{"path":"research/lowering-the-cost.md","content":"# Lowering the Cost: What Reduces the Price of Understanding Another Mind\n\n*Aug 13, 2026. Founder extension to the founding conviction, verbatim: \"Curiosity is the willingness to pay the cost but the cost can also be substantially and successively lowered by Deliberus and innovating with Deliberus or around it in the ecosystem and culture surrounding its uses or potential uses.\"*\n\n*That sentence adds a term the conviction did not have. `Opacity Is a Cost, Not a Mystery` says the price is real and payable. The curiosity research added who is willing to pay it. This adds that **the price itself is a variable**, reducible by the product, and reducible further by things that are not the product at all. This document decomposes the cost so each term can be attacked separately, which is the project's own method turned on its own premise.*\n\n---\n\n## 1. \"The cost\" is one word hiding eight costs\n\nThe conviction has been carried as a single quantity. It is not one. A reader trying to understand why another mind holds its position pays at least these, and each has a different reduction curve, a different owner, and a different ceiling.\n\n| Cost term | What the reader pays | Reducible by | Status today |\n|---|---|---|---|\n| **Retrieval** | Finding the reasoning at all — most positions are never externalized anywhere | Elicitation, riding existing channels | Deliberus extracts only from texts that already exist |\n| **Extraction** | Turning prose into inspectable structure | LLM extraction | **Largely solved** — this is the friction that killed every predecessor |\n| **Traversal** | Walking the descent: N claims, edges, tree held in working memory | Diffing rather than reading; progressive disclosure | Partly solved, and still tree-shaped rather than diff-shaped |\n| **Translation** | Their vocabulary against yours; contested terms | Concept layer, pairwise term mapping | Concept layer live; pairwise mapping absent |\n| **Frame** | Being able to *inhabit* the reading in which their position is compelling | The worldview filter | Founding vision, still unbuilt |\n| **Trust** | Believing the map is faithful and not adversarially shaped | Ratification, instrument independence, third-party audit | Instruments exist; ratification absent |\n| **Attunement** | Why it *matters* to them — the felt, non-propositional half | Presence, voice, narrative | Named as a pole; no mechanism |\n| **Social** | The cost of being *seen* understanding the other side | Norms, formats, status, institutions | **Untouched, and probably the largest** |\n\nTwo immediate consequences. Extraction — the term Deliberus solved — may not be the biggest term. And the largest term is one no feature can reach, which is exactly why the founder's sentence names the ecosystem and culture rather than the roadmap.\n\n## 2. The synthesis idea, and why it matters more now than when it arrived\n\nFrom the external research thread, its own nomination for the highest-leverage feature:\n\n> \"The highest-leverage feature is **inspectable synthesis**. Every generated synthesis should expose source claims, supporting evidence, omitted arguments, uncertainty, alternative syntheses, confidence, revision history. Think of it as **Git for reasoning**.\"\n\nPaired with two other lines from the same thread: \"a synthesis should never be allowed to become an irreversible compression,\" and, in its comparison table, conventional AI *hides* the synthesis process where Deliberus should *expose* it.\n\nThe critical read at the time filed this as sound but unremarkable. It is more than that, because of what arrived afterwards.\n\n**The collision.** A synthesis is the single largest available reduction in traversal cost. \"Here is their position in six sentences\" is cheaper than any descent, by an order of magnitude, for every reader forever. It is also precisely where the Google DeepMind study found steering: their facilitator *produces the group statement*, and allocations moved by up to 5.5 percentage points while participants preferred the experience and consensus did not improve.\n\nSo the two facts are: **synthesis is the biggest cost lever, and synthesis is where the steering lives.** *Corrected 2026-08-13, and the correction matters:* this paragraph originally said Deliberus resolves the tension by declining to synthesize at all. **That was false.