{"path":"research/scheme-bounded-decomposition-and-evidence-as-subgraph.md","content":"# Scheme-Bounded Decomposition and Evidence-as-Subgraph\n\n> **Status update (Aug 13, 2026).** The meta-analysis thinking here was revisited when the founder asked whether evidence weighing had been discussed before — it had, and this is the document. What has shipped since: the methodology ladder now feeds QBAF base weight (`METHODOLOGY_BASE_WEIGHT`, meta-analysis at the top), and `study_count`, `heterogeneity`, `effect_size`, CI bounds and per-claim `source_span` are all extracted and reach the synthesis prompt. **What has not**: the meta-analysis-to-studies decomposition sketched in § Evidence-as-Subgraph, and the `[CQ] Effect size heterogeneity?` idea — heterogeneity is recorded and displayed but weighs nothing. A twelve-study meta-analysis at I²=0.73 currently scores identically to a forty-study one at I²=0.05. Continuation, including the same structure arriving as a long-form video and five specific things the ladder gets wrong: **`long-form-sources-and-meta-analysis-weighing.md`**.\n\n**Date**: March 29, 2026 (Session 7)\n**Status**: Exploratory architectural reasoning — not a decision. See CLAUDE.md §Critical Design Philosophy.\n**Type**: Deep design exploration arising from founder questioning\n\n---\n\n## What evidence currently *is*, and the finding that resolves the open question (2026-08-16)\n\nThe founder asked what evidence means or should mean, and the honest answer is that the word currently names **three different objects** in the system, which is the source of the confusion.\n\n**Properties on a claim** — `evidence_type`, `methodology_tier`, `effect_size`, confidence-interval bounds, `sample_size`, `study_count`, `heterogeneity`, `source_span`. All extracted, all shipped.\n**A claim linked by a support edge** — what `POST /claims/{id}/add-evidence` creates.\n**A decomposed subgraph from a document** — the vision sketched below, not built.\n\n**The first two are the same object at different granularities, and nobody had noticed.** `heterogeneity: 0.73` sits on a claim as a number nobody can argue with. Stated as a claim it reads *\"the twelve studies disagree substantially with each other\"* — which is attackable, decomposable, and can carry evidence of its own. **Every evidence property is a claim that has not been minted yet**, which is structurally the same finding as the reported-speech blind spot: real content trapped inside something that is not a node.\n\n**That answers the founder's own open question** — *\"What actually constitutes un-decomposable, more atomic evidence is an open question to me, scientifically, rationally, philosophically\"* — in a usable direction. **There is no atomic evidence, only claims whose decomposition nobody has demanded yet.** That is the self-similar principle applied to evidence, which he had already said it should be.\n\n**And it dissolves a problem filed as something else.** The status note above records that `heterogeneity` weighs nothing, so a twelve-study meta-analysis at I²=0.73 scores identically to a forty-study one at I²=0.05, and treats this as a missing weighting formula. It is not: it is a property that should be a claim. Mint it with an attack edge on the parent and the shipped machinery holds it — **but the expectation as first written here was REFUTED when run** (stress-suite wave two, 2026-08-22): an attack edge whose critical questions are unanswered contributes exactly zero by design, so minting buys *contestability*, not weight. The heterogeneity claim now presses on the meta-analysis only once someone scores the edge — the badge-boundary two-step (attached ≠ scored) applies to minted evidence-properties like everything else. Still no new formula and no new weights; but also no automatic effect. The same holds for sample size, which becomes a contestable claim about generalisability, and methodology tier, which becomes a claim about what that design can establish.\n\n**Consequence for the evidence-decomposition pass**: its first and cheapest step is not deeper document parsing but **turning the already-extracted evidence properties into claims**. Document decomposition is the second and expensive step. The founder's own constraint governs when to stop — *\"maximum granularity that is helpful to the rest of humanity\"*, which is demand-bounded rather than maximal, and matches the rule reached independently in [decomposition-axes.md](decomposition-axes.md): decompose along an axis only when someone contests along it.