{"path":"research/the-reuse-flywheel-prior.md","content":"# Should we expect the flywheel? The cross-scale prior, and what the corpus already saw\n\n**Date**: 2026-08-29 · **Type**: prior-examination + measurement + corpus archaeology\n**Question (founder)**: *\"even if we've not made the flywheel/asymptote appear concretely as of yet,\ncan we EXPECT it to appear if we design everything just right? … which also connects to whether there\nare fractal priors indicating we should have such expectations.\"*\n\n---\n\n## 0. The answer in one paragraph\n\n**Yes, expect it — the pattern recurs at seven independent scales and has never failed where its\ncondition held.** The condition is the same every time: **identity must be cheap to decide.** Where\nidentity is free (an atom of carbon, a codon, a type in a proof library, a DOI) the flywheel runs\nalmost frictionlessly. Where identity is *judged* rather than computed, the flywheel still runs — legal\nprecedent is the proof — but its rate is set by the quality of the identity work, and an entire\nprofession exists to do that work. **Deliberus is in the second bucket.** So the honest expectation is\nnot *the flywheel may not come*; it is **the flywheel will come at the rate identity work improves,\nand identity work is therefore the permanent core activity rather than plumbing to get past.**\n\nThree independent lines converge on that, which is why it is worth trusting: the corpus reasoned to it\nin **April 2026**, the cross-scale prior predicts it, and today's measurement shows it.\n\n---\n\n## 1. What the corpus already saw (the founder's prior instruction: this matters more than the build)\n\n**March 2026, the founder's own words** ([scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md),\nSession 7, verbatim):\n\n> *\"If the system would auto-decompose incoming new claims in a way that automatically prunes them\n> based on some criteria, like only finding or creating new premises that could one step removed be\n> allowing connections to other existing claims. Then perhaps it would be more computationally\n> feasible and helpful.\"*\n\nThe same doc then sharpens it into the move that matters, and the sharpening is the deepest thing in\nthis whole line:\n\n> **\"The better inversion — Auto-CONNECT, not auto-decompose.\"** … *\"The system doesn't GENERATE\n> premises — it DISCOVERS them in the existing graph. … This avoids hallucination entirely — you're\n> never generating premises, only finding real ones.\"*\n\nIt also names the three layers (extraction-time implicit premises · on-demand human-directed descent ·\nfull recursive auto) and gives four reasons layer three fails alone: infinite regress, plausible-not-\nreal premises, inversion of the sorry model, and most arguments not being deductive.\n\n**April 3 2026** ([soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md))\nstates the asymptote four months before this session, and states it *better*:\n\n> *\"once the graph is rich enough, recursion may get cheaper per marginal claim because more\n> decomposition steps terminate in already-known reusable substructure\"*\n>\n> the frontier changes from *\"how many layers can we generate?\"* to **\"how quickly can we recognize\n> that a new decomposition path has entered already-inhabited graph territory?\"**\n>\n> **\"The limiting factor becomes: identity confidence, cluster precision, reversible canonicalization,\n> reuse selection — not brute-force recursive generation depth.\"**\n\nPlus the design constraints that go with it: **never hard-merge by default**, prefer canonical clusters\nover node collapse, preserve lineage, **make splitting easier than merging** (false merges are worse\nthan missed merges), and the **Lean `@[simp]` flywheel** as the structural analogy — *\"the point of a\nmature library is not that every new proof must be generated from scratch.\"*\n\n**August 25 2026** ([kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md)) turns it\ninto a workload: eager recursive decomposition into a **draft tier, machine provenance, shown only on\ncontest**; adversarial-balance on fetched research; `worth_asking` as the machine's own stopping\nprice; four build pieces, none large.\n\n**Session 25** rules the identity half: anyone types anything, **careful auto-merge afterwards**,\nanyone can split — with originals never destroyed and merges visible where the claim is read.\n\n**So the design is not missing. It has been essentially complete since April, in the founder's own\nformulation, with the limiting factor correctly identified.** What this session contributed was a\nrediscovery of the April conclusion, which is the uncollected-idea pattern again.