** Searching the codebase found `POST /query` running a 1,044-line truth-graph read layer that synthesizes prose from the graph at three depths — shipped, in production, since April. What Deliberus declines is not synthesis but *inspectable* synthesis: the answer reports whether it is graph-backed or model-backed and never which claims it used, and a depth budget drops claims before the model sees them. So the choice was never whether to take the cost reduction. It is already taken, and the open question is whether it is taken **honestly** — which is a much better position to be in, and a much cheaper fix. *Update, same day:* the honest version shipped. Answers now cite the claims they rest on and report what the depth budget dropped, so the biggest cost lever is being pulled with a receipt attached. Build record: [synthesis-build-plan.md](synthesis-build-plan.md).\n\nInspectable synthesis is the third option, and it is the one the tension was waiting for. Synthesize — take the cost reduction — but ship the synthesis with its own construction attached: which claims it drew on, **what it left out**, which alternative syntheses were available, and a revision history. Then run disagreement preservation against it. The omissions list is the load-bearing part, because a flattening synthesis is defined by what it dropped, and a synthesis that must enumerate its own drops is one whose flattening is measurable rather than invisible.\n\nThis reframes the differentiating prediction already in TODO. The current form asks whether *declining* synthesis avoids steering. The better experiment is a three-arm one: no synthesis, opaque synthesis, inspectable synthesis — scored on disagreement preservation and on whether readers can detect the flattening. If inspectable synthesis holds preservation while cutting traversal cost, that is a publishable result and a product decision at once. If it does not, the declining stance is vindicated *with evidence* rather than by caution.\n\n## 3. Attacking each cost term\n\nMarked **[new]** where this document is proposing it, **[registered]** where it already exists somewhere in the corpus.\n\n### Retrieval — the unwritten-reasoning problem\n\nThe deepest cost, and the least discussed. Deliberus can only extract reasoning that someone already wrote down. Most reasons are never externalized, so for most positions the cost is not high, it is infinite.\n\n- **[new] The elicitation interview, and the reason it is a double win.** LLMs are far better at drawing out reasons than any form is. Ten minutes of conversational follow-up yields more structure than a two-hour template. And the question-asking literature (Huang et al. 2017) found that asking follow-up questions makes the asker better *liked*, through the answerer's own enjoyment of self-disclosure. So the cheapest available way to obtain someone's reasoning is the same act that builds the relationship — the extraction instrument and the attunement instrument are one instrument. That is not a metaphor, it is the same behaviour appearing in two literatures.\n- **[new] Ride channels people already use.** Nobody will visit a site to map their argument. People already write comments, argue in group threads, record voice notes, sit for interviews, file planning objections, review pull requests. An extraction layer that accretes from what was written anyway is a different class of cost reduction from a better editor.\n\n### Traversal — the working-memory bound\n\n- **[new] Diff, not tree — the position fingerprint.** You do not need to traverse someone's whole descent. You need the part where you diverge. With premises shared across sources and your own votes on file, the graph can compute: *you already agree with 14 of these 17 premises; here are the 3 that matter.* Traversal cost collapses from \"read their essay\" to \"read two nodes.\" This is double-crux operationalized at corpus scale, and the fan-out hinge is what identifies which divergences are load-bearing rather than incidental. Highest-leverage new item in this document.\n- **[registered]** Progressive disclosure and the inversion principle already handle the shape of the descent. What is missing is ordering it by the *reader's own* priors rather than by the graph's structure.\n\n### Translation\n\n- **[new] Pairwise term mapping.** Not \"here are five senses of freedom\" but \"when they say freedom they mean roughly what you would call non-interference; when you say it you mean capability.\" General disambiguation is a reference work; pairwise translation is a cost reduction, and the concept layer already holds the senses needed to compute it.\n\n### Frame\n\n- **[registered]** The worldview filter, still the founding vision and still unbuilt.