--- slug: the-dictionary-that-eats-itself passed: 2026-10-18 by: the assay line (research/assay-line/claude-1018c/), Claude fixer (claude-opus-5-5) directed by claude-assaying-boyle-1014; source record by a Claude agent (claude-sonnet-5-5), checked by a Claude agent (claude-opus-5-5) checked: 83 confirmed: 56 wrong: 25 # 8 WRONG + 17 MINOR unverifiable: 0 observed: 2 stale: 0 --- This was a directed repair pass over the 25 triage items in research/assay-line/claude-1018c/the-dictionary-that-eats-itself/triage.json (the checker read the layer at 87ef051577; the fixer at 98fe4927af). The checker raised 8 WRONG lines and 17 MINOR. The fixer re-read the evidence for every WRONG line itself: Crossref for Harnad 1990 (issued June 1990), the shipped lexicon.json with the verifier's own parse (digital, machine and network are headwords; effect and result have no arrow to the other five), the full text of arXiv 1411.0129 (each MinSet "part-Core and part Satellite"; "our algorithms are able to compute their MinSets", with CPLEX for Longman and Cambridge) and the Gutenberg catalogue pages for #660, #671, #672, #673 and #29765. All eight WRONG lines held and were fixed, with one Corrected 2026-10-18 line on the page and in the markdown; sixteen MINOR lines were fixed in a few words each, and one (the recompute-it-yourself instruction) was listed. The CONFIRMED and OBSERVED lines below are the checker's, carried over unchanged. ## Claims - MINOR | [dek-724-knot] Pick any word and follow its definition, word by word: you fall into a knot of 724 words that all define each other, with no exit to anything grounded. | research/the-dictionary-that-eats-itself/lexicon.json | Core 724 recomputed, but from 48 of the 1,200 words no chain of lookups ever reaches the Core (the 47 out-degree-0 words plus menu, whose only arrow goes to the sink bill). In 20,000 random click trails from "word", 23.8% stop at such a word before any repeat. The page's own "false floor" box concedes these; the dek's "any word" does not. (fixed: any-word-reaches-core) - WRONG | [dek-harnad-26] Under it all sits Harnad's 26-year-old open question, left open | https://api.crossref.org/works/10.1016/0167-2789(90)90087-6 | Crossref: Physica D 42(1-3):335-346, issued June 1990. The layer is dated 2026-07-03, when the question was 36 years old. 26 is the 1990-to-2016 interval the body uses correctly ("Twenty-six years later"); the dek turned it into an age. Live on /archive/. (fixed: harnad-26-year-old) - MINOR | [body-core-98] run Tarjan's algorithm in-page to see the whole dictionary collapse into a single 724-word strongly-connected Core that 98.6% of lookups never leave | src/content/strata/the-dictionary-that-eats-itself.md | 724 and 98.6 recomputed (3,800 lookups from Core words, 52 leave). "The whole dictionary" is 60.3% (724 of 1,200; the verifier's own "Core as % of vocabulary 60.3"). Markdown body; I did not find it rendered on the live page or /archive/. (fixed: core-is-whole-dictionary) - MINOR | [body-every-number] Every number on screen is recomputed from first principles by verify-the-dictionary-that-eats-itself.mjs (20/20 green) | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | Not every number: the verifier never computes the greedy percentage (38.8%) or its first six words, the per-word stay percentages in the trail status, or the bloom ring counts, and it asserts closure(word) only as ">= core size" (it prints 739). Same wording in the footer ("Every number shown is recomputed by ..."). (fixed: every-number-verified) - MINOR | [html-proper-names] every proper name, archaic term and rare word is dropped | research/the-dictionary-that-eats-itself/lexicon.json | Proper names that are headwords survive as arrows: hall -> london, setting -> paris, between -> boston, search/list/jack -> john, football -> india, chinese -> china, canadian -> canada (recomputed with the page parse). (fixed: proper-names-dropped) - MINOR | [html-handful-clicks] Within a handful of clicks the trail turns red: you're back at a word you already passed. | research/the-dictionary-that-eats-itself/lexicon.json | Shortest loop from the presets is 2 to 6 clicks, but in 20,000 random trails per preset the median is 11 to 15 clicks (90th percentile 24 to 30), and 21% to 32% of trails reach a dead-end word before any repeat. Matches N1. (fixed: handful-of-clicks) - MINOR | [html-paper-shape] That is the shape the paper by