\n\n---\n\n## 2. The cross-scale prior, and the one condition\n\nSeven instances of *accumulated shared substructure makes new work cheaper*, each with its identity\nmechanism named — because the mechanism is what decides whether the pattern transfers.\n\n| Scale | The flywheel | How identity is settled | Cost of identity |\n|---|---|---|---|\n| Chemistry | ~100 elements compose everything; no new compound needs a new element | physics — two carbon atoms **are** identical | free |\n| Genetics | proteins are recombinations of existing domains; one code, ~20 amino acids | literal sequence match | free |\n| Language | small phoneme inventory; new words are mostly compounds of old morphemes | form | near-free |\n| **Mathematics (mathlib)** | a mature library lets new proofs terminate in retrieval (`exact?`, `@[simp]`) | **a lemma has a TYPE, and type equality is decidable** | **computed** |\n| Software | new programs are mostly calls into libraries | name + signature | free |\n| Science | you cite rather than re-derive | DOI | free |\n| **Legal precedent** | case law accumulates; new cases cite holdings rather than re-deriving them | **judged** — distinguishing, overruling, contested | **an entire profession** |\n\n**The pattern has a perfect record where identity is decidable and a slower, contested, still-real\nrecord where identity is judged.** It has no instance of outright failure — which is a genuinely\nencouraging prior, and it is the kind that survives the convictions-sweep transfer test rather than\nbeing a hopeful analogy.\n\n**Which bucket is Deliberus in?** The legal one, unambiguously. A claim's identity is judged, not\ncomputed; the sameness literature gives it a known ceiling (cross-document event coreference at 88%\nclean, 70–77% messy; sense granularity where human agreement is 56% fine-grained against 91% coarse);\nand the founder's own merge ruling treats identity as an ongoing, reversible, contestable activity\nrather than a preprocessing step.\n\n**Three consequences follow, and they are architectural rather than encouraging:**\n\n1. **Identity work is the permanent core activity, not plumbing.** In the one comparable case, it is\n   what the professionals spend their careers on. A system in this bucket should be *designed around*\n   sameness and distinction, which is what soft canonical clustering already proposed.\n2. **The rate is set by identity quality, not by corpus size.** Volume without matching does not bend\n   the curve — the April doc's own conclusion, and § 3 measures it.\n3. ~~**The mathlib case names the design direction** — make identity structural rather than textual;\n   nothing in the corpus has proposed structural matching keys yet.~~ **⚠ WRONG on both halves,\n   corrected 2026-08-29 after reading properly.** (a) It is the **scheme-relative hypothesis**, held\n   since 2026-04-19 in [claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md)\n   as explicitly *\"NOT a proposal, NOT a decision\"*, with **six named overreach risks** — one of which\n   is aimed precisely at the move I made: *\"The hypothesis is seductive: it has the shape of an elegant\n   reconciliation of all the traditions above, and elegant reconciliations across traditions are\n   exactly what Žižek warned about as premature synthesis.\"* The others bite too: cardinality blowup,\n   scheme authorship (Walton's catalogue is a standpoint, not a neutral basis), schemes blurring at the\n   edges, and render-time collapse reintroducing the violence the storage layer avoided. (b) Its\n   deterministic first rung **is already built**: `candidate_generation.structural_kinship_candidates`\n   (2026-08-20) signs each claim as the multiset of *(direction, edge type, scheme)* over its incident\n   scheme-bearing edges plus claim type, matched by weighted Jaccard, explainable by construction. It\n   was measured before shipping and it is the **weakest** of the three generators on hand precision\n   (≈3.5/6 against concept-route's ≈6/6), with a diagnosed failure mode — single-edge claims carry\n   information-free signatures and scored spurious 1.0 matches — fixed the same hour by a minimum\n   signature mass of four atoms. So the direction is not a hope, an insight, or unproposed: it is\n   shipped, measured, and known to be the *least* precise of what we have.\n\n---\n\n## 3. What is measured today, and the two corrections it forces\n\n**Correction one: the discovery mechanism exists and fires.** Earlier in this session I wrote that no\npath looks up an existing claim before minting. **Wrong.** `auto_connect`'s classifier offers\n`decomposes_into` among five outcomes with explicit two-direction handling (*could A serve as a PREMISE\nthat B logically depends on?