\n\n## The Founding Question\n\n**Fredrik (verbatim, Session 7):**\n\n> \"What if we need even more atomic claims? Could the system include auto-finding premises for each new claim submitted, so that all claims would be the conclusion of a few logical premises? That's the way I (or my left hemisphere at least) always imagined it. But perhaps I need to let go of that rigorous analytical hope. Perhaps that's a pipedream.\"\n\nAnd later, pushing back on the distinction between \"premises\" and \"strength modifiers\":\n\n> \"Are they truly not premises? Are they not decomposable into subquestions about the expert's experience, education etc?\"\n\nAnd the deeper conviction:\n\n> \"I don't want any copouts, I want people and AI agents surfing around in Deliberus to see exactly what the best minds on our planet think in detail and what the best evidence shows - evidence that also is decomposed into its constituent parts/arguments. This is similar to the conviction I have that value premises are often decomposable and evidence is often too. What actually constitutes un-decomposable, more atomic evidence is an open question to me, scientifically, rationally, philosophically.\"\n\n> \"In accordance with the progressive complexification, uploading a meta-analysis as a PDF to Deliberus as support of a claim should be allowed. But then that PDF should be decomposed as far as possible - maximum granularity that is helpful to the rest of humanity.\"\n\n---\n\n## The Three Layers of Auto-Decomposition\n\n### Layer 1: Implicit Premise Detection (extraction-time)\n\nSurfacing what's ALREADY IMPLIED in the source text but not stated. The MArgE research calls this \"implicit premise recovery.\"\n\nSource text: \"UBI would reduce poverty because it provides a basic income floor.\"\nCurrently extracted: two claims + supports relationship.\nWith implicit premise detection: also surfaces \"a basic income floor lifts people above the poverty line\" — the unstated assumption connecting the two explicit claims.\n\nThis stays within the source material. It makes the author's actual reasoning more visible without inventing reasoning they didn't have. The pipeline could add a pass for this.\n\n### Layer 2: On-Demand Decomposition (interaction-time, human-directed)\n\nWhen a user clicks ◇ on a value premise, the system generates suggested sub-premises. \"You said 'every person deserves a subsistence floor.' Here are possible foundations — which resonates?\"\n\n- Human dignity is inherent (deontological)\n- Subsistence floors produce better societal outcomes (consequentialist)\n- Social contract requires mutual provision (contractarian)\n\nThe user picks, edits, or writes their own. The LLM suggests; the human commits. This IS the sorry model working as designed.\n\n### Layer 3: Full Recursive Auto-Decomposition\n\nEvery claim auto-generates premises, recursively. Problems:\n\n1. **Infinite regress**: Every premise needs premises. A 74-claim extraction becomes 300-1200 after one round, 5000+ after two.\n2. **LLM premises are plausible, not real**: Generated premises may not be the author's actual reasoning. Mercier & Sperber's insight — reasoning is post-hoc rationalization — means auto-generated premise chains find the rider's story, not the elephant's motivation.\n3. **Inverts the sorry model**: The system does the cognitive work that should be the user's (P14 interpassivity).\n4. **Most human arguments aren't deductive**: Arguments from analogy, expert opinion, consequences don't reduce to syllogisms. Forcing them into premise-conclusion form distorts them.\n\n**However**: Fredrik's pruning idea transforms Layer 3 from pipedream to feasible...\n\n---\n\n## The Pruning Insight: Graph-Aware Auto-Decomposition\n\n**Fredrik's synthesis (verbatim, Session 7):**\n\n> \"If the system would auto-decompose incoming new claims in a way that automatically prunes them based on some criteria, like only finding or creating new premises that could one step removed be allowing connections to other existing claims. Then perhaps it would be more computationally feasible and helpful.\"\n\nThis is a different operation than Layer 3. Instead of recursively decomposing EVERYTHING, only auto-decompose in directions that CREATE CONNECTIONS to existing claims.