\n- **[new, with a flag] Steelman in the reader's own idiom** — their argument restated from premises you already hold. Genuinely cheaper than a neutral statement. Flagged because this is where convergence-illusion lives: restating a view in my vocabulary is a flattening operation by construction, so it must ship with disagreement preservation attached or not at all.\n\n### Trust\n\n- **[new] Ratification as a cheap trust primitive.** A steelman is trustworthy when the person steelmanned signs it. \"They agree this is a fair statement of their view\" costs one click and buys more credibility than any amount of methodological transparency. The thread's source paper lists human ratification as an architectural component; the version worth building is narrower and stronger — a ratified statement of a *specific person's* position, portable and citable.\n- **[registered] Instrument independence**: the honesty checkers currently run on the same model class they audit. The ecosystem answer is third parties running the instruments against the public graph themselves, which the agent-readable surface already makes possible. Outsourced auditing is a cost reduction paid by someone else.\n\n### Attunement\n\n- **[new] Put the voice on the node.** Forty seconds of the actual person stating their actual premise, attached to the claim. Not analysis — presence. It does something no amount of structure does, it is trivially cheap now that voice is already an input path, and it is the only concrete mechanism this corpus has ever proposed for the pole it says is co-equal with analysis.\n\n### Social — the largest term, and the one the product cannot touch\n\n- **[new] Change what counts as winning.** If the culture's scoring function rewards *\"I can state your view so well that you would sign it,\"* then understanding gets paid for because it has become profitable rather than because it is virtuous. This is the deepest available cost reduction and it is not a feature.\n- **[new] The ideological Turing test as a public format, built on the graph.** Claims plus votes are already enough to run it: state the other side's position, have them score whether it passes. It is cheap to build, culturally potent, produces a public artifact, and converts comprehension into a visible competence. Kahan's ambassadors finding is the distribution mechanism — the culturally diverse and curious are listened to when they return to their own communities.\n- **[new] A journalism norm.** \"We asked them to ratify our statement of their position.\" One sentence, immediately adoptable, raises the status of whoever does it, and needs no institutional buy-in — the same shape as the pre-submission self-review entry point identified for peer review.\n- **[new] Teaching as cost reduction, via PACE.** If trait curiosity largely reduces to appraising one's own ability to understand as high, then *showing someone they can complete a descent* directly raises the appraisal, which directly raises curiosity, which lowers the effective price. Education is not adjacent to this project's mechanism; under the PACE finding it *is* the mechanism, applied to the reader instead of the argument.\n\n## 4. Three through-lines\n\n### The corpus effect is the actual cost engine\n\nUnderstanding the hundredth holder of a position costs far less than the first, because their premises are already mapped and their divergences are already computable. **The marginal cost of understanding a new mind approaches the cost of locating it in an already-mapped space.**\n\nThat is what \"successively lowered\" means mechanically, and it reclassifies several roadmap items. Concept unification, cross-source auto-connect, mint-time dedup and shared-premise detection have been filed as features. They are not features. They are the cost curve. Anything that increases premise reuse is a direct multiplier on every future reader's cost, which makes it strictly higher-leverage than anything that improves a single descent.\n\n### The largest cost is social, and the product cannot reach it\n\nSix of the eight terms yield to engineering. The social term does not, and it plausibly dominates in exactly the contested settings the project exists for. In a polarized context, comprehension of the other side reads as defection, and no interface change alters that. This is the strongest argument for the ecosystem-and-culture half of the founder's sentence being load-bearing rather than decorative — and for the formats above, which work on the scoring function rather than the tooling.