{"path":"research/consensus-path-forward.md","content":"# Consensus Consult: Path Forward for Deliberus\n\n**Date**: March 28, 2026\n**Models consulted**: Gemini 3 Pro (skeptic stance), Gemini 3 Pro (optimist stance), GPT-5.2 (neutral/practical stance)\n**Confidence**: High (8/10 across all models)\n\n---\n\n## The Consensus (All 3 Models Agree)\n\n### 1. Single-Player MVP First — Unanimous\n**\"Help me think about X\"** — personal cognitive prosthesis, not collaborative platform. The cold-start problem is fatal for any multiplayer-first approach. Every failed argumentation platform died here. Build a tool that delivers value to ONE person analyzing ONE text/decision.\n\nEntry point: **\"Paste URL/text → get an interactive, editable argument draft.\"** This competes with summaries by offering *structure* and competes with argument maps by offering *purpose*.\n\n### 2. No VC — Unanimous\nThe 50K-200K TAM makes venture capital structurally misaligned. VC demands engagement-maximizing algorithms that are toxic to truth-seeking. Self-fund → grants → philanthropic sustainability.\n\n### 3. Niche Beachhead — Unanimous\nTarget Rationalist/EA + policy researchers exclusively for 3+ years. They tolerate early UX friction if epistemic yield is high. Do NOT attempt mass market.\n\n### 4. Human-in-the-Loop as Primary Workflow — Unanimous\nTreat 50-60% NLI F1 on real-world text as a **physical constraint**, not a bug to fix. The LLM is an \"eager but flawed intern\" that *drafts*; the human corrects. The correction UX IS the product.\n\n### 5. Frame Humbly — Unanimous\nOutputs are \"community-mapped argument structures\" / \"mapped perspectives\" — NOT objective truth. Epistemic humility as architectural principle. Private-by-default, share-by-link.\n\n### 6. No Probabilistic/Reputation Features in MVP — Unanimous\nBridging arguments, two-axis voting, calibration scoring, prediction markets = Phase 2. Ship the extraction + correction loop first. Storage choices are reversible; user workflow mistakes are not.\n\n---\n\n## The Key Disagreement: Solo vs. Accountability\n\n### Gemini (Both Stances): Stay Solo\n\"Premature team formation is your biggest organizational risk. Adding co-founders will dilute architectural coherence and accelerate runway burn.\"\n\n### GPT-5.2: Solo Is the Highest-Risk Path\n\"The biggest execution risk is not coding — it's the founder pattern: 15 years of research without shipping. That is an accountability/system design problem, not a technical one.\"\n\n**GPT-5.2's recommendation**: Add ONE collaborator or structured external accountability:\n- **Option A**: One builder (frontend/product) with weekly demo cadence\n- **Option B**: One product/accountability operator\n- **Option C (solo fallback)**: \"Synthetic accountability\" — weekly recorded demos, pre-committed scope, small advisory group that will actually call slippage. But humans work better than rituals.\n\n**The synthesis**: GPT-5.2 is right that the 15-year pattern is the real risk. But Gemini is right that premature team formation dilutes vision (the Nils Janse lesson). The resolution: **solo for the first 6-8 weeks (clickable prototype), then evaluate whether external accountability is needed based on actual shipping velocity.**\n\n---\n\n## Concrete Timeline (GPT-5.2)\n\nStarting April 2026, with one experienced dev + heavy AI leverage:\n\n| Phase | Duration | Deliverable |\n|-------|----------|-------------|\n| **0. Research wrap** | Now → April | Complete simmering. Run claim extraction experiment. Decision on ontology approach |\n| **1. Thin slice** | 6-8 weeks | Clickable prototype: URL/text ingestion → basic extraction → simple tree/graph view |\n| **2. Correction UX** | 10-14 weeks | Editing UI that makes output trustworthy: accept/reject/merge/split claims, fix edges, confidence display |\n| **3. MVP v1** | 16-24 weeks | One killer workflow (\"analyze this paper for my decision\"), onboarding, exports, share links, private alpha |\n\n**~4-6 months to an MVP people can actually use**, if scope stays narrow.