\n\n**Benefits**:\n- Natural pruning: most generated premises are discarded (no combinatorial explosion)\n- Connection discovery: kept premises bridge otherwise disconnected argument structures\n- `@[simp]` flywheel: as the graph grows, more premises find connections, so more are kept. System gets BETTER over time\n- On an empty graph, almost nothing kept (good — no filler). On a 10,000-claim graph, many connections found (powerful)\n\n**The better inversion — Auto-CONNECT, not auto-decompose**:\n\nInstead of \"generate premises, then check if they connect,\" flip it:\n\n\"For each new claim, search the existing graph for claims that could serve as its premises, evidence, or qualifications.\"\n\nThis is P13 (composing = retrieval) applied at the backend. The system doesn't GENERATE premises — it DISCOVERS them in the existing graph. \"UBI reduces poverty\" enters the graph, and the system finds: \"There's already a claim 'direct cash transfers reduce poverty indicators' from the economics extraction — could that be a supporting premise?\"\n\nWhen no existing claims connect, the sorry marker appears: \"This claim has no supporting structure in the graph yet.\"\n\nThis avoids hallucination entirely — you're never generating premises, only finding real ones. The auto-linking already does a primitive version (SIMILAR_TO at 0.80 cosine). The extension: also search for potential SUPPORTS, ATTACKS, and DECOMPOSES_INTO relationships. An LLM evaluates: \"Does existing claim X logically support new claim Y?\" — classification (high accuracy), not generation (hallucination-prone).\n\n---\n\n## Critical Questions ARE Premises\n\nThe initial framing treated Walton's critical questions as \"strength modifiers\" — meta-questions about an argument rather than part of the argument itself. **This was wrong.** Fredrik's pushback was correct.\n\n\"Is the expert credible?\" isn't floating above the argument. It decomposes:\n\n```\n\"Dr. Smith says UBI works\" (argument from expert opinion)\n  └── CQ: \"Is Smith credible in economics?\"\n        └── \"Smith has published 47 papers in labor economics\" (empirical, verifiable)\n        └── \"Smith holds a chair at MIT\" (empirical, verifiable)\n        └── \"Smith consulted for three governments\" (empirical, verifiable)\n        └── CQ of THOSE: \"Are peer-reviewed papers reliable indicators of expertise?\"\n              └── \"Peer review has a replication crisis in social sciences\" (empirical)\n              └── \"Despite flaws, peer review remains the best available filter\" (normative)\n              └── ...\n```\n\nIn ASPIC+ formalization, critical questions map to **undercutting attacks** — arguments that weaken the inference. Each undercutter is itself an argument with premises. The argument scheme tells you WHICH premises to look for — not \"what are some premises?\" (open-ended, hallucination-prone) but \"has the expert's domain relevance been established?\" (specific, answerable, structured).\n\n**An argument scheme is a premise-generation template.** It reveals not just what was said (explicit premises) but what MUST be true for the argument to work (the critical questions). This is scheme-bounded decomposition: bounded by the scheme's finite CQ set, not recursive to infinity.\n\n---\n\n## What Is Truly Atomic?\n\n**Almost nothing.** Every candidate for \"atomic evidence\" decomposes further:\n\n| Candidate | Decomposes into... |\n|---|---|\n| \"I observed X\" | Senses functioning + measurement calibrated + observation conditions adequate |\n| \"Study found d=0.4\" | Sample size + methodology + statistical test + confounders + replication status |\n| \"2+2=4\" | Axioms of arithmetic (which are CHOSEN — Peano, ZFC, constructive make different choices) |\n| \"I feel pain\" | The Cartesian cogito. Genuinely hard to decompose. But even: \"am I interpreting sensation correctly?\" |\n| A meta-analysis | Individual studies → each study's methodology → each methodology's assumptions |\n\n**The graph has no natural bottom. But it doesn't NEED one.** It needs a **current bottom** — sorry markers — that is always invitable to go deeper. The depth at any point reflects how much the community has cared to decompose that particular thread.\n\nThis extends the \"no copout axioms\" principle universally: not just to value premises, but to empirical evidence, expert testimony, causal claims — everything. The label is always \"currently undecomposed,\" never \"bedrock.\"\n\nSome threads will be decomposed to near-bedrock levels (heavily scrutinized medical evidence with individual study methodologies examined). Others will have sorry markers at the first level (claims nobody has yet examined). That asymmetry IS the information — it shows where epistemic effort has been invested.