{"path":"research/drowning-in-claims.md","content":"# Drowning in Claims\n\n**Date**: 2026-08-30 · **Type**: founder question, answered from the corpus and then synthesized —\nthe question had never been asked directly, and its pieces were sitting in six different docs.\n\nThe founder, verbatim:\n\n> \"Could Deliberus drown in claims and contributions? Could the sheer amount of text content grow\n> unmanageable and too hard to bring order to? What are natural limitations and mitigations of\n> this? Even barring scenarios where we're spammed with nonsensical claims and questions intended\n> to confuse and sabotage or that aren't earnestly contributing... Could this come to happen?\"\n\n---\n\n## 1. Never asked directly — explored indirectly in six places\n\nA sweep for prior art finds no doc that poses the drowning question as such (the word appears only\nin Singer's drowning child and in *science drowning in slop*). The pieces exist separately: the\ncompression ceiling ([the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md)), **ratification\ndebt** ([incentives-analysis.md](incentives-analysis.md) Stage A — *\"unremarkable at 25 sources and\na legitimacy problem at 10,000\"*), the spaghetti-graph problem and its four scalability techniques\n([academic-foundations.md](../academic-foundations.md) § Scheuer), Kialo's 90-9-1 participation\nshape ([engagement-gradient-priors.md](engagement-gradient-priors.md)), identity scattering\n([the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) class 3), and staged maturity\n(the lifecycle quarantine). This doc is their synthesis.\n\n## 2. It already happened once — and the machine did it, not humans\n\nThe calibrating fact: **the graph has already drowned itself in miniature.** Roughly two thirds of\nall claims are machine-minted critical-question scaffolding (1,741 substantive of 4,852; the\nboth-polarity CQ design creates on the order of five hundred nodes per extraction), and the\ndogfood-run finding F12 was exactly *CQ scaffolding dominates the reading surface*. Meanwhile\nhumans have never been the flood anywhere: argument-mapping's nearest neighbor confirms the classic\n90-9-1 shape, and Kialo's largest debates hold thousands of claims, not millions. **The firehose is\nmachine minting — extraction, scaffolding, and every future daemon — not earnest human\ncontribution.** That is also where the governors already point: the ingest-by-debate-cluster gate,\nspend gates, propose-only minting with rejection-rate pause, and the lifecycle quarantine that\nlets raw material flood in without touching default surfaces. **And the structural fix is proposed\n(2026-08-30, unruled): latent-until-touched scaffolding — the type–token split\n([latent-scaffolding-and-the-type-token-split.md](latent-scaffolding-and-the-type-token-split.md)).**\n\n## 3. The natural limits, and they are real\n\n- **Distinct content saturates even when text does not.** The measured compression ceiling: seven\n  key points, written blind, covered 72.5% of what 233 people independently said about a topic —\n  and the curve flattens hard. Past saturation, earnest contributions are increasingly\n  *re-statements*, so the problem converts from volume into **sameness**. Topics are unbounded, but\n  each topic's map converges; mature-region growth shifts from new nodes to new **edges, evidence\n  and answers** — which is the reuse asymptote seen from the flood side.\n- **Attention is the bounded input.** Contribution follows attention (Simon: a wealth of\n  information creates a poverty of attention), and the engagement gradient means readers vastly\n  outnumber writers by nature. The library test already rules that a reader who never contributes\n  is a success.\n- **Additivity protects pebbles from burial.** The strength model's test-discovered property (I4):\n  small claims are not drowned out arithmetically by crowds of siblings.\n\n## 4. The reframe: volume only costs when unconnected or unmerged\n\nIn a graph, a new claim that is *distinct* and *connected* adds value at any scale (nonrivalry).\nOnly two forms are pathological: the **duplicate** (identity scattering — the same claim in forty\nwordings, evidence diluted across them, false disagreement manufactured between them) and the\n**orphan** (noise that touches nothing). So the anti-drowning programme is not a new workstream —\n**it is claim-sameness plus auto-connect plus honest strength, which are already the named core\nactivities.** Drowning is the reuse flywheel's failure mode restated, and the flywheel doc's\ncondition (identity must be cheap to decide) is the drain that keeps the pool from overflowing.\n\n## 5. The threat ordering, even barring bad faith\n\n| Rank | Flood | Existing answer | State |\n|---|---|---|---|\n| 1 | Machine minting (extraction, scaffolding, daemons) | cluster-gated ingestion; propose-only + rejection pause; lifecycle quarantine; spend gates | policies ruled; quarantine live at retrieval |\n| 2 | Sameness scatter (duplicates diluting evidence) | the merge design (add-never-overwrite, reversible, one-step invariant); structural identity | designed; auto-merge gated on the review-based eval (seeded wrongs + blind subset, founder-revised 2026-08-31) |\n| 3 | Ratification debt (minting outpaces examining, by construction) | materiality — most entries never examined, *by design*, honestly recorded; the gray band; the triage feed routing scarce human warrant | states designed, unbuilt; feed unruled |\n| 4 | Reader drowning (the hairball) | nobody reads a graph linearly: entry by question, hinge-ordered traversal, badges as compression, progressive disclosure, bundles, semantic zoom, lifecycle-filtered retrieval | partly live, partly design |\n| 5 | Maintenance drowning (staleness, drift — the class-killer per Star & Ruhleder) | the staleness daemon (deliberately the first