{"path":"research/unbuilt-features-research-mar31.md","content":"# Unbuilt Features Research: UX + Technical Frontier\n\n**Date**: March 31, 2026 (Session 7)\n**Status**: Research complete. Prioritized roadmap established.\n**Type**: Web search + existing doc synthesis across 8 feature areas\n\n---\n\n## Part 1: High-Impact UX Features\n\n### Two-Axis Voting (P2)\n\n**State of the art**: LessWrong shipped agree/disagree + karma in June 2022. Three years of production data validates the approach.\n\n**What worked** (Ben Pace, LessWrong): Comments that would have been \"downvoted to hell\" now sit on positive karma but are \"disagreed to hell\" — obvious improvement in signal clarity. Users feel comfortable upvoting good arguments they disagree with.\n\n**What didn't work**: Visual confusion for newcomers (~5% found it confusing). \"What does agree mean?\" for non-propositional comments (jokes, process). Most votes go the same direction (>95% of the time).\n\n**Key insight**: Comments with high karma + high disagreement are the most epistemically valuable — arguments that people find genuinely update-worthy even while disagreeing. **This IS the bridging argument signal.** Surface it as a discovery feature.\n\n**Deliberus advantage**: Claims are already atomic with explicit types (empirical/normative/definitional/value_premise), so \"what does agree mean?\" is largely solved. Each claim has clear propositional content.\n\n**Implementation recommendation**:\n- Two vote buttons per claim: well-argued / poorly-argued (left), agree / disagree (right)\n- Agreement displayed as percentage with Bayesian smoothing (virtual prior: 5 up + 5 down → starts at 50%)\n- Agreement does NOT affect visibility or ranking (LessWrong's validated decision)\n- Surface \"high-quality + high-disagreement\" quadrant as first-class feature\n\n**Sources**: LessWrong announcement (lesswrong.com/posts/HALKHS4pMbfghxsjD), EA Forum two-factor proposal (forum.effectivealtruism.org/posts/e7rWnAFGjWyPeQvwT)\n\n---\n\n### Feed Algorithm for Deliberation\n\n**State of the art**: Active research area in 2025-2026. Three paradigms beyond engagement:\n\n1. **Justified Representation (JR)** — Revel et al., ICML 2025. From social choice theory: ensure every significant opinion group has at least one representative in top-K results. Proven to increase viewpoint diversity while remaining compatible with quality optimization. **Most rigorous formal treatment.**\n\n2. **Bayesian Surprise Ranking** — Hasan & Bunescu, RecSys 2023. Rank by KL divergence between user's prior and posterior beliefs. Surface content that maximally informs rather than confirms.\n\n3. **Bluesky Algorithmic Marketplace** — AT Protocol feed generators. Users choose between algorithms. Open-source starter kit. The pattern: let users switch between \"most contested\", \"bridging\", \"newest\", \"most surprising to me.\"\n\n**Key finding** (Knight-Georgetown Institute, Mar 2025): Employee performance metrics at platforms are tied to engagement. Alternative feeds require alternative business models.\n\n**Recommendation for Deliberus**: Start with JR constraint (formal grounding) + Bayesian surprise (personalized). Never optimize for time-on-platform. Use stated preferences (\"was this valuable?\") not behavioral proxies (clicks, dwell time). Let users choose feed algorithm (Bluesky pattern).\n\n**Sources**: Revel et al. ICML 2025 (proceedings.mlr.press/v267/revel25a), KGI report (kgi.georgetown.edu), Bluesky feed-generator (github.com/bluesky-social/feed-generator)\n\n---\n\n### Bridging Argument Detection\n\n**State of the art**: The formal problem is solved by Blair et al. (\"The Structure of Bridging\", Harvard/Toronto):\n\n- **Problem**: Polis's group-aware consensus depends on arbitrary clustering (k-means on PCA). Different k or random seeds → different \"bridging\" statements.