*), and the graph holds **84 cross-source DECOMPOSES_INTO edges**. The real\ngap is narrower: it is called only from the two extraction paths, scoped to one source's claims against\n**other** sources, so claims the correction UX mints never pass through it and same-source reuse is\ninvisible.\n\n**Correction two, and it is the bigger one: the concept-layer reuse figure the corpus cites is\nwithin-source.** Measured 2026-08-29: 90 concepts, **79 reused by more than one claim (88%)** — the\nnumber `lowering-the-cost.md` and the frontier both cite as evidence that reuse runs through the\nconcept layer. But **only 4 concepts span more than one SOURCE (4%), and no concept in the graph spans\nmore than two.** *Justice* appears in the death-penalty, minimum-wage, assisted-dying and\nIsrael/Palestine material and is recorded at two sources. Concept labels are minted per extraction with\nsource-specific phrasing (*public good / welfare*, *sovereignty / effective control*), so **concept\nidentity has the same matching problem as claim identity** — and it went unnoticed because the\nwithin-source number looked healthy. **Identity is the binding constraint at every layer, not only the\nclaim layer.**\n\n**⚠ Every measurement in this section was taken on a linker that has been superseded in code and\nnever run.** The 2026-08-20 auto-connect upgrade shipped three additions — concept-route candidates\nwith a polysemy guard, structural-kinship candidates, and a write-time stance guard — described in\ntheir own doc as *live in code, exercised at next extraction (extraction currently paused)*. Verified\n2026-08-29: **zero edges in the graph carry `r.generator`, `r.via_concept` or `r.stance_caution`**, so\nevery connection in the corpus was made by cosine alone, on exactly the linker that doc was written to\nreplace. Measured before shipping, the new generators surfaced **58% of their candidate pool as pairs\ncosine never found** (76% across a wider sweep, peaking at 98% on one source). And\n[retrieval-instruments-beyond-cosine.md](retrieval-instruments-beyond-cosine.md) § 0 warns explicitly\nthat reading a low cosine as *not related* **is not a safe inference**, and that *\"the corpus may have\nless evidence of cross-domain sparseness than it believes\"* — a caveat the paragraph below leaned on\nthat evidence without carrying.\n\n**The flywheel's shape as measured on the OLD linker**: per-source cross-source connection yield does not rise with\ncorpus size. It tracks whether an **adversarial counterpart is already present** — death-penalty pair\n0.65 and 0.68 edges per claim, assisted-dying 0.60 and 0.94, minimum wage 0.44, against 0.02 for a lone\nWikipedia piece, 0.04 for Graeber, 0.02 for a solitary unified-theory article. **The flywheel is paired,\nnot cumulative.** (31 sources, so suggestive rather than settled; the pattern is not subtle.) This is\ndirect support for the standing ingest-by-debate-cluster rule and against the naive *it gets better as\nit grows* reading.\n\n---\n\n## 4. ~~A design tension nobody has named~~ — RETRACTED 2026-08-29, it was backwards\n\nThe first version of this section claimed the stranger test is **anti-reuse**, because decontextualising\nbakes source specifics into the claim text. **That inverts the finding it cites.** The production\nclaim-matching literature calls decontextualisation **normalization** and treats it as *the enabling\npreprocessing* — and the corpus recorded the consequence a month ago:\n[sameness-merge-and-split-across-fields.md § 5](sameness-merge-and-split-across-fields.md) §3d/§5,\nwhere the architecturally decisive result is that **a frontier model asked directly whether two claims\nmatch underperforms by nearly 10 CoNLL F1 points**, and the winning architecture is *collaborative* —\nthe model normalises and contextualises each claim, a dedicated matcher decides. Its own words:\n**\"the extraction pipeline's decontextualization pass is already that first half, so the recommended\narchitecture is half-built here by accident.\"**\n\nSo the stranger test is not in tension with reuse; **it is the first stage of the recommended matching\narchitecture.** The error is recorded rather than deleted because it is a clean instance of the class\nthis corpus keeps catching: a grep hit read as confirming the opposite of what its sentence said.\n\n## 4b. What the corpus holds that is better than anything above\n\nRead properly, three things in\n[sameness-merge-and-split-across-fields.md](sameness-merge-and-split-across-fields.md) answer the\nfounder's question more sharply than the cross-scale prior does.\n\n**The flywheel and the catastrophe are one mechanism.