{"path":"research/combinatorial-mvp-reasoning.md","content":"# The Combinatorial MVP: Why the Thin Slice Enters a Dead Category\n\n**Date**: March 28, 2026\n**Origin**: Conversation between Fredrik and Claude, applying Fredrik's \"combinatorial exploration\" principle to the MVP strategy. Emerged as a critique of the multi-model consensus recommendation.\n\n---\n\n## The Problem with the Consensus MVP\n\nThe multi-model consensus ([consensus-path-forward.md](consensus-path-forward.md)) recommended: \"paste URL → editable argument draft\" as the thinnest viable slice. GPT-5.2's second opinion ([second-opinion-mvp-strategy.md](second-opinion-mvp-strategy.md)) reinforced this with strong arguments about execution proof and workflow truth.\n\nBut both evaluated components **in isolation**. Fredrik's combinatorial exploration principle (documented in global CLAUDE.md) states:\n\n> \"What emergent capabilities arise from combining A+B that neither provides alone? Don't evaluate components in isolation — evaluate the interaction space. The most differentiated products come from novel combinations of individually-unremarkable parts.\"\n\n## The Dead-Predecessor Table\n\nEach component of the thin-slice MVP, evaluated alone, has a dead predecessor:\n\n| Component alone | Already exists as | Outcome |\n|----------------|-------------------|---------|\n| Claim extraction | Claimify, ClaimsMCP | Academic tool, niche |\n| Argument map editor | Kialo, Argdown, Rationale | Platform graveyard (20+) |\n| Correction UX | Every annotation tool | Commodity |\n| Single-player thinking tool | Obsidian, Roam, Miro | Crowded market |\n| Opinion clustering | Polis | Exists, 10M+ users, but no argument structure |\n| Epistemic scoring | Metaculus | Exists, active, but no argumentation |\n\n**None of these are what makes Deliberus novel.** A \"paste URL → editable draft\" MVP enters the argument-mapper category, which has 20+ dead predecessors, and offers no visible reason why \"this time is different.\"\n\n## The Steve Jobs Principle\n\nSteve Jobs didn't ship a phone, then add a browser, then add music. He showed the *combination* because **the combination was the insight**. A phone that just makes calls is just a phone. Phone + web + music + touch = a new category.\n\nSimilarly:\n- Argument extraction alone = Claimify (exists)\n- Opinion clustering alone = Polis (exists)\n- Argument extraction + opinion clustering + two-axis voting = **bridging arguments** (exists nowhere)\n\nThe combination creates something qualitatively different — the ability to see which reasoning structures are compelling across opinion groups, even when those groups disagree on conclusions. This is invisible to both Kialo AND Polis. It's the emergent property that justifies building a new platform.\n\n## What Would the MVP Prove? — A Sharper Analysis\n\nThe standard \"Lean Startup\" question is \"what's the riskiest assumption?\" The consensus identified the correction UX as the product risk. But there's a prior question: **does anyone care?**\n\n| What MVP supposedly proves | Already known? | Cheapest test | What ACTUALLY needs proving |\n|---------------------------|---------------|---------------|---------------------------|\n| LLMs extract arguments | Yes (Claimify, F1 scores) | Script + 10 texts (days) | Settled — not the risk |\n| Users tolerate correction | Analogous evidence (Wikipedia, Metaculus) | Wizard of Oz (n=20) | Whether correction is rewarding as SELF-SERVICE (WoZ lies about this) |\n| Maps help decisions | Decades of evidence (van Gelder, DeliData) | Share manual maps | Settled — not the risk |\n| Retention | Genuine unknown | Manual service | Whether people return when THEY do the work |\n| \"Aha moment\" exists | Unknown | Figma / demo video | **Time-to-value in real environment** |\n| **The combination creates new insight** | **UNKNOWN — this is the real question** | **Only testable by building the combination** | **Whether extraction + classification + two-axis voting reveals something invisible to Kialo AND Polis** |\n\nThe last row is what the thin-slice MVP misses. A \"paste URL → draft\" proves the correction loop works but does NOT prove the combinatorial thesis. And the combinatorial thesis is the entire reason to build Deliberus rather than contributing to Kialo or Polis.\n\n## The Combinatorial MVP\n\nThe smallest artifact that demonstrates the combinatorial novelty:\n\n1. **Extract claims** from a real debate text (same as thin slice)\n2. **Classify factual vs normative** (the fact/value boundary — one LLM call per claim)\n3. **Show two-axis evaluation**: agree/disagree AND well-argued/poorly-argued (even as simple UI toggles)\n4. **Demonstrate what emerges**: show how the interaction of these axes reveals bridging arguments, value-premise divergence, and discourse-level forks — things invisible to any existing tool\n\nThis is slightly MORE than the thin slice (adds classification + two-axis UI) but dramatically MORE differentiated. It shows the gap in the market, not just another entry into a dead category.\n\n## The Tension: Shipping vs Mattering\n\nThe consensus optimized for **\"highest probability of shipping\"** — ruthless scope reduction, weekly demos, execution forcing.\n\nThe combinatorial principle optimizes for **\"highest probability of mattering\"** — demonstrate the emergent property that justifies existence.\n\nThese are in tension. The resolution: **the combinatorial demo is only marginally larger than the thin slice** (one classification step + one extra voting axis). At 8-15x AI-augmented development velocity, the difference is days, not weeks. The cost of including the combinatorial insight is low; the cost of omitting it is entering a dead category.\n\n## The Revised Build Sequence\n\n1. **Claim extraction experiment** (days) — test extraction + fact/value classification on real texts\n2. **Two-axis evaluation UI** (days) — even crude, show agree/disagree AND well-argued/poorly-argued\n3. **Combinatorial demo** (1-2 weeks total) — show what emerges from the interaction that no existing tool reveals\n4. **Correction UX** (the product) — make the above editable, persistent, shareable\n5. **Storage, infrastructure, scaling** — only after the above is proven\n\nThis is the same thin-slice → correction-UX → MVP sequence the consensus recommended, but with the classification and two-axis evaluation included from step 1 — a small addition that transforms the artifact from \"yet another argument mapper\" to \"something genuinely new.\"\n\n## Cross-References\n\n- [consensus-path-forward.md](consensus-path-forward.md) — the original multi-model consensus this critiques\n- [second-opinion-mvp-strategy.md](second-opinion-mvp-strategy.md) — GPT-5.2's emphasis on execution proof\n- [assumption-ranking.md](assumption-ranking.md) — the weakest assumptions (#1-5) are exactly what the combinatorial demo tests\n- [bridging-arguments.md](bridging-arguments.md) — the novel theoretical contribution at stake\n- [steelmanned-critiques.md](steelmanned-critiques.md) — the platform graveyard critique that motivates differentiation\n- [conceptual-threads.md](../conceptual-threads.md) §Why Now? — the combinatorial capability table\n- vision.md §The Combinatorial Bet — the strategic principle\n"}