{"path":"research/world-completeness-and-its-relatives.md","content":"# World completeness: the third check, and everything in the corpus it touches\n\n*2026-10-14. Written at the founder's ask to relate the third kind of completeness to everything it\nis related to, so he can read through it and ponder it. The three kinds come from the outside audit\nof the same morning (`claim-parts-and-inferential-reasons-2026-10-14.md`): content completeness,\nargument completeness, evidence completeness. He calls the third \"world completeness\", and the name\nis better than the audit's, since it says what the check is against. The corpus graph was asked\nfirst (search, then the notes it named), and each passage below was opened. Nothing here is ruled;\nthe question it lands on is closing question 3.*\n\n## 0. The short version\n\nWorld completeness asks: has everything in the world that bears on this claim been brought into the\nrecord? No test can pass it, since the record cannot contain what nobody brought, and that is the\none law every relative below is a consequence of: **absence of a record is not a record of\nabsence.** What *can* be recorded is the search: what was looked for, by whom, with which moves, and\nwhen it stopped. So every instrument this project has for the third completeness is an instrument\nabout the search, never about the world; every principle is about who may be trusted to say the\nsearch was enough; and every adversary is somebody with a reason to stop looking early. The one\nhuman-reserved error class that nothing inside the graph can see, coherent absence, is this\ncompleteness failing silently.\n\n## 1. The three completenesses, told apart\n\n| | Asks whether | Tested by | Can it be passed |\n|---|---|---|---|\n| **Content** (of a split) | the parts say everything the claim said, and nothing more | the rebuild check: can the claim be reconstructed from its parts alone (`CoverageReport.reconstructed`; `the-jhu-decomposition-line.md`) | yes, for one cut |\n| **Argument** (of a step) | the chain has no gap and nothing removable, and each step applies one stated rule once | the refutation question and the removal test (ledger AG14; the third clause ruled 2026-10-14) | yes, for one chain |\n| **World** (of a case) | every consideration and piece of evidence that bears on the claim is in the record | nothing; only the search can be tested | never; approached |\n\nThe first two are closed checks on an object in front of you. The third is open: its object is\nwhat is *not* in front of you. That is why it has no instrument of its own and dozens of relatives.\n\n## 2. The law, and where the corpus states it\n\n- **Open world.** A horizon reading is open-world: *no cross-attack exists* reads as *unexamined\n  between*, never as *agreed* (`what-belongs-in-the-ontology.md` § 6c, outside names; ledger O11: a\n  computed reading is defeasible, never final, and says so).\n- **The Tractatus's totality clause** (1.11, 1.12): the world is the facts *and by these being all\n  the facts*; the totality determines also what is not the case. The corpus's reading: that is the\n  one assumption a record cannot make about itself (`the-tractatus-and-deliberus.md`, the table\n  row on totality; ledger Arc AJ, unruled).\n- **A null result is informative only if the record shows what was looked for**: the\n  intelligence, forensic and historical professions all keep the search beside the finding\n  (`what-counts-as-evidence/law-intelligence-history-journalism-forensics.md`; the reader's finding\n  on `convergence.md`, 2026-10-11).\n- **Keynes's weight of an argument** (1921): how much has been examined is a quantity apart from\n  which way it points; the code keeps them apart (`evidential_weight.py`), and the invariant that an\n  unexamined claim carries no number is the law applied to the badge.\n\n## 3. What the code touches, and what each instrument can and cannot see\n\nEvery instrument below reads the record. None reads the world. Each is listed with its blind spot.\n\n- **The completeness oracle** (`deliberus/completeness.py`, July): walks a claim's parts, reports\n  value premises with nothing under them as copout axioms, and returns `fully_exposed` when every\n  leaf bottoms out. It reads only what was written, so a lossy split reports the same completeness\n  as a faithful one (`the-residual-error-taxonomy.md` § 2, the same class one level down), and\n  `fully_exposed` at depth 0 flatters by construction (dogfood run 4). It once called the corpus's\n  first classified terminus a copout axiom because its query never selected the terminus type.