** The Semantic Web's `owl:sameAs` was strict\nidentity used sloppily at scale: 3–20% of identity links erroneous, and because **identity is\ntransitive, single bad links chained** — one closure falsely unified **177,000 names** for different\ncountries, cities and people. The doc's formulation is the one to keep: *\"identity infrastructure\npropagates errors exactly as efficiently as it propagates benefits.\"* **Every increment of reuse is an\nincrement of error-corridor.** You cannot buy one without the other, and the design consequence is\nalready recorded: no silent transitive chaining — cluster composition is itself a merge decision,\nnever an inference.\n\n**And a second thing the flywheel costs, from the philosophical side.** Gadamer: understanding is a\nfusion of horizons that happens *in* the encounter, so pre-clustering *\"does not just risk error; it\nprevents the event of understanding from occurring.\"* The flywheel's core move — recognising that a\ndescent has entered already-inhabited territory — is the same move that can foreclose the encounter\nthe product exists to host. Neither of these is an argument against the flywheel; both are the reason\nidentity work has to be good rather than merely fast.\n\n**The dial that no dictionary has.** Lexicography's answer to sense granularity was governance rather\nthan discovery: OntoNotes *\"iteratively merged senses until 90% inter-annotator agreement was\nreached\"* — granularity as the control variable, reliability as the target. **Deliberus has a strictly\nbetter dial: demand-driven granularity.** Two senses need separating exactly when the claims using\nthem **diverge downstream** — which is Wittgenstein's use-relativity and James's *a difference that\nmakes no difference is no difference* arriving at the same operational rule, and it is the reason the\ndownstream-divergence split alarm is *not only a safety net but the granularity-setting mechanism*.\nA dictionary must fix granularity a priori for all future uses; we never have to.\n\n**And the ceilings are numbers, not vibes**: cross-document event coreference 88.4% / 85.2% on curated\ncorpora against **~70–77% on messy ones**; fine-grained sense agreement **56% against 91%** coarse,\nwith **96 of 191 words** having no reliable middle granularity at all; **22.8%** of argument-to-key-point\nmatches human-ambiguous. The gray zone is a measured property of the judgment, not a weakness of the\nmachinery — so a borderline band is descriptively correct rather than a concession.\n\n**The synthesis that supersedes my two-bucket framing**: *\"Every field that hit the wall built process\ninstead of algorithm\"* — taxonomy's `sec.` convention, library authority records, graded identity\npredicates after the sameAs catastrophes, lexicography's agreement target with an underspecified escape\nhatch. **A process layer for contested judgment is what Deliberus already is**, which is a much\nstronger position than \"we are in the judged bucket.\" With its honest cost stated in the same place:\n**the risk transfers rather than disappearing — from \"no algorithm can do this\" to \"nobody shows up to\ndo the judging\"** — landing on the engagement gradient and on ratification decay.\n\n**One calibration I should have applied and did not**: those literatures are **entropy-class** (dirty\ndata, honest variation, nobody modelling the matcher). Our eventual setting is **strategy-class** —\nparties who want a claim merged or split for advantage. By the adversary-class rule that is a\n**rebuild, not a port**, so the maturity is real and its scope is narrower than it looks.\n\n---\n\n## 5. What would falsify the expectation\n\nStated so this section is not decoration. The prior predicts that with identity work improving, reuse\nrises. It is refuted if: cross-source concept span stays at ~2 after the next twenty **paired** sources;\nor the sameness detector, once switched on with a measured false-merge rate, fails to raise cross-source\n`DECOMPOSES_INTO` per claim; or the yield stays flat across a doubling of a *single* debate cluster,\nwhere every condition the prior names is met at once.\n\n**See also**: [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md)\n(the mechanism audit) · [lowering-the-cost.md](lowering-the-cost.md) (the cost curve and its formalism) ·\n[sameness-merge-and-split-across-fields.md](sameness-merge-and-split-across-fields.md) (the six-field\ngrounding and the known ceilings) · [assembly-theory-and-the-reuse-mechanism.md](assembly-theory-and-the-reuse-mechanism.md)\n(construction-with-reuse as physics) · [lean-deliberus-analogies.md](lean-deliberus-analogies.md) (the\n`@[simp]` flywheel) · [convictions-across-scales.md](convictions-across-scales.md) (the transfer test\nthis section uses).