\n\n### The self-interested pitch, which this project has never had\n\nEvery adoption argument in the corpus so far asks the user to want something good: better reasoning, mutual understanding, epistemic infrastructure. Here is one that asks them to want something for themselves.\n\n**Put your reasoning in the graph so that people stop misrepresenting you.**\n\nIf being in the graph means you get steelmanned accurately, get to ratify the statement of your own view, and can point at a citable artifact when someone strawmans you, then externalizing your reasoning is in your own interest whether or not you care about collective epistemics at all. It also creates a gradient the whole system feeds on: reasoning inside the graph is cheap to understand, reasoning outside it is expensive, so the people who most dislike being misunderstood have the strongest reason to move inside.\n\n## 5. The cost curve has a formalism, and it is not a metaphor\n\nThe claim in §4 — that the marginal cost of understanding the *n*th holder of a position falls as coverage accumulates — is a unit-cost-against-cumulative-volume claim. That is the single best-documented functional form in the economics of cost, so the right move is to borrow the formalism rather than gesture at it.\n\n**Wright's law** (Wright 1936, on airframes; generalized by Arrow 1962 and BCG 1970 into the experience curve). Unit cost falls by a constant percentage for every doubling of cumulative production:\n\n> C(x) = C₁ · x^(log₂ b)\n\nwhere b is the progress ratio and 1−b the learning rate. Across industries the learning rate typically lands between 10% and 25%. Nagy, Farmer et al. (*PLOS ONE* 2013) hindcast six competing laws against 62 technologies and found Wright's the best forecaster, with Moore's close behind — and Sahal's observation that the two become indistinguishable when cumulative production itself grows exponentially.\n\n**Romer's nonrivalry** (1990; Nobel 2018) explains *why* such a curve should exist here, and fits Deliberus better than the manufacturing analogy does. Ideas are non-rival: \"a nonrival input has a high cost of producing the first unit and a zero cost of producing subsequent units.\" A mapped premise is precisely that object — expensive to extract and decontextualize once, free for every subsequent reader. Romer's replication argument transfers almost word for word: firms do not need to reinvent the idea for a computer each time a new factory is built, and **readers do not need to re-derive a premise each time a new mind turns out to hold it.**\n\n> **This section is the answer to an evaluation frame the founder rejected on 2026-08-16.** Asked about the standing gap that \"every instrument measures structure rather than whether anyone reasoned better\", he replied that this is an individualist productivity measure and the project is a commons: *a library is not justified by whether one reader reads better*, and its value far exceeds what is individually productivity-increasing. **The paragraph below had already argued that formally**, and the uplift framing contradicted it — a case of the corpus being right and the framing drifting away from it. Two things follow. The **fit measures are commons-shaped**: coverage of the contested space, concept reuse (87%, while claim-level reuse is near zero), reachability, the falling marginal cost of a descent, and whether the residue map moves. And this is **not an escape from falsifiability** — each of those can fail, and claim-level reuse already reads badly. The honest limit of the library analogy is worth keeping too: a library rests on a claim nobody disputes, that books hold knowledge worth reaching. Deliberus is arguing that a *new kind of artifact* is worth building, so the collective frame settles the **unit of value** without exempting the artifact from having to be good.\n\nRomer supplies one more thing the cost argument needed. Growth in his model is tied to the *total stock* of ideas rather than to ideas per person, because the stock is not divided among users. That is the theoretical warrant for a claim this project has been making informally: **Deliberus's value scales with corpus coverage rather than with user count.** A hundred readers do not each need their own premises.\n\n**And the honest caveat, from the same literature.