\n- **The horizon reading** (`deliberus/horizon.py`, 2026-09-16; ruled *read, never set*): an\n  arrangement nobody has tested reads as *unexamined*, never as a horizon. What it cannot see:\n  *the arrangement looks settled because the move that would unsettle it has not been contributed*\n  (`session33-the-horizon-vocabulary.md`; the ontology note § 6c.3). Invitation is the only\n  instrument for that.\n- **`worth_asking` and the hinge** (`deliberus/hinge.py`, `question_ordering.py`): rank the open\n  questions by how much an answer would move the verdict. They order the search and never end it:\n  the only stopping rules today are *answered everything* and *the user walked away*\n  (`active-inference-context-acquisition-and-deliberus.md` § 2, which supplies the missing cost\n  term; Terenin's Pandora's boxes are the non-myopic form).\n- **The standing-questions count**: 2,084 standing questions on the live graph's reasons, one\n  answered (`checks-on-machines-and-people.md` § 6, 2026-10-12). The awareness of what has been\n  tried is a count, and the count today says almost nothing has.\n- **The cross-source premises pass** (`cross_source_premises.py`, propose-only): finds absences\n  *within* supplied material, which a within-source pass cannot see by construction. It cannot see\n  that the supplied material is unrepresentative.\n- **The disagreement-preservation instrument**: its denominator is the conflict edges that already\n  exist, so it is silent exactly where the linker found nothing (`the-scrutiny-gap.md`, read in\n  code).\n- **The map's one marked place for the region's rarest claim** (`specs/map-view/design.md`): the\n  only surface that shows a reader where the record is thin on purpose.\n- **The measurable form the evidence dossier proposed**: *no attacks on a claim whose sources are\n  all one-sided* (`what-counts-as-evidence.md`, the line on coherent absence given a measurable\n  form), and the per-cluster ingestion-balance count (counting only, no model), proposed and\n  unbuilt (`the-residual-error-taxonomy.md` § 2).\n\n## 4. The rulings and principles that govern it\n\n- **Q34 (ruled 2026-09-30, for now)**: a reason that fails drops out; failure counts against a\n  claim only through a contribution that says why, and the third such contribution is **a thorough\n  search that found nothing, argued as an argument from ignorance**, the catalogue's `ignorance`\n  scheme, whose critical question is whether the search was thorough. With it, the display state\n  *nothing stands (N checked)*. `checks-on-machines-and-people.md` § 6.5 places it: *the search was\n  thorough* is the claim that turns *every listed reason failed* into *probably false*, and it is a\n  claim, so it can be attacked. **None of it is built.**\n- **The two human-reserved classes** (the harness conviction, ratified 2026-08-30): coherent\n  absence (the machine cannot know what its corpus silently omits) and frame lock-in (it cannot\n  stand outside the frame the graph crystallized). World completeness is the first of these, named\n  as an error class (`the-residual-error-taxonomy.md` § 2; `what-human-judgment-is-for.md` § 1, the\n  row *no, in principle*).\n- **Never ask a mind to audit its own frame**: the absence judgement comes from someone other than\n  the author; a model asked what is missing searches the frame it is inside. The derivability\n  ladder measured this one-sidedly: on thirteen real arguments a model raised every consideration\n  human commenters raised, and its zero at the corpus-absent rung is model-graded and biased toward\n  zero by construction (`derivability-of-missing-considerations.md` § 10). What it named as beyond\n  its reach was evidence from a position.\n- **\"Nobody here has said X\"** is the first of the four human acts and a low-rung one: it needs a\n  position, not ontology fluency (`what-human-judgment-is-for.md`, Act one). Minted claims are\n  anti-hidden-profile; this act is anti-cascade (ledger E16).\n- **No gap displayed without a move offered**, and its 2026-09-09 extension: a frame gap, once a\n  person voices one, gets a contribution path like any other gap. The standing violation is the\n  oracle's `unsupported_value_premises`, which reports and offers nothing.