\n- **Solution**: Pairwise disagreement metric — partition-agnostic, axiomatically justified, handles sparse vote matrices. Measures how much a statement reduces expected disagreement between pairs of users.\n\n**PAKT** (COMMA 2024): Perspectivized Argumentation Knowledge Graph. Neo4j-based, annotates arguments with stakeholder perspectives. Architecturally closest to Deliberus.\n\n**The vagueness-bridging tension** (Puig Larrauri, ICIP 2025): Maximizing common ground incentivizes vague platitudes. Deliberus's four-type classification is a structural countermeasure — empirical claims are harder to make vague. \"Bridging empirical claims\" could resolve this tension.\n\n**Deliberus's unique contribution**: Extending bridging from STATEMENTS to REASONING STRUCTURE is genuinely novel. No existing system computes bridging metrics over argument graphs where reasoning quality is a separate dimension from agreement. This remains Deliberus's most differentiated theoretical contribution.\n\n**Recommendation**: Implement pairwise disagreement (Blair et al.) as primary bridging metric. Use Google Jigsaw classifiers as bootstrap before vote data accumulates. Apply bridging per claim type (empirical bridging claims are most valuable).\n\n**Sources**: Blair et al. (cs.toronto.edu/~nisarg/papers/bridging.pdf), PAKT (arxiv.org/abs/2404.10570), Puig Larrauri (icip.cat)\n\n---\n\n### Onboarding\n\n**Key insight**: The extraction pipeline IS the onboarding. User pastes URL → watches extraction → browses structured claims. Time to first value: ~30 seconds. No tutorials needed.\n\n**Progressive disclosure of engagement** (validated by Kialo, Polis, AskVerdict patterns):\n- Level 0: Read extracted claims, browse argument map (no account)\n- Level 1: Vote on claims — agree/disagree + well-argued (requires account)\n- Level 2: Add evidence, answer CQs (now built — correction UX)\n- Level 3: Submit new texts for extraction\n- Level 4: Edit the argument graph directly\n\n**Anti-pattern** (Kialo): 23+ help articles means the system isn't self-explanatory. If you need a tooltip, the design isn't done.\n\n**Duolingo principle**: Micro-sessions — let users evaluate 5-10 claims in 2 minutes. Show surprising insight: \"You agreed with 3 claims from the opposing side.\" The reward is the insight, not points.\n\n---\n\n## Part 2: High-Impact Technical Features\n\n### Semantic Zoom for Argument Graphs\n\n**Best library**: Cytoscape.js with compound nodes + cytoscape-expand-collapse extension. Supports three zoom levels:\n1. Topic clusters (compound nodes, labels + claim counts)\n2. Argument bundles (expand topic → see arguments with support/attack edges)\n3. Individual claims (full metadata, type badges, evidence links)\n\n**Closest published work**: Conch (Chen et al., IEEE VIS 2025) — hierarchical debate visualization with LLM-extracted \"clash points\" at multiple levels. Also: Society Library's \"Junto\" debate mapping system.\n\n**Recommendation**: Cytoscape.js for detail views, Sigma.js for >10K node overview. Use Louvain community detection on SIMILAR_TO edges for auto-discovering topic clusters.\n\n**Complexity**: Medium-high (2-3 weeks).\n\n---\n\n### Claim Deduplication and Merge UX\n\n**State of the art**: CLEF CheckThat! 2025 Task 2 (claim normalization). Best approach: retrieval via embeddings (Deliberus already has this) → LLM confirmation (\"are these the same assertion?\").\n\n**Wikipedia merge UX pattern**: Banner on potential duplicates → user confirms merge/keep separate → merged claim gets MERGED_INTO edge, relationships transfer.\n\n**Key principle**: Never auto-merge. The distinction between \"same conclusion, different evidence\" and \"true duplicate\" requires human judgment and is core to the epistemology.\n\n**Complexity**: Medium (1-2 weeks). Deliberus already has embeddings + Gemini Flash.