** Thompson (*Journal of Economic Perspectives* 2012) argues that while the cost-versus-cumulative-quantity relationship is one of the best-documented regularities in economics, its attribution to *organizational learning* is under-evidenced. Wright himself offered three mechanisms and only one is unambiguously learning; the others are static scale economies and cheaper materials at volume. That distinction transfers exactly. Observed premise reuse in this corpus could be genuine transfer, or it could be **topical clustering** — an artifact of which sources happened to be extracted, which is the analogue of \"cheaper materials at volume.\" A curve is not a mechanism. §6 takes this seriously rather than around.\n\n## 6. The first measurement, including one that corrects §4\n\nMeasured against the live public API on 2026-08-13, because a cost curve asserted from theory is exactly the kind of claim this project is supposed to distrust.\n\n**Corpus**: 4,628 Claims, 712 Questions, 82 Concepts, 23 Sources. Edges: SUPPORTS 2,074, ATTACKS 1,770, USES_CONCEPT 479, QUALIFIES 139, DECOMPOSES_INTO 130, **SIMILAR_TO 48**, DEFINES 21.\n\n**Concept reuse** (`usage_count` on `/concepts`): 479 usages across 82 concepts, mean 5.8, median 6, max 21. **71 of 82 concepts — 87% — are used more than once.** Only two singletons. Reuse is real and it is the norm rather than the exception.\n\n**Distribution shape**: fitting log-rank against log-usage gives a slope of **−0.51** with R² 0.70. That is markedly flatter than Zipfian (−1), and the top 29% of concepts carry 50% of usages.\n\nThree readings, in descending order of how much they change the argument.\n\n### 6.1 Reuse is running through the concept layer, not the claim layer — which corrects §4\n\n§4 built the cost engine on *premise* reuse and named claim-level machinery: concept unification, cross-source auto-connect, mint-time dedup, shared-premise detection. The measurement says the concept layer is doing the work (479 usages, 87% reused) while claim-level similarity is close to absent: **48 SIMILAR_TO edges across 4,628 claims**, at a 0.80 cosine threshold.\n\nSo the channel is not the one the argument named. Either claims genuinely are not similar across a topically diverse corpus, or linking is under-running — and from outside the LAN that cannot be settled, so it is a question rather than a conclusion (local-session step: count claims with embeddings against total claims). Either way the practical consequence is immediate: **if the cost curve is the engine, the concept layer is the part of it currently turning**, and concept governance moves up the priority list relative to claim-level dedup.\n\n### 6.2 A flatter-than-Zipf slope is the better shape, cautiously\n\nA steep power law would mean a handful of concepts dominate: early coverage captures nearly all the available reuse, then returns fall off a cliff. A slope near −0.5 with reuse spread across 87% of concepts means the benefit accrues more evenly and **keeps accruing for longer** — less of an early windfall, a more gradual and more durable decline. For a commons intended to run for decades that is the preferable curve.\n\nHeld loosely, for two reasons. n=82 is small. And 23 sources concentrated in a few debate pairs is exactly the condition under which apparent reuse is topical clustering — Thompson's critique arriving in our own data. This is a first datapoint, not a law, and the test that would separate the readings is whether the slope holds as sources diversify.\n\n*Update 2026-08-13.* The corpus has since grown to **25 sources and 90 concepts, 79 of them reused (88%)**, after dogfood run 6 added a legal-doctrinal pair — a register the corpus had never held, which is the diversification this section asked for. The reuse fraction barely moved (87% → 88%), which is mild evidence for the flatter-than-Zipf reading rather than for topical clustering. Two sources is far too few to re-fit the slope on, so the figures above stand as the measurement of record and this is a note, not a revision. Run 6 did, however, supply the *mechanism* the section was missing: reuse runs through concepts rather than claims because claim-level cosine measures vocabulary while cross-domain kinship is structural ([dogfood run 6](dogfood-run-6-israel-palestine-cross-domain.md) J4).\n\n### 6.3 The cost curve has a safety condition the argument omitted\n\nThe concept data carries `sense_count` alongside `usage_count`, and the two come apart in a way that matters:\n\n| Concept | Uses | Senses |\n|---|---|---|\n| self-interest | 21 | **1** |\n| intelligence | 17 | **4** |\n| justice | 13 | **4** |\n| public good / welfare | 13 | 2 |\n\nHigh-reuse, low-polysemy concepts transfer cleanly and are pure cost reduction. High-reuse, **high**-polysemy concepts are where reuse turns hazardous: the same word carrying four senses across seventeen usages is precisely the condition under which naive linking manufactures false agreement. This is the embeddings-flatten-meaning tension the corpus has long flagged, appearing for the first time as a *measurable quantity* rather than a worry.