\n- **The sacredness re-scope** (ratified in substance 2026-08-25) and the deep-descent finding: only\n  the **moves-tried record** distinguishes bedrock from arrested inquiry (`deep-descent-experiment.md`,\n  the specimen *the badness is not a further fact beyond how it feels*). A horizon is where mapping\n  currently ends *under the current moves*, dated: the move-set is part of the record of the search.\n- **Every taxonomy breaks on the next register**: a missing slot in a schema is coherent absence\n  inside the ontology, which is why every enum carries a confession value rather than a slot per\n  register.\n- **If a system cannot confess failure, it will report success**: an oracle that cannot say what it\n  did not look at reports *fully exposed*. The confession principle is the third completeness's\n  honesty requirement stated for instruments.\n- **Every published count labels its provenance**: a count that does not say what it is a count of\n  overstates what is held, and *how much has been examined* is itself such a count.\n- **Visibility plus invitation** (the founder, 2026-08-21): a lean in a synthesis must be\n  attributable to the contributed reasoning, and any unattributable lean is a platform-level\n  finding; the 2026-08-25 caveat is this note's § 5 below, since a lean caused by what got ingested\n  is indistinguishable in-graph from a lean in the reasoning.\n\n## 5. The adversaries of the third completeness\n\nEach is somebody, or something, with a reason for the search to stop early\n(`fractal-scales-and-temporal-frame.md` § Three classes of adversary).\n\n- **Entropy**: input friction, staleness, the long tail. The head of a debate saturates fast and\n  the tail never closes (`saturation-and-the-long-tail.md`: Heaps-shaped growth of distinct content,\n  more than half of noun phrases unique to low-frequency arguments), so world completeness is\n  asymptotic by the shape of language, before any adversary.\n- **Motivated stopping** (the stakeholder class): the agent stops when the case looks good. The\n  reasoner rung's dominant failure, with the completeness oracle and the sorry frontier as its\n  instruments (the scale ladder's second row).\n- **Selective legibility** (the strategy class): map your strong arguments in full and leave the\n  weak ones unstated; the graph shows you as thoroughly examined because it measures a *claim's*\n  exposure and never an *actor's*; no instrument sees it (`incentives-analysis.md` § 5). The\n  payoff-removal form is the opponent-proposes-you-ratify design, not a detector.\n- **Asymmetric ingestion**: an absent premise accumulates no attacks, so what got ingested becomes\n  strength (`the-residual-error-taxonomy.md` § 2, class 2, the least defended). Policy defence:\n  ingest by debate cluster, never by source (`incentives-analysis.md` § 6b).\n- **Obfuscated arguments** (Barnes & Christiano): error spread thin across many barely-checkable\n  steps, a completeness-of-checking attack (the standing threat model, entry 1).\n- **Shared origin** (ledger Q4): three reports from one source counted as three; the record of\n  where each piece came from is the missing half, and what every field that combines evidence keeps.\n\n## 6. The outside sources, each with what it adds\n\n- **Pollock** (`pollock-against-accrual-and-the-overlap-check.md` § 1): the one place a whole\n  exceeds its parts is enumerative induction, where the *number* of instances matters, so more\n  looking legitimately raises a claim of that kind; and two reasons for one claim are *almost\n  invariably* one direct argument and one statistical syllogism, so what a second look adds is\n  usually a reliability claim, not a second fact.\n- **Prakken's second principle** (2005 § 3.2): an accrual makes its elements inapplicable *because a\n  comparison between accruals is meant to consider all available information*. World completeness\n  is the premise under which accruals may be compared at all.\n- **Walton's argument from ignorance** (the `ignorance` scheme in the catalogue): the scheme whose\n  critical questions are *how thorough was the search?* and *would the evidence have turned up if\n  the claim were true?* It is Q34's third contribution and the only scheme whose subject is the\n  search itself.\n- **The open-world assumption** (knowledge representation): the ontology note's outside name for the\n  law in § 2.\n- **Sunstein's hidden profiles; Lu, Yuan and McLeod**: groups bury what only one member knows;\n  every reason minted as a claim is the structural answer, and the group-premise evidence note\n  grades that condition *supplied* (`the-group-premise-evidence.md`, row 3).