggests Deliberus might blend expert moderators with crowd participants.\n\n---\n\n### 9. Predictive Liquid Democracy\n\n**Source:** File not found on FERMI volume.\n\nThis paper was not located in the archive. It likely covers mechanisms for delegating voting power transitively in democratic systems, with predictive elements.\n\n---\n\n## Interrelationship Map\n\n```\n                    THEORETICAL FOUNDATION\n                    =====================\n\n    Mercier & Sperber (7)          Monk & van Gelder (2)\n    \"Reasoning is social\"          \"Working memory is the bottleneck\"\n           |                              |\n           |  WHY deliberation works      |  WHY visualization is needed\n           v                              v\n    +--------------------------------------------------+\n    |            DELIBERUS CORE THESIS                  |\n    |  Group argumentation + visual externalization     |\n    |  = better collective decisions                    |\n    +--------------------------------------------------+\n           |                              |\n           |                              |\n    SYSTEM DESIGN                  DATA MODEL\n    ============                   ==========\n\n    Scheuer et al. (1)            AIF / Chesnevar et al. (6)\n    \"Design space survey\"          \"I-nodes, S-nodes, schemes\"\n           |                              |\n    van Gelder (3,4)              van Gelder (3)\n    \"Decision mapping\"            \"Decision mapping ontology\"\n           |                              |\n    Yang/Kirschner (5)\n    \"CSAV + social design\"\n           |\n    Shneiderman (8)\n    \"Participation tiers\"\n```\n\n### Cross-Paper Connections\n\n| Connection | Papers | Insight |\n|---|---|---|\n| Cognitive bottleneck + social reasoning | 2 + 7 | Working memory limits explain why individuals reason poorly; social exchange of arguments compensates by distributing cognitive load |\n| Ontology design + AIF | 1 + 6 | Scheuer's survey of ontology tradeoffs (granularity, domain-specificity) is resolved by AIF's core+extension architecture |\n| Visualization + deliberation quality | 2 + 5 | Monk/van Gelder's theoretical argument for maps is empirically validated by the CSAV case studies |\n| Group dynamics + participation design | 7 + 8 | Mercier/Sperber explain why groups outperform individuals; Shneiderman provides the Reader-to-Leader framework for nurturing that participation |\n| Decision vs. argument mapping | 3 + 1 | Decision mapping (van Gelder) is a specific mode within the broader design space (Scheuer). Deliberus needs both general argumentation and decision-focused views |\n| Computational limits of maps | 4 + 6 | Van Gelder notes argument maps resist computation (natural language); AIF provides the formal structure needed for machine processing. These are complementary layers |\n| Confirmation bias as design constraint | 7 + 2 | Mercier/Sperber show confirmation bias is inherent to reasoning; Monk/van Gelder argue maps help counteract it by making the full argument structure visible |\n\n---\n\n## Most Actionable Ideas for Deliberus\n\n### Tier 1: Core Architecture Decisions\n\n1. **Adopt AIF as the data model** (Paper 6). The I-node/S-node distinction with explicit reasoning schemes is the right abstraction. It enforces argument quality structurally and enables interoperability with future tools.\n\n2. **Dual-layer representation** (Papers 4, 6). Human-readable argument maps on top of machine-processable AIF structures underneath. This resolves van Gelder's observation that maps resist computation -- the visualization is for humans, the graph is for AI.\n\n3. **Support both argument mapping and decision mapping modes** (Papers 1, 3). Arguments explore what is true; decisions choose what to do. Different views of the same underlying data.\n\n### Tier 2: Interaction Design\n\n4. **Graph representation with aggressive scalability solutions** (Papers 1, 2). Graphs are the right default, but Deliberus must solve the spaghetti problem. Techniques from the survey: filtering by acceptability threshold, local views (DebateGraph-style focus+context), mini overview maps, and semantic zoom.\n\n5. **Reader-to-Leader participation tiers** (Paper 8). Design for passive reading, active contributing (adding claims/evidence), collaborative structuring (organizing argument maps), and leadership (moderating, synthesizing).\n\n6. **Moderate ontology granularity** (Papers 1, 6). Start with ~8-10 extensible types. Use AIF's core+extension model: claims, evidence, inference, conflict, preference as core; domain-specific types as extensions.\n\n7. **Design for genuine disagreement** (Paper 7). The platform must actively surface opposing viewpoints and counterarguments. If all participants agree, reasoning degrades (group polarization). Consider adversarial pairing, mandatory counterargument prompts, or devil's advocate features.\n\n### Tier 3: AI-Assisted Features\n\n8. **Automated argument analysis** (Paper 1). Use pattern detection for: circular arguments, missing rebuttals, unsupported claims, and argument strength assessment. Scheuer's taxonomy (domain-specific patterns, problem-specific patterns, simulated reasoning) provides the roadmap.