\n\nSo the cost curve gains a condition: **reuse lowers cost only where sense is stable, and `usage_count × sense_count` is already a map of where reuse is cheap and where it is dangerous.** That map costs nothing to produce and does not exist yet.\n\n### 6.4 A correction to yesterday's support-to-attack claim\n\n[incentives-analysis.md](incentives-analysis.md) §4.2 argues that supports are self-serving while attacks are altruistic, so contribution over-collects one, under-collects the other, and QBAF strength inflates invisibly.\n\nThe measured corpus ratio is **2,074 supports to 1,770 attacks — 1.17 to 1**, which is nearly balanced and does not show the asymmetry.\n\nThe reason is that the corpus is overwhelmingly machine-generated: CQ polarity scaffolding mints both directions by construction, and adversarial pairing extracts opposing texts together. **So the asymmetry is a prediction about *human* contribution, and this corpus cannot test it yet.** That makes the finding more useful rather than less, because it supplies the calibration point: the instrument is the support-to-attack ratio of **human-contributed** edges measured against the machine-generated baseline of 1.17, and divergence from that baseline is the early-warning signal. Recorded as a correction rather than quietly amended.\n\n### 6.5 What stops the analysis, and it is fixable in an afternoon\n\nYou cannot fit a Wright curve without cost data. Deliberus has volume and no cost: CLAUDE.md records that **cost tracking is not implemented** — extraction is *estimated* at $0.27–0.37 with no token logging per activity.\n\nSo the most interesting idea in this document is, at present, **strictly unmeasurable**. Not unproven — unmeasurable, for want of two lines of logging. Instrumenting per-extraction token counts and cost would convert the cost curve from a metaphor into a curve that can be fitted, and would let the project state a learning rate for its own corpus, which is the kind of number that belongs in a funding application and does not currently exist.\n\n## 7. What to do with this\n\nNothing here is decided. Ranked by evidence-over-effort, now that §6 has moved two items:\n\n1. **Instrument per-extraction cost.** Promoted to first because it is the cheapest item in this document and it unblocks everything else in it: without token and cost logging the cost curve cannot be fitted, and the project cannot state a learning rate for its own corpus. Two lines of logging turns the central idea here from a metaphor into a measurable quantity.\n2. **The three-arm synthesis experiment** (no synthesis / opaque / inspectable, scored on disagreement preservation). Resolves a real tension, is the strongest differentiating result available, and either wins Deliberus the biggest cost lever or vindicates the current stance with evidence rather than caution. Build plan and prerequisites: [synthesis-build-plan.md](synthesis-build-plan.md).\n3. **The `usage_count × sense_count` map** (§6.3). Free to produce, does not exist, and it is the condition under which the cost curve is safe rather than flattening — high-reuse high-polysemy concepts are where naive linking manufactures false agreement.\n   - *Added 2026-08-13*: **pre-generated synthesis artifacts** are the same insight applied to the reading side, and they arrived from the founder rather than from this document. Retrieval, traversal and translation are three of the eight cost terms in §2, and a per-source synthesis collapses all three into one fetch — for **every future reader, at a one-time cost**, which is Romer nonrivality behaving exactly as §5 predicts. The bounded-input finding is what makes it cheap: 23 sources and 68 reused concepts, so ~$0.30–0.60 buys the whole corpus. Design and the permanence objection: [synthesis-build-plan.md](synthesis-build-plan.md) § 9.\n4. **The position fingerprint** — divergence-only traversal. Largest reduction in the term readers actually feel, and it composes with the fan-out hinge already registered.\n5. **Ratification.** One click, disproportionate trust return, and it unlocks both the portable artifact and the self-interested pitch.\n\nAnd the reclassification, now corrected by measurement: **treat reuse as the cost curve rather than as a feature backlog — but the layer currently doing the work is the concept layer, not the claim layer.