\n\n---\n\n## 6. The founder's mechanism, and the measurement that already exists for it\n\n*Founder, 2026-08-29: the current graph state matters less than whether the flywheel gets up to speed,\nand it likely will, because **\"people don't have that many different claims and concepts that they\nthrow around\"** — the operative reason being that **people need to be understood**, and that need\nseverely narrows the space of claims that circulate, and likewise the space of concept meanings that\ntake hold, even across subcultures using a word differently.*\n\n**This is not a hopeful analogy. It has been measured, on exactly this object, and the result is\nstronger than the claim needed.**\n\nBar-Haim et al. (IBM Research, ACL 2020, *From Arguments to Key Points*) took **28 controversial\ntopics and 6,515 crowd-contributed arguments** — about 233 per topic — and had a professional debater\nwrite **at most seven \"key points\" per side, in ten minutes per side, without seeing a single one of\nthe arguments.** Then they had eight annotators map every argument to those key points.\n\n**Written blind, seven sentences covered 72.5% of what 233 people independently said.** The coverage\ncurve is the finding:\n\n| key points | 1 | 2 | 3 | 4 | 5 | 6 | 7 |\n|---|---|---|---|---|---|---|---|\n| argument coverage | **41.3%** | 56.3% | 64.2% | 69.3% | 71.5% | 72.3% | **72.5%** |\n\n**One sentence covers 41% of a topic's arguments.** And the authors note the marginal contribution\ndecaying at six and seven, concluding that seven suffice.\n\n**The founder's mechanism is what explains the blind part.** The expert could write the key points\n*before seeing the data* only because the space is narrow and shared — which is what the need to be\nunderstood produces. An open space is not predictable in advance; this one was.\n\n### The limit, measured in the same table, and it is the important half\n\n**The curve saturates at ~72.5% and does not close.** Going from five key points to seven bought\n**one percentage point**. So roughly **a quarter of what people say does not compress**, and a further\n**22.8% of arguments were *ambiguous*** — trained annotators split, with no key point reaching the 60%\nthreshold (the same 22.8% the sameness doc already cites as the measured gray zone).\n\nTwo honest scope notes: these are crowd-contributed arguments under strict length limits on\ncontroversial topics, which is the register most likely to be conventional — a scholarly text, a legal\nbrief or an original position may compress far less. And the study measures *coverage by a human\nexpert's key points*, not machine matching, so it bounds the **compressibility of the space**, not our\nability to detect it.\n\n### What this settles, and it changes the payoff rather than the answer\n\n**The flywheel is real and it is front-loaded.** The head of any topic is narrow, heavy and\npredictable — a handful of debates per topic plausibly exhausts most of it, and the marginal value of\nmore volume drops off exactly as that curve does. **That is the answer to how much backfill is worth\nbuying: the coverage curve is the budget.**\n\n**And the residual quarter is where the crux lives.** Run 6's finding was that the crux of a real\ndebate is typically *unwritten* — which is by definition not in the circulating stock, so it sits in\nthe uncompressed tail rather than the reusable head. Compose the two results and the division of\nlabour falls out:\n\n> **Machine reuse makes the ~72% that recurs nearly free. Human attention goes to the quarter that\n> does not compress — which is exactly where the disagreement actually is.**\n\nSo the flywheel's payoff is not *less work*. It is **work concentrated where it matters**, which is\nthe same conclusion the engagement gradient reaches from the other side, and a considerably better\noutcome than a uniform speedup.\n\n**Source**: [Bar-Haim, Eden, Friedman, Kantor, Lahav & Slonim, *From Arguments to Key Points: Towards\nAutomatic Argument Summarization*, ACL 2020](https://arxiv.org/pdf/2005.01619) — coverage figures read\nfrom the paper directly (Figure 1, §3.2.2, §3.3).\n\n**Refinement (2026-08-30, from the chart–ledger comparison)**: the free-vs-judged split is not\n*geometry versus meaning* — **identity is cheap where nobody has a stake, and stakes make identity\njudged even in cartography** (OpenStreetMap border edit wars; Google Maps showing different national\nborders per country). The nautical chart's centuries of frictionless reuse ran on waters nobody\ndisputed. Sharpens why Deliberus sits in the judged bucket: its subject matter is staked by\nconstruction. [the-harness-conviction.md](the-harness-conviction.md) § 8.5b.\n"}