\n\n---\n\n## The 96 Schemes Clustered by Decomposition Type\n\nEach cluster has a characteristic pattern — the scheme tells you WHERE to dig and WHAT KIND of premises to look for.\n\n### Source-Based: \"Who says it, and why should we believe them?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Expert Opinion | WHO → credentials → domain match → reliability → peer agreement → evidence basis |\n| Position to Know | WHO → access → honesty → actual assertion |\n| Witness Testimony | WHO → presence → perception conditions → memory → bias |\n| Popular Opinion | HOW MANY → evidence for popularity → information quality → independent merit |\n| Popular Practice | HOW COMMON → descriptive accuracy → normative relevance |\n\n**Evidence tracking**: Source claims decompose into credential chains. \"Smith is an expert\" → publications, institutional affiliation, peer recognition. Each is empirical and verifiable. The inference pattern: credibility of source transfers (defeasibly) to credibility of assertion.\n\n### Causal: \"What causes what, and how do we know?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Cause to Effect | MECHANISM → causal generalization strength → presence of cause → interfering factors |\n| Effect to Cause | OBSERVATION → alternative explanations → base rates → mechanism plausibility |\n| Correlation to Cause | CO-OCCURRENCE → confounders → temporal ordering → dose-response → mechanism |\n| Sign | INDICATOR → correlation strength → alternative explanations |\n\n**Evidence tracking**: Causal claims decompose into mechanistic chains + statistical evidence. \"Smoking causes cancer\" → epidemiological studies (sample sizes, effect sizes, confounders) + biological mechanism (tar → DNA damage → mutation → uncontrolled growth). Each step has its own evidence quality. This is where meta-analysis decomposition lives — a meta-analysis IS a causal evidence claim.\n\n### Consequence-Based: \"What happens if we do X?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Positive Consequences | ACTION → causal link to outcome → outcome desirability → alternatives |\n| Negative Consequences | ACTION → causal link to harm → harm severity → mitigation possibility → alternatives |\n| Practical Reasoning | GOAL → action as means → efficiency → feasibility → side effects |\n| Danger/Fear Appeal | THREAT → probability → severity → proposed response → alternatives |\n| Waste/Sunk Costs | PAST INVESTMENT → goal achievability → forward cost-benefit (ignoring sunk) |\n\n**Evidence tracking**: Consequence claims chain two sub-arguments: (1) a causal claim (\"action A leads to outcome O\") — which is itself a causal scheme with all its CQs, and (2) an evaluative claim (\"outcome O is good/bad\") — which is a value premise inviting decomposition.\n\n### Analogy/Example: \"What's it like?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Analogy | SOURCE CASE → target case → similarity mapping → relevant differences → counter-analogies |\n| Example | INSTANCE → generalization → representativeness → counter-examples |\n| Precedent | PAST CASE → ruling/outcome → similarity → changed circumstances |\n\n**Evidence tracking**: Analogy claims decompose into similarity assessments. \"Finland's UBI experiment applies to Sweden\" → population similarity, economic similarity, policy design similarity, cultural similarity. Each is empirical and assessable.\n\n### Rule/Classification: \"What category does this fall in?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Verbal Classification | OBJECT → category definition → fit → edge cases |\n| Established Rule | RULE → applicability → exceptions → authority of rule-maker |\n| Exception to Rule | CASE → rule → exception criteria → qualification |\n| Precedent | PAST DECISION → current case → relevant similarity → authority |\n\n**Evidence tracking**: Classification claims decompose into definitional disputes — where semantic disambiguation becomes critical. \"This counts as hate speech\" → definition of hate speech → this utterance's properties → fit assessment. The definition IS a contested claim (P18 concept tracking).\n\n### Character/Ad Hominem: \"Who are you to say?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Ethotic Ad Hominem | PERSON → character flaw → relevance to claim → bearing on truth |\n| Circumstantial Ad Hominem | PERSON → circumstances → inconsistency → bearing on truth |\n| Bias | PERSON → interest → influence mechanism → independence of claim |\n| Inconsistent Commitment | PERSON → past commitment → current assertion → contradiction → explanation |\n\n**Evidence tracking**: Decompose into empirical claims about persons + inferential claims about relevance. The CQ \"Does the person's flaw bear on the TRUTH of the claim?