\n\n9. **AI-assisted argument extraction** (Papers 2, 4). Natural language arguments in prose are hard to parse. Use LLMs to help users extract argument structure from existing documents, discussions, or meeting transcripts into the structured AIF format.\n\n10. **Epistemic vigilance support** (Paper 7). Mercier/Sperber's concept of \"epistemic vigilance\" -- evaluating both source reliability and argument coherence -- can be operationalized: source credibility indicators, coherence checking, and fallacy detection.\n\n---\n\n## Tensions and Open Questions\n\n### Formality vs. Accessibility\nPapers 1 and 4 identify a fundamental tension: formal argument structures enable analysis and computation, but informal natural language is how people actually think and communicate. Deliberus must find the right level of structure that improves reasoning without creating adoption barriers. The AIF's I-node/S-node model may be too formal for casual users.\n\n### Individual Bias vs. Group Wisdom\nMercier & Sperber (7) show individuals are hopelessly biased but groups are remarkably effective. However, groups only outperform when there is genuine disagreement. How does Deliberus ensure diverse participation without descending into unproductive conflict? The survey (1) notes that community-scale systems typically need more rigid structures (threads over graphs) to manage coordination -- but this may sacrifice the visual advantages that Monk/van Gelder (2) argue are essential.\n\n### Scale vs. Depth\nThe systems in Paper 1 split into small-group (deep, synchronous, graph-based) and community (broad, asynchronous, threaded). Can Deliberus bridge these? Decision mapping (Paper 3) works best in small teams; collective intelligence (Paper 8) works best at scale. The platform may need distinct modes for different scales.\n\n### Computational Argument Processing\nVan Gelder (Paper 4) notes argument maps blend natural language and structure, resisting computation. The AIF (Paper 6) provides formal structure, but the natural language content of nodes still requires NLP/LLM processing. In 2026, LLMs make this tractable -- but the question of how much AI assistance is helpful vs. how much short-circuits the critical thinking benefits (Paper 4's Deliberate Practice Hypothesis) remains open.\n\n### Motivation and Adoption\nShneiderman (Paper 8) identifies the participation challenge but the CSAV case studies (Paper 5) show most adoption has been in education or small organizational settings. Large-scale public argument mapping has rarely succeeded (DebateGraph being a notable exception). Deliberus needs a compelling \"why should I use this?\" answer that goes beyond \"it's better for reasoning.\"\n\n---\n\n## Later Literature Touchpoints\n\n### Procaccia, Konya, et al. -- \"Using Collective Dialogues and AI to Find Common Ground Between Israeli and Palestinian Peacebuilders\" (FAccT '25 / arXiv:2503.01769, 2025)\n\n**Key findings:**\n- Reports collective statements, including demands to world leaders, with **at least 84% agreement from participants on each side** in an AI-assisted dialogue with Israeli and Palestinian peacebuilders.\n- Utilizes **bridging-based ranking** (Diverse Approval / group-aware scoring) to surface statements that resonate across deep divide lines.\n- Proves that \"bridging\" is a computable signal that can identify areas of consensus previously invisible to participants.\n\n**Relevance to Deliberus:** Direct empirical validation of the bridging and convergence thesis. This paper provides a methodological template for Deliberus's ranking mechanisms and cooperation-discovery experiments.\n\n---\n\n### Klingefjord, Lowe & Edelman -- \"What are human values, and how do we align AI to them? (Moral Graph Elicitation)\" (arXiv:2404.10636, 2024)\n\n**Key findings:**\n- Introduces **Moral Graph Elicitation (MGE)**: using LLMs to interview humans and map their values into directional graphs (\"value A is wiser than value B in context C\").\n- Found that a representative sample of 500 Americans often felt well represented by the process and regarded the final moral graph as fair, even when their own value was not selected as wisest.\n\n**Relevance to Deliberus:** Provides a close precedent for the convergence thesis: values that appear irreconcilable as abstractions can be elicited, compared, and reconciled in context-aware reasoning graphs.\n\n---\n\n## Source File Locations\n\nAll PDFs are located in the FERMI drive Dropbox PDF Archive:\n\n| # | Paper | Path (relative to PDF Archive) |\n|---|---|---|\n| 1 | Scheuer et al. | `Argument mapping/Computer-supported argumentation.pdf` |\n| 2 | Monk & van Gelder | `Argument mapping/enhancing-our-grasp-of-complex-arguments.pdf` |\n| 3 | van Gelder | `Argument mapping/What is DecisionMapping.pdf` |\n| 4 | van Gelder interview | `Argument mapping/Interview with Tim van Gelder - TheReasoner-4(2).pdf` |\n| 5 | Yang (book review) | `Health Psychology/Social & Collective/tools for collaborative sense-making, argument visualization.pdf` |\n| 6 | Chesnevar et al. | `Health Psychology/Social & Collective/towards argument interchange format.pdf` |\n| 7 | Mercier & Sperber | `MercierSperberWhydohumansreason.pdf` |\n| 8 | Shneiderman | `collectiveintelligence-2012-shneiderman-v2.pdf` |\n| 9 | Predictive Liquid Democracy | **Not found on FERMI volume** |\n"}