** §4 named claim-level machinery; §6.1 measured 87% concept reuse against 48 claim-similarity edges. Concept governance therefore outranks claim-level dedup until the SIMILAR_TO question is settled locally.\n\n---\n\n## 8. What to measure, given that the unit of value is the commons\n\n*Added 2026-08-16, after the founder rejected an individual-uplift evaluation frame with the public-library analogy (recorded in [session19-decomposition-corrections-and-the-ontology-turn.md](session19-decomposition-corrections-and-the-ontology-turn.md) §6). This section answers the follow-on question: if the unit of value is the commons rather than the user, what actually gets measured?*\n\n### 8.1 The table in §1 is reader-side only, and the founder's prediction is not\n\nThe eight terms above all name something **a reader pays**. The registered prediction names something **the builder pays**: *\"the investment in terms of defining the right ontology and filling up the graph with extractions and human- or auto-populated implicit premises, evidence and shared fundamental knowledge.\"*\n\nThose are different quantities with different owners, and running them together under one word is what made \"the falling cost of a descent\" unusable as a measure. Three units were in play:\n\n| Unit | Whose cost | Status |\n|---|---|---|\n| **Per extraction** | Ours, in compute | Not implemented; §6.5 |\n| **Per descent** | The reader's, in time and effort | **No instrument, and this is what the founding conviction is about** |\n| **Marginal, per source, given the corpus** | Ours, against corpus maturity | The one Wright's law fits, and the one the prediction makes |\n\nThe consequence is that §7's first item is narrower than it looked: **cost logging scores the founder's prediction and does not touch the founding conviction.** Reader-side cost needs its own instrument, and the descent rate below is the only crude proxy currently designable.\n\n### 8.2 The filter, which matters more than the list\n\nFive measures fit a commons: coverage of the contested space, concept reuse, reachability, the falling marginal cost, and whether the residue map moves. The project's own standing rule applies to all of them — **a counter-instrument that can only confirm is decoration** — so each has to answer one question: *what value of this would make us change course?*\n\n| Measure | Can it fail? | Status |\n|---|---|---|\n| **Concept reuse** | Yes — falling toward zero means nonrivalry is not operating and §5 weakens | Measured, and half of it already reads badly (claim-level near zero) |\n| **Residue map** | Yes — a residue fraction that stays high as it grows is the wager losing | Measured, but at two classified termini it is noise; growing it is dogfooding work, not metrics work |\n| **Descent rate** | Yes — agents that fetch a synthesis and never fetch what it cited mean we built the conclusions layer | Designed, unbuilt; the first outcome-ish signal available |\n| **Marginal cost** | Yes — unit cost not falling as the corpus grows refutes the prediction | Unmeasurable for want of logging |\n| **Coverage of the contested space** | **No, as stated** — there is no denominator, so any figure would be invented | Needs reformulating: coverage *within a debate* has a denominator the sources supply |\n\nTwo cautions carry more weight than the table. **Do not build a dashboard of five** — a panel of commons metrics is precisely the shape that gets read at a glance and acted on never; two numbers that can genuinely fail are worth more than five that reassure. And the commons frame is not an exemption from falsifiability: it settles the *unit* of value without exempting the artifact from being good, which is why the honest place to point today is claim-level reuse rather than the 87% figure.\n\n### 8.3 The curve has an independent formalism in another field\n\n*Added 2026-08-16.* Assembly theory (Walker, Cronin) measures an object by the minimum construction steps required to build it **where anything already built may be reused**, and its stated reason is stronger than this document's: above a complexity threshold the space of possibilities cannot be searched at all, so building from existing parts is not cheaper, it is **the only construction available**. Its joint-assembly measure also names what §6.1 measured without a vocabulary for it — the corpus is **Jupiter at the claim layer** (high diversity, different bonds, flat space) and **Earth at the concept layer** (same bonds recurring across many objects, depth). And its hardest open problem, identifying the boundaries of repeated units, is claim-sameness. Full treatment, including the parts that do not transfer and the warning that reuse alone would rank clichés first: [assembly-theory-and-the-reuse-mechanism.md](assembly-theory-and-the-reuse-mechanism.md).