\" is itself a critical premise — ad hominem can be valid (bias is relevant) or fallacious (character is irrelevant to truth).\n\n### Slippery Slope: \"Where does this lead?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Causal Slope | A0 → A1 → A2 → ... → An: each link's causal plausibility + breakpoints |\n| Precedent Slope | DECISION → future interpretations → institutional dynamics → endpoint |\n| Gradualism | SMALL STEP → direction → accumulation → threshold |\n| Full Slope | Combined causal + precedent + psychological momentum |\n\n**Evidence tracking**: Each link is a separate causal claim with its own evidence. The slope is only as strong as its weakest link. The system adds real value by making each link individually examinable.\n\n### Abductive: \"What explains this?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Evidence to Hypothesis | EVIDENCE → explanatory power → competing hypotheses → search thoroughness |\n| Abductive Reasoning | OBSERVATION → best explanation → alternatives eliminated → mechanism |\n\n**Evidence tracking**: Decomposes into evidence quality + explanatory comparison + search thoroughness. This is the scientific method formalized.\n\n### Composition/Division: \"Parts vs whole\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Composition | PARTS have property P → WHOLE has property P? → valid transfer conditions |\n| Division | WHOLE has property P → PARTS have property P? → valid transfer conditions |\n\n### Values-Based: \"What matters?\"\n\n| Scheme | Decomposition pattern |\n|---|---|\n| Argument from Values | VALUE → priority relative to competing values → scope → application |\n\n**Evidence tracking**: This is where \"no copout axioms\" bites hardest. \"Liberty matters more than equality\" decomposes into WHY — and the WHY decomposes further. Value premises are the deepest sorry markers, but they're NOT un-decomposable.\n\n---\n\n## Deductive vs Inductive: Not Different Systems, Different Schemes\n\nFredrik's concern (verbatim):\n\n> \"Any claim could be supported by different arguments, but each argument needed its premises in order to be a logical argument. But of course there's also inductive arguments, right? And the whole thing about what constitutes a good bayesian inductive argument and so on.\"\n\nThe argument scheme absorbs this distinction. Different schemes have different premise-conclusion relationships:\n\n| Type | Example | Inference relationship | \"Premises\" means... |\n|---|---|---|---|\n| **Deductive** | Syllogism | Necessarily follows | Logical necessities |\n| **Inductive** | \"N studies show X\" | Probably follows | Evidence + quality indicators (sample size, methodology, replication) |\n| **Abductive** | \"Best explanation for X is Y\" | Explains best | Evidence + explanatory comparison + alternatives eliminated |\n| **From analogy** | \"Worked in Finland → works in Sweden\" | Structurally similar | Source case + similarity mapping + disanalogies |\n| **From authority** | \"Expert says X\" | Credibly asserted | Credentials + domain match + peer agreement |\n| **From consequences** | \"Policy P causes outcome O\" | Causally connected | Causal chain + evidence for each link |\n\nYou don't need separate systems for deductive vs inductive. The scheme TYPE determines:\n- What the CQs are (what questions to ask)\n- What \"premises\" means for THIS argument (what kind of support it needs)\n- What \"VALID\" means for THIS argument bundle (the badge from Sketch 2)\n- What sorry markers appear (which CQs are unanswered)\n\n### Bayesian Integration\n\nFor inductive arguments, Bayesian reasoning provides the formal backbone:\n\n- **Prior probability** of the conclusion (before this evidence)\n- **Likelihood ratio**: how much more likely is this evidence if the conclusion is true vs false?\n- **Posterior probability**: updated belief after considering this evidence\n\nThe CQs for inductive arguments map to likelihood ratio assessments: \"How strong is the causal generalization?\" → affects how much this evidence should update our belief. \"Are there confounders?\" → if yes, the likelihood ratio drops. \"Has this replicated?\" → affects our confidence in the likelihood ratio itself.\n\nThis connects to Thread 5 (Probabilistic Argument Graph): QBAFs can represent both deductive and inductive strength. Deductive arguments with verified premises → strength approaching 1.0. Inductive arguments from one case study → strength ~0.3, increasing with replications.