\n\n---\n\n### Epistemic Calibration Scoring\n\n**Gold standard**: Metaculus's log scoring rule (strictly proper). Score = ln(p) where p is probability assigned to actual outcome.\n\n**Implementation for Deliberus**: Extend claims with resolution field. Store user predictions with timestamps. Compute log score + peer score on resolution. Display calibration curves.\n\n**Key design decision**: What are users predicting? Best fit: \"How likely is this claim to be true given current evidence?\" (epistemic assessment, not social prediction).\n\n**Complexity**: Medium (1-2 weeks for core). Needs claim resolution mechanism (governance decision).\n\n---\n\n### Real-Time Collaboration\n\n**Best approach**: Yjs CRDTs with Y.Map for nodes/edges.\n\n**Phased**:\n- Phase 1 (1 week): Shared cursor awareness only — see where others are looking\n- Phase 2 (2 weeks): Collaborative voting/annotation (append-only, simple for CRDTs)\n- Phase 3 (2-3 weeks): Full collaborative graph editing (requires FalkorDB sync layer)\n\n**Recommendation**: Defer Phase 3 until multi-user. Phases 1-2 provide 80% of collaborative value at 30% complexity.\n\n---\n\n## Part 3: Academic Alignment\n\n**Francesca Toni's group at Imperial College London** is the closest academic alignment to Deliberus:\n- ArgLLM-App (Feb 2026): Working web system combining LLMs + computational argumentation\n- \"Argumentative Human-AI Decision-Making\" (AAMAS 2026): Deliberus's thesis in academic form — \"AI should reason WITH humans, not FOR them; decisions must be contestable and revisable\"\n- Potyka (QEM inventor) is adjacent\n\n**Deliberus goes significantly beyond ArgLLM-App**: extraction pipeline + scheme detection + CQ generation + QEM + correction UX + Temporal workflows. ArgLLM-App handles binary decisions; Deliberus targets multi-perspective deliberation.\n\n**Other aligned groups**: UKP Lab at TU Darmstadt (ClaimFlow), Serena Villata at Inria (argument mining), Stanford Digital Economy Lab (AI deliberation at scale).\n\n---\n\n## Priority Roadmap\n\n| Priority | Feature | Effort | Dependencies | Why now |\n|---|---|---|---|---|\n| **1** | Two-axis voting | 1 week | Google OAuth (done) | Generates the data all other features need |\n| **2** | Claim deduplication | 1-2 weeks | Embeddings (done) | Improves extraction quality immediately |\n| **3** | Bridging detection | 1-2 weeks | Two-axis voting data | The novel contribution — prove it works |\n| **4** | Semantic zoom | 2-3 weeks | Cytoscape.js integration | Transforms the navigation UX |\n| **5** | Feed algorithm | 1-2 weeks | Voting data + bridging | Makes the graph browsable at scale |\n| **6** | Calibration scoring | 1-2 weeks | Claim resolution governance | Needs users predicting |\n| **7** | Real-time collab (Phase 1) | 1 week | Yjs integration | Shared awareness, low complexity |\n\n**The critical path**: Two-axis voting → bridging detection → feed algorithm. This sequence generates the data each subsequent feature needs, and proves the combinatorial bet (the novel value emerges from their interaction).\n\n---\n\n## Cross-References\n\n- [ux-principles.md](../ux-principles.md) — P2 (two-axis voting), P9 (epistemic notifications), P17 (feed as dashboard)\n- [conceptual-threads.md](../conceptual-threads.md) — Thread 4 (dedup), Thread 5 (probabilistic), Thread 6 (visualization), Thread 7 (gaming/reputation)\n- [bridging-arguments.md](bridging-arguments.md) — The original bridging argument research\n- [feed-algorithm-design.md](feed-algorithm-design.md) — Original feed algorithm research\n- [epistemic-gamification.md](epistemic-gamification.md) — Calibration scoring design\n- [crdt-collaborative-graphs.md](crdt-collaborative-graphs.md) — CRDT architecture research\n- [graph-visualization.md](graph-visualization.md) — Semantic zoom research\n- [qbaf-gradual-semantics-research.md](qbaf-gradual-semantics-research.md) — QEM as foundation for all of the above\n"}