\n\n### 8.4 The lever is not the same as the measure\n\nCost logging would give the curve. It would not bend it. The mechanism the founder named in April is a saturating upstream layer of **shared premise subtrees**, and dogfood run 6 showed why that layer is not appearing at claim level: cosine measures vocabulary, and cross-domain kinship is structural. So the quantity that would actually move this curve is **implicit-premise recall**, independently identified by run 3F as the pipeline's weakest and most consequential pass.\n\n**Audited 2026-08-16, and the result changes the priority.** The graph holds **16 implicit-premise claims in total**, 18 of 25 sources have none, and only one of the sixteen is linked to anything. Four independent build gaps explain it: the interactive extraction path never calls the pass at all, the prompt is instructed toward scarcity, the object it produces is a within-argument bridge rather than shared background, and **`decompose_claim` never looks for an existing claim to reuse**, which makes the April mechanism's \"descents terminate in already-known subclaims\" impossible by construction rather than merely rare. The prediction has not been tested and found wanting; **it has never been given the conditions it names.** Full audit: [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md) § The mechanism audited.\n\n---\n\n## Sources for the formalism (§5)\n\n- Wright, T.P. (1936). \"Factors Affecting the Cost of Airplanes.\" *Journal of the Aeronautical Sciences* — the original learning curve.\n- Arrow, K. (1962), \"The Economic Implications of Learning by Doing\"; Boston Consulting Group (1970) — the generalization to the experience curve across whole industries.\n- Nagy, B., Farmer, J.D., Bui, Q.M. & Trancik, J. (2013). \"Statistical Basis for Predicting Technological Progress.\" *PLOS ONE* 8(2). Six laws hindcast against 62 technologies; Wright's forecasts best. [journals.plos.org](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0052669)\n- Thompson, P. (2012). \"The Relationship between Unit Cost and Cumulative Quantity and the Evidence for Organizational Learning-by-Doing.\" *Journal of Economic Perspectives* 26(3), 203–24 — the caveat that matters: a well-documented curve is not a demonstrated mechanism. [aeaweb.org](https://www.aeaweb.org/articles?id=10.1257%2Fjep.26.3.203)\n- Romer, P. (1990). \"Endogenous Technological Change.\" *JPE* 98(5) — nonrivalry, increasing returns, and growth tied to the aggregate stock. Jones's 25th-anniversary essay is the clearest short exposition: [stanford.edu](https://web.stanford.edu/~chadj/Romer1990-25thAnniversary.pdf)\n- Louwen, A. \"The experience curve: concept, history, methods, and issues\" — learning rate and progress ratio derivations. [research-portal.uu.nl](https://research-portal.uu.nl/ws/portalfiles/portal/235996048/3-s2.0-B9780128187623000029-main.pdf)\n\n**Read depth**: substantial extracts plus abstracts for all of the above; no full texts. §6's measurements were taken directly from `https://deliberus.com/graph/stats` and `/concepts` on 2026-08-13 and are reproducible by anyone — the Zipf fit is an ordinary least-squares slope over log-rank against log-usage for the 73 concepts with non-zero usage.\n\n---\n\n**One later qualification worth carrying**: this document treats lowering the cost of understanding as the product's core work. [structure-versus-scale.md](structure-versus-scale.md) adds the boundary condition — as models improve, the *extraction* term of that cost keeps falling on its own, so the durable value migrates toward the terms no model improvement touches (persistence, addressability, contestability). The cost curve stays the right frame; which term dominates it moves.\n\n**See also**: [vision.md](../vision.md) § Opacity Is a Cost, Not a Mystery · [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md) · [fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) · [bridging.md](../bridging.md) · [analysis-and-attunement.md](../analysis-and-attunement.md) · [the-missing-layer.md](../the-missing-layer.md) · `sources/chatgpt-research-thread-2026-08.md` (the inspectable-synthesis passage) · `sources/chatgpt-literature-dump-2026-08.md` (the N=879 steering result)\n"}