daemon), supersession, the temporal rung | staleness half built |\n\n## 6. What would show drowning approaching — measurable, no LLM needed\n\nCandidate early-warning dials, each countable from the graph: **duplicate rate at mint time**\n(near-matches per new claim) · **orphan rate** (claims with zero edges after auto-connect) ·\n**the debt ratio** (minted : examined) · **the unclassified rate** (`does_not_fit` frequency) ·\n**scaffolding fraction** (minted : substantive). A sixth arrived from the follow-up research sweep: **the\nGood-Turing novelty rate** — the measured probability that the next earnest contribution adds\nsomething distinct (≈ the singleton fraction from mint-time dedup, free to compute), read TOGETHER\nwith the `does_not_fit` rate so apparent saturation can be told apart from frame ossification. A\ndrowning dashboard is these six numbers; every one is computable today. Unruled proposal. Deepening\nof § 3's ceiling claim — the two-regime curve, the value-bearing tail, the recombination era:\n[saturation-and-the-long-tail.md](saturation-and-the-long-tail.md).\n\n## 7. Honest state\n\nThe mitigations are mostly **designed, not built**: the merge machinery is gated on evidence that\nhas not been collected, ratification has zero actions, the triage feed is unruled, and the\ncompression ceiling was measured on someone else's corpus. So the answer to *\"could this happen?\"*\nis: **yes — through the machine door, and through sameness — and the reason for calm is not that\nthe risk is small but that its remedies are exactly the already-prioritized work.** Nothing new\nneeds inventing; the priority stack (sameness, strength honesty, staleness, triage) IS the\nanti-drowning programme. That alignment, rather than any single mechanism, is this doc's finding.\n\n---\n\nCross-references: [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) (the ceiling; the\ncondition) · [incentives-analysis.md](incentives-analysis.md) § 6b (ratification debt; supply-side\ngates) · [soft-canonical-clustering-and-reversible-merge-semantics.md](soft-canonical-clustering-and-reversible-merge-semantics.md)\n(the merge design) · [the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) (identity\nscattering) · [the-ratification-record-and-the-triage-feed.md](the-ratification-record-and-the-triage-feed.md)\n(the warrant allocator) · [engagement-gradient-priors.md](engagement-gradient-priors.md) (90-9-1;\nthe library test) · [../academic-foundations.md](../academic-foundations.md) (Scheuer's spaghetti\nproblem; Bar-Haim's ceiling) · [dogfood-run-1-friction-log.md](dogfood-run-1-friction-log.md)\n(F12; I4)\n\n## 8. The founder's flood: irrelevant-but-connected text (2026-09-03)\n\nFounder: *\"What I'm imagining is drowning in irrelevant content/text. Which can definitely\ndestabilize and confuse and derail.\"* Section 4 named two pathological forms, the duplicate and\nthe orphan. This is a third: **the tangent** — content that is distinct, that auto-connect *will*\nattach (a 0.65 similarity and a judge saying \"qualifies\"), and that is nevertheless noise: the\nplausible-but-pointless question, the near-topic ramble, the earnest derail. It is not chiefly a\nsaboteur's weapon. The machine makes it itself — the 2025 critical-question shared task measured\nroughly half of generated questions as unhelpful or invalid, and a model judging its own rated\n75–85% of them useful — and earnest contributors make it constantly.\n\n**Three victims, three different answers.**\n\n*The numbers.* Irrelevant content moves no strength unless it sits on an edge, and an edge's\neffect is governed by its own strength. The correction this forces on the deflation defense\n(derivability doc § 4, § 9): a requirement *inside* the derivable space was said to cap\n\"automatically\". Too generous — the machine's own list is half junk, so a materialized junk\nquestion at 0.5 would cap the whole at neutral forever. **Relevance is a graded, contestable\nproperty of the edge, and the cap weight is the relevance strength**: presumed-but-rebuttable and\nmodest for a scheme's standard question, argued for anything else, and answerable by *not\napplicable* with a reason — a first-class answer, attackable like any other, which a skeptical\nsecond-family judge can propose at scale. Irrelevant content has no lever on the arithmetic\nbecause relevance *is* the lever, and it is never free.\n\n*The structure.* Tangents accumulate as low-edge, low-temperature regions. They do not spread,\nbecause nothing points at them; the orphan rate and the debt ratio (§ 6) count them; they cost\nbytes, not coherence.\n\n*The attention.* This is the real one, and the answer is that **the reader never sees the store**.\nSurfaces show the hinge-ranked few — a question is offered only if a decisive answer could move a\nbadge across a display band (`worth_asking`), the triage feed shows uncertainty tails rather than\neverything, disclosure is progressive, and the UX directive already rules that *conservative\nrelevance beats semantic reach — tangential matches are worse than honest thinness*. Volume only\ncosts when it reaches a surface, and the surfaces are rationed by what could change a reading. In\na live session the same holds: a derail lands as a claim with no edges into the debate's cluster,\nvisibly off-map, and the map itself says *this touches nothing here yet*.\n\n**The honest residual.** An earnest, argued, connected tangent is indistinguishable from a real\nconsideration until someone judges its relevance — and that judgment is human-performable (low\nrung: *does this bear on the question?*) and machine-proposable, but every such judgment spends\nattention, which is the scarce input. So the flood's real cost is the triage load, and the triage\nfeed's rationing is the defense. That sits at rank 3 of § 5's ordering, where it already was.\n"}