\n\n---\n\n## What to Build FIRST (GPT-5.2 — Specific Order)\n\n1. **\"Paste URL/text → argument draft\" frontend** (even ugly) that returns something visible and editable\n2. **Minimal extraction pipeline**: Claim nodes + Support/Attack edges + source spans + confidence scores\n3. **Editing/correction UI** (this IS the product): accept/reject/merge/split claims; fix edge directions; mark normative vs descriptive\n4. Only THEN choose durable storage (graph DB vs relational) based on actual usage patterns\n\n**Do NOT start with**: graph storage, probabilistic semantics, reputation systems, or the full ontology. Start with the user workflow.\n\n---\n\n## Financing Strategy\n\n### Phase 0-1: Self-Funded (Now → Prototype)\nNo external dependencies during critical design/build phase.\n\n### Phase 2: Build-in-Public Lite\n- One weekly devlog (30-60 min cap)\n- One monthly deeper essay (repurpose existing research docs)\n- **Patreon ONLY after working prototype loop** (URL → draft → edit → share). Otherwise you monetize the *idea*, reinforcing \"simmering\"\n- If Patreon early, set explicit deliverables (\"ship X by date Y\"), not \"support the vision\"\n\n### Phase 3: Grants (Mid-2026, After Prototype)\n- **Vinnova**: Frame as democracy tech / AI for public sector / misinformation resilience. Stronger with consortium (municipality, university partner)\n- **EU Horizon**: Join as technical work package lead, not consortium coordinator. Keep product execution ring-fenced\n- **Knight Foundation / EA Infrastructure Fund**: Pitch with working demo\n- **Ideell förening**: For legitimacy + grants, but consider dual structure (association for mission + lean entity for execution)\n\n### Phase 4: Sustainability\nGrant-funded public good + optional API/enterprise licensing for institutional users. NOT advertising, NOT engagement optimization.\n\n---\n\n## Organizational Structure Recommendation\n\n**Now**: Solo developer, no formal entity\n**After prototype**: Evaluate PBC (Bluesky model) vs ideell förening (Swedish nonprofit) vs dual structure\n**The Mastodon precedent**: Eugen Rochko → solo dev → Patreon → gGmbH (German nonprofit LLC). This is the closest analog to Fredrik's situation.\n\n---\n\n## Top 3 Risks\n\n1. **The 15-Year Pattern** (Existential) — Endless research without shipping. Mitigation: ruthless scope control, weekly demos, external accountability (collaborator or advisory group), hard deadline for first public artifact.\n\n2. **LLM Extraction Quality** (Serious) — 50-60% F1 on real political text. Mitigation: human-in-the-loop as primary workflow; start with narrow domains (papers, policy docs) where extraction is reliable; surface uncertainty visibly.\n\n3. **The Platform Graveyard** (Existential) — 20+ predecessors all stagnated. Mitigation: single-player utility dissolves the cold-start; don't build multiplayer until 1,000 solo users proven; the \"digital gardens\" asymmetric sharing model avoids the engagement trap.\n\n---\n\n## ARG-tech / AI4Deliberation: Don't Merge, Collaborate\n\nKeep product ownership independent. Partner for:\n- **Evaluation**: user studies, measurement, academic credibility\n- **Pilot contexts**: universities, policy labs\n- **Datasets/benchmarks**: argument mining evaluation\n- **Funding leverage**: stronger grant applications with academic partners\n\n---\n\n## The One-Line Summary\n\n**Build a personal thinking tool that produces shareable argument maps, starting with \"paste URL → editable draft,\" and let the collaborative platform emerge from proven single-player utility.**\n\n---\n\n## Sources\n\n- Gemini 3 Pro Preview (skeptic + optimist stances, March 28, 2026)\n- GPT-5.2 (neutral/practical stance, March 28, 2026)\n- Cross-referenced with: adoption-problem.md, steelmanned-critiques.md, single-player-utility.md, conceptual-threads.md\n"}