\n- **Mercier and Sperber**: reasoning is largely post hoc, so the reasons a single mind generates\n  are its own; the other side's reasons arrive only from the other side, which is why the third\n  completeness needs *many different somewheres* (`derivability-of-missing-considerations.md`\n  § 7's phrase) rather than smarter reviewers.\n- **Kahan**: curiosity alone reverses polarization; who goes looking decides what is found\n  (`curiosity-as-growth-fuel.md`).\n- **Terenin, Pandora's box** (Weitzman's rule): when the best remaining box is no longer worth\n  opening, stop; the stopping rule the descent lacks (the active-inference note § 2).\n- **Keynes** on weight (§ 2 above), and **the Tractatus** on totality (§ 2 above).\n\n## 7. Where it lands: closing question 3, and the number's limit\n\nTwo things this note was written to feed.\n\n**Closing question 3**: *when every argument for a claim has failed, who may say the search was\nthorough?* Everything above narrows it. The search is a claim (Q34), so the question is who may\nassert it with what standing: the author (the frame problem says no, in principle); a machine\n(the derivability ladder says it finds what a frame contains and cannot grade what it lacks; a\nmachine's claim would carry machine provenance and move nothing until taken up, AG12); someone\npositioned differently on the thing in question (the harness conviction's answer, and the only one\nthe corpus endorses); or nobody, with the record showing the moves tried and the count of standing\nquestions, and the display saying *N checked*. The recommendation that follows from the corpus as\nit stands: **the search is thorough is a claim anyone may file, it carries its author's provenance,\nits critical question is the catalogue's, and the moves-tried record and the standing-questions\ncount are its evidence**; a machine may propose it and may not assert it.\n\n**The number's limit** (`what-the-number-on-a-claim-means.md` § 4): reading the number as a\nprobability presumes world completeness, since a probability is conditional on *everything*\nrecorded being everything that bears; reading it as the confidence the examined case warrants\nscopes the number to the record and leaves world completeness as the open question it is. That is\nthe deepest reason the pick there is the second reading with probability as the limit: the limit\nis reached exactly when the third completeness has been, and the dial between them is the record of\nthe search.\n\n**What a well-defined mother claim buys** (2026-10-15, on the founder's question whether the\nsensing of need fooled us about this completeness, and whether a well-defined mother claim makes it\nmore feasible): the *space* of what would bear on it, not whether anything in the space does. The\nderivable requirement space (each part's negation, the scheme's questions, one definitional question\nper contested term) is enumerable from the claim alone and can be pre-populated for every claim\nalike; the answers, and the tail outside it (evidence from a position), cannot be. The sensing of\nneed answered two closed questions (a part's, a premise's) and the filing of reasons as parts from\nApril to October made that look like the open one; nobody wrote that it was.\n`retrofitting-the-reasons-channel.md` § 2.\n\n## 8. A reading order, for pondering\n\n1. `the-residual-error-taxonomy.md` § 1 (the law: inspection catches incoherence, never coherent\n   absence) and § 2 (asymmetric ingestion).\n2. `what-human-judgment-is-for.md` § 1 (the table of classes with *human* in the last column) and\n   Act one.\n3. `checks-on-machines-and-people.md` § 6 and § 6.5 (the count of what has been tried, and where\n   the thorough search fits).\n4. `relevance-and-the-two-factors.md` § 7 (Q34 as ruled; *nothing stands*).\n5. `what-belongs-in-the-ontology.md` § 6c (the open-world reading; the horizon as the edge of the\n   moves tried).\n6. `derivability-of-missing-considerations.md` § 3, § 7, § 10 (what a machine reaches; what it\n   named as beyond it; why its zero is biased).\n7. `incentives-analysis.md` § 5 and `fractal-scales-and-temporal-frame.md` § Three classes of\n   adversary (who stops the search).\n8. `active-inference-context-acquisition-and-deliberus.md` § 2 (when to stop asking).\n9. `what-counts-as-evidence.md`, the two coherent-absence passages and § 8c (what every field that\n   combines evidence records).\n10. `the-tractatus-and-deliberus.md`, the totality row (what no record can assume about itself).\n"}