\n\n---\n\n## The Sketch Architecture Made Concrete\n\n**Sketch 1** (Contention diagram): A \"Contention\" at the center, surrounded by argument bundles, with three scoring axes: **truthiness, relevance, controversiality**. Some nodes with dashed borders (proto-sorry markers for incomplete claims).\n\n**Sketch 2** (Argument bundles): A \"conclusion\" at top, three VALID badges connecting to \"speed argument,\" \"dream argument,\" \"cobalt argument\" — each with premise cards. Tab navigation at bottom.\n\nThese sketches map to the scheme-bounded architecture:\n\n- Each **argument bundle** (speed, dream, cobalt) = one argument scheme instance\n- Each bundle's **premises** = the scheme's explicit premises + CQ-revealed implicit premises\n- The **VALID badge** = all CQs for this scheme addressed (green), some addressed (blue/amber), none (orange), one failed (red)\n- The **three axes** = scoring at the bundle level:\n  - **Truthiness**: How well are the CQs answered? (evidence quality)\n  - **Relevance**: Does this bundle actually bear on the contention? (inference strength)\n  - **Controversiality**: How contested are the CQ answers? (community disagreement)\n- **Sorry markers** (dashed borders in Sketch 1) = CQs not yet addressed\n\n---\n\n## Evidence-as-Subgraph: The PDF Decomposition Vision\n\nWhen someone uploads a meta-analysis as evidence for a claim, the system should decompose it — not treat it as a black-box attachment:\n\n```\nClaim: \"UBI reduces poverty\"\n  ├── [SCHEME: argument from evidence]\n  ├── Evidence: meta-analysis (uploaded PDF, extracted by pipeline)\n  │     ├── [EXTRACTED] \"12 RCTs with combined N=41,000 show poverty reduction of 14-27%\"\n  │     │     ├── [CQ] Sample representativeness? ← sorry\n  │     │     ├── [CQ] Effect size heterogeneity? → \"I²=73%, high heterogeneity\" (extracted)\n  │     │     └── [CQ] Publication bias? → \"No funnel plot asymmetry\" (extracted)\n  │     ├── [EXTRACTED] \"Inclusion criteria: randomized, >6 months, income-based outcome\"\n  │     │     └── [CQ] Were criteria appropriate? ← sorry\n  │     ├── [SORRY] Individual study methodologies — not yet decomposed\n  │     │     (clicking this: runs extraction pipeline on individual study PDFs)\n  │     └── [SORRY] Statistical model choice — not yet examined\n  ├── Evidence: Finland KELA experiment (extracted from earlier URL submission)\n  │     ├── [EXTRACTED] \"2,000 participants, 2-year duration\"\n  │     ├── [EXTRACTED] \"Employment effects null; well-being effects positive\"\n  │     └── [CQ, auto-connected] \"Does a Finnish sample generalize?\"\n  │           → links to existing claim about Nordic exceptionalism from another extraction\n  └── [SORRY] Competing evidence — not yet examined\n```\n\nEach layer uses the SAME extraction pipeline. Evidence documents become subgraphs. Sorry markers show where decomposition hasn't happened yet. Auto-connect finds bridges. The community fills in gaps that matter most.\n\n**This is the extraction pipeline's natural extension**: it already extracts claims from text. A meta-analysis IS text. Run the pipeline on it, link the extracted claims to the higher-level claim they support. Recursive, but bounded by the actual content of the evidence document.\n\n---\n\n## How Close to the Dream: 85%\n\nThe fully decomposed graph is achievable through:\n\n1. **Scheme-bounded decomposition** — each scheme has a finite CQ set (not infinite recursion)\n2. **Auto-connect** — discovers existing answers in the graph (not generating new claims)\n3. **Evidence-as-subgraph** — uploaded documents are extracted by the pipeline, becoming linked subgraphs\n4. **Sorry markers** — show the current bottom, always inviting deeper decomposition\n5. **Community contribution** — humans fill in the gaps that matter most\n6. **`@[simp]` flywheel** — the system gets better at auto-connecting as the graph grows\n\nThe 15% gap:\n- LLM-generated decompositions will sometimes be wrong (human validation needed)\n- Not all documents decompose cleanly (some are narratives, not argument structures)\n- The infinite depth means the system is NEVER \"done\" — but this is a feature. It means \"currently undecomposed\" is always the honest label\n- Some evidence is genuinely difficult to decompose further (the Cartesian cogito, mathematical axiom choices, direct phenomenal experience)\n\n**The dream is an emergent property of the system operating over time, not a feature of the extraction pipeline.** It grows like Lean's Mathlib — claim by claim, community contribution by contribution, with the system getting better at auto-connecting as density increases. The depth of any thread reflects actual epistemic effort invested. That asymmetry IS the information.\n\n---\n\n## Extraction Cost Analysis (March 30, 2026)\n\nGemini 3 Flash pricing (verified March 30, 2026): $0.50/1M input, $3.00/1M output. Cached input (85% reduction): ~$0.075/1M.\n\nSource: Wikipedia UBI article (~5,000 words ≈ ~7,000 tokens). Produces ~14 arguments → ~74 claims → ~26 relationships.\n\n### Current Pipeline (6 passes, 32 LLM calls)\n\n| Pass | Calls | Input cost | Output cost | Subtotal |\n|------|-------|-----------|------------|----------|\n| 1 Scout | 1 | $0.004 | $0.006 | $0.010 |\n| 2a Decompose | 14 | $0.017 | $0.034 | $0.051 |\n| 2b Decontext | 14 | $0.017 | $0.042 | $0.059 |\n| 3a Relationships | 1 | $0.003 | $0.006 | $0.009 |\n| 3b Concepts | 1 | $0.003 | $0.005 | $0.007 |\n| 4 Self-eval | 1 | $0.003 | $0.003 | $0.006 |\n| 5 Embed + link | 0 LLM | — | — | $0.000 |\n| **Total** | **32** | **$0.047** | **$0.095** | **$0.142** |\n\n### New Pipeline (8 passes, ~403 LLM calls)\n\n| Pass | Calls | Input cost | Output cost | Subtotal |\n|------|-------|-----------|------------|----------|\n| 1 Scout | 1 | $0.004 | $0.006 | $0.010 |\n| 2a Decompose | 14 | $0.017 | $0.034 | $0.051 |\n| 2b Decontext | 14 | $0.017 | $0.042 | $0.059 |\n| 3a Rels+Schemes | 1 | $0.003 | $0.012 | $0.015 |\n| 3b Concepts | 1 | $0.003 | $0.005 | $0.007 |\n| CQ Parameterize | 1 | $0.002 | $0.015 | $0.017 |\n| 4 Self-eval v2 | 1 | $0.005 | $0.006 | $0.011 |\n| 5 Embed + link | 0 LLM | — | — | $0.000 |\n| 6 Auto-connect | 370 | $0.093 | $0.111 | $0.204 |\n| **Total** | **~403** | **$0.143** | **$0.230** | **$0.373** |\n\n### Cost driver analysis\n\n- **Output tokens dominate**: $0.230 of $0.373 (62%) is output cost. Gemini charges 6x more for output than input.\n- **Auto-connect is the majority**: $0.204 of $0.373 (55%). 370 individual LLM classification calls.\n- **Batching auto-connect** (20 pairs per call → 19 calls): reduces to $0.171, total $0.340. Saves ~$0.03.\n- **Flywheel recursion** (one round, ~60 more calls): adds ~$0.04, total ~$0.41.\n- **Passes 1-4 barely change**: $0.170 vs $0.142 — scheme detection + CQ parameterization adds only $0.03.\n\n### At scale\n\n| Extractions | Current | New | New + flywheel |\n|-------------|---------|-----|----------------|\n| 1 | $0.14 | $0.37 | $0.41 |\n| 10 | $1.40 | $3.70 | $4.10 |\n| 100 | $14 | $37 | $41 |\n| 1,000 | $140 | $370 | $410 |\n\nVery affordable for development and early use. Auto-connect costs grow linearly with graph size (more existing claims = more candidates), but the top-K pre-filter caps LLM calls at `new_claims × K` regardless of graph density.\n\nSources: [Gemini 3 Flash Pricing](https://ai.google.dev/gemini-api/docs/pricing), [pricepertoken.com](https://pricepertoken.com/pricing-page/model/google-gemini-3-flash-preview)\n\n---\n\n## Cross-References\n\n- [vision.md §No Copout Axioms](../vision.md) — value premises always decomposable; extended here to ALL claim types\n- [vision.md §Core Dialectic](../vision.md) — analysis↔attunement; the left hemisphere's dream of total decomposition in service of the right hemisphere's understanding\n- [walton-argument-schemes.md](walton-argument-schemes.md) — the 96 schemes, CQs, LLM classification accuracy, VALID badge concept\n- [lean-deliberus-analogies.md](lean-deliberus-analogies.md) — sorry model, `@[simp]` flywheel, trustless collaboration\n- [extraction-pipeline-design.md](extraction-pipeline-design.md) — three-pass pipeline, no chunking, RLM for large texts\n- [ux-principles.md §P4](../ux-principles.md) — sorry model as UX principle\n- [ux-principles.md §P13](../ux-principles.md) — composing = retrieval\n- [ux-principles.md §P14](../ux-principles.md) — interpassivity warning (LLMs scaffold, never replace)\n- [ux-principles.md §P15](../ux-principles.md) — argument bundles as tab-navigable threads (Sketch 2)\n- [conceptual-threads.md §Thread 5](../conceptual-threads.md) — probabilistic argument graph, QBAFs\n- [inductive-reasoning-research.md](../inductive-reasoning-research.md) — BIRD framework, Bayesian integration\n- [sketches.md](../sketches.md) — the original hand-drawn contention diagram and argument bundle sketches\n- [session7-deep-review-and-strategic-assessment.md](session7-deep-review-and-strategic-assessment.md) — the session context for this exploration\n"}