Philippe Vincent-Lamarre, Stevan Harnad and colleagues found in 2016 in learner dictionaries and WordNet | https://arxiv.org/pdf/1411.0129 | The paper found it in four dictionaries: Longman and Cambridge (learner), Merriam-Webster (248,466 words) and WordNet ("The Kernel of each of the four dictionary graphs turns out to contain an SCC much larger than all the others"). The page drops Merriam-Webster, the one closest to Webster's 1913. Also README line 98 area. (fixed: paper-dictionaries-omits-webster) - WRONG | [html-false-floor-why] modern terms whose real defining words (digital, machine, network) simply fell outside this 1,200-word slice. | research/the-dictionary-that-eats-itself/lexicon.json | digital, machine and network are all inside the slice (digital is headword 191 of lexicon.json's 1,200, network 154, machine 765). The five glosses are the 1913 senses and use none of them; their missing words are datum, use (a stopword), compute, serve, photograph. N1's note "all three are in the Core" is itself wrong: only machine is; digital and network are outside the Kernel (in-degree 0). (fixed: defining-words-outside-slice) - MINOR | [html-widen] Widen the vocabulary and the arrows reappear, pointing right back into the Core. | research/the-dictionary-that-eats-itself/lexicon.json (recreated in scratch: vocabulary widened along google-10000-english by README step 1, shipped glosses kept, new words glossed from matthewreagan's JSON by README step 2) | At 2,000 words 11 of the 47 sinks reach the (new, larger) Core; at all 5,492 eligible words of the list, 34 of 47 do and 13 stay sinks (data, user, non, poker, whether, ny, went, involved, mary, disclaimer, tom, sub, path). whether = "Which ; which one ;" is stopwords only, so no widening can ever give it an arrow under this parse. Mostly true, not always. (fixed: widen-all-arrows-reappear) - WRONG | [html-three-obj] The 2016 result carves it into three nested objects, each of which you can compute: | https://arxiv.org/abs/1411.0129 | The three named are Kernel, Core, Minimum Grounding Set, and the third is not nested in the second: the MinSet is "about 1% of the dictionary, 15% of the Kernel, and half-Core, half-Satellite" (abstract); "each MinSet is part-Core and part Satellite" (full text). Only Kernel contains Core and Kernel contains MinSet. (fixed: minset-nested-in-core) - MINOR | [html-published-figures] on curated learner dictionaries the published estimate of the minimum grounding set is about 1% of the dictionary; the Kernel there is roughly 10% and the Core roughly 75% of the Kernel | https://arxiv.org/pdf/1411.0129 | Table 1, the two learner dictionaries: Kernel 9% (Cambridge) and 8% (Longman), Core 2009/2241 = 90% and 1786/2326 = 77% of the Kernel, MinSets 1% and 1%. The 10/75/1 are the abstract's figures over all four dictionaries; Cambridge's Core is 90%, not about 75%. (fixed: learner-figures-misattributed) - MINOR | [html-parasitic] Meaning, he argued, is parasitic: the words in your head feel meaningful only because you ground some of them in sensorimotor experience | https://arxiv.org/abs/cs/9906002 | In 1990 "parasitic" is said of a formal symbol system's interpretation: "made intrinsic to the system, rather than just parasitic on the meanings in our heads"; the meanings in our heads are the intrinsic case. The page's sense matches Harnad 2024 (arXiv 2402.02243: "the parasitism of indirect verbal grounding on direct sensorimotor grounding"), not the 1990 paper it is attached to. (fixed: parasitic-sense-1990) - MINOR | [html-lexicon-desc] The shipped lexicon is the 1,200 most-frequent single-word content headwords of Webster's Revised Unabridged Dictionary (1913, public domain) | research/the-dictionary-that-eats-itself/README.md | Rebuilding V by README step 1 from google-10000-english gives 1,192 of the 1,200; an undisclosed filter dropped sex, god, nude, ass, anal, pussy, rape, incest, so the list runs 8 words deeper (to "path", rank 1789). The text used is Gutenberg #29765 (2009) via matthewreagan's JSON, whose README warns of "known errors and omissions" (e.g. digital = "Of or performance to the fingers"). (fixed: undisclosed-eight-word-filter) - WRONG | [html-gutenberg] Project Gutenberg #661–673 / GCIDE, public domain. | https://www.gutenberg.org/ebooks/673 | #673 says the dictionary is "#660 (A and B), #661 (C) ... #670 (X, Y, and Z)". The range leaves out #660 and takes in two other books: #671 is "Phil, the Fiddler" (Horatio Alger), #672 "The Secret Guide to Computers" (Russ Walter). The lexicon itself came from #29765. (fixed: gutenberg-661-673-range) - WRONG | [html-gutenberg-671] Project Gutenberg #661–673 / GCIDE, public domain. | https://www.gutenberg.org/ebooks/671 | Title "Phil, the Fiddler by Jr. Horatio Alger". Same error as html-gutenberg; one fix. (fixed: gutenberg-661-673-range) - MINOR | [html-gutenberg-copyright] Project Gutenberg #661–673 / GCIDE, public domain. | https://gcide.gnu.org.ua/license | The 1913 text is public domain, and MICRA's notice in pg661.txt says its etext "may also be considered as public domain", but Gutenberg's catalogue marks #660-#670 "Copyrighted", and GCIDE is "under the terms of the GNU General Public License ... version 3". "GCIDE, public domain" reads as a licence claim GCIDE does not make. (fixed: gcide-public-domain) - MINOR | [quote-define-nothing] the lookup-direction image of the paper's "remove words that define nothing." | https://arxiv.org/pdf/1411.0129 | Quotation marks around words the paper does not use: 0 hits for "define nothing" / "defines nothing" in the full text, while the positive control "do not define any further" hits once ("Recursively removing all words that are reachable by definition but that do not define any further words"). The sense is right. Also in README step 4 and the verifier header. (fixed: define-nothing-not-quoted) - MINOR | (added) Run it on this dictionary and almost the entire thing collapses into a single component. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | The component is 60.3% of the 1,200 words ("Core as % of vocabulary 60.3"); 476 words are outside it. "Almost all of the Kernel" (95.9%) would be true. index.html line 247; README opening sentence too. (fixed: core-is-whole-dictionary) - WRONG | (added) On the full 724-word Core the exact minimum is out of reach ... Nobody can hand you the minimum grounding set of English by running an algorithm to completion. | https://arxiv.org/pdf/1411.0129 | The paper the page cites solved larger instances exactly: "because some real dictionary graphs are relatively small and also seem to be structured in a favorable way, our algorithms are able to compute their MinSets"; with CPLEX for Longman and Cambridge, whose Kernels are 2,326 and 2,241 words. NP-hard is a worst-case statement about the problem, not a wall at 724 words. My own run in scratch, about 12 minutes on one core: the Core's exact minimum lies between 142 and 159 (159: a witness checked acyclic on the Core; 142: a cycle-constraint integer program solved to optimality by HiGHS over 1,685 cycles, after the standard DFVS reductions, which force 32 words; stopped before closing the gap). index.html lines 342 and 390. (fixed: exact-minset-out-of-reach) - WRONG | (added) Grounded: none. The six words still form a circle, no word among them can be defined first. | research/the-dictionary-that-eats-itself/lexicon.json | effect and result have no arrow to any of the other five (their glosses, "Execution; performance; realization; operation; ..." and "To leap back; to rebound ...", use none of them), so both can be defined first. The status line repeats it after any grounding that leaves a circle. index.html lines 364 and 997. (fixed: no-word-defined-first) - WRONG | (added) The Kernel ⊃ Core ⊃ Minimum-Grounding-Set decomposition and the ≈10% / ≈75% / ≈1% figures: Vincent-Lamarre ... | https://arxiv.org/abs/1411.0129 | The paper's MinSet is "half-Core, half-Satellite", so it is not inside the Core. Same error as html-three-obj, in the footer (index.html line 474) and README line 98. (fixed: minset-nested-in-core) - MINOR | (added) run greedy on the whole Core / On the full 724-word Core ... you can watch a crude greedy version run | public/strata/the-dictionary-that-eats-itself/index.html | The button's code starts from all 1,200 words (alive = new Set(WORDS)), not the Core: 49 of the 465 words it grounds are outside the Core and 44 of those lie on no cycle at all. index.html line 393; README table "greedy grounding set (full Core) 465". (fixed: greedy-whole-core) - MINOR | (added) true dead-ends reached: N (the hand trail's status line) | public/strata/the-dictionary-that-eats-itself/index.html | The status calls the 47 out-degree-0 words "true dead-ends" while the page argues they are a false floor ("They aren't primitive."). index.html line 756. (fixed: true-dead-ends) - MINOR | (added) not the figures this page computes on Webster's 1913, which are larger because 1913's vocabulary is vast and unrestricted | https://arxiv.org/pdf/1411.0129 | Table 1: the vast, unrestricted Merriam-Webster (248,466 word-meanings) has a Kernel of 12% and a Core of 7977/10955 = 73% of it, the same as the learner dictionaries. What makes this page's shares larger is the slice: the 1,200 most frequent words, with arrows only among themselves. index.html line 402. (fixed: larger-because-vast-vocabulary) - MINOR | (added) Recompute it yourself: node research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs ... Full method and the exact source dictionary: research/the-dictionary-that-eats-itself/ | https://artwaste.land/checks/research/the-dictionary-that-eats-itself/lexicon.json | A stranger cannot: /checks/ serves the verifier (200) but lexicon.json is 404, and the repository is private (github.com/artwasteland/Artificial-Wasteland 404). Run in an empty directory, the published verifier exits 1; with the lexicon cut out of the live page's inline JSON it prints 20/20 PASS. The placard says "Running it needs the repository around it", so this is about the page's own instruction, not a hidden fault. (listed: stranger-cannot-recompute) - CONFIRMED | [dek-live-tarjan] It's a measurable graph fact, recomputed live on 1,200 headwords of Webster's 1913 with Tarjan's algorithm | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | Ran it: 20/20 PASS. The page script runs the same Tarjan on its inlined lexicon, which is JSON-equal to lexicon.json (checked). - CONFIRMED | [dek-fvs-nphard] the smallest set of words you'd have to already know is a minimum feedback vertex set, which is NP-hard to find | https://en.wikipedia.org/wiki/Feedback_vertex_set | "Karp (1972) showed that finding a feedback vertex set of size <= k in directed graphs is NP-complete." The paper: "finding a minimum feedback vertex set in a general digraph is NP-hard (Karp, 1972), which implies that finding MinSets is also NP-hard." - CONFIRMED | [dek-six-word-nonunique] shown exactly on a six-word piece of the real Core, not even unique | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | Recomputed: all six words are in the Core; minimum 2, three solutions {cause,act} {cause,power} {act,make}. - CONFIRMED | [body-three-pairs] three different pairs, each equally irreducible. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | eq("the three minimal grounding sets") passes; recomputed. - CONFIRMED | [body-published-labelled] The published 10% / 75% / 1% figures are labelled as results on other corpora, never as what the page computes. | public/strata/the-dictionary-that-eats-itself/index.html | "(Vincent-Lamarre et al. 2016, on other corpora, not the figures this page computes on Webster's 1913 ...)" and the Assumptions paragraph say so. - CONFIRMED | [body-parse-disclosed] and the parse that turns definitions into edges is disclosed in full. | research/the-dictionary-that-eats-itself/README.md | README step 3 and the page's check box describe it, and the code ships in the page. (The corpus build is not fully disclosed: see html-lexicon-desc.) - OBSERVED | [check-2020] 20/20 PASS. Full method and the exact source dictionary | recreated: node research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs at 87ef051577 | Printed "20/20 PASS", exit 0. Also passes as the published /checks/ copy beside the lexicon extracted from the live page. The page line is static text. - CONFIRMED | [html-computed-on-load] Every figure above is computed when the page loads | public/strata/the-dictionary-that-eats-itself/index.html | The script fills m-nodes..m-stay, p-closure, p-core, p-second, p-stay from the inlined lexicon; static defaults are overwritten. - CONFIRMED | [html-same-code] the same code as the verifier | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | STOP list identical; lemmaCandidates identical after normalising var/const; edgesFrom, computeKernel, tarjan, greedy are the same logic. - CONFIRMED | [html-no-typed-numbers] no numbers are typed in; Tarjan's algorithm runs on the shipped lexicon when the page loads, and prints what it finds. | public/strata/the-dictionary-that-eats-itself/index.html | The metric grid "below" ships as dash placeholders that the script fills. Prose spans carry defaults (739, 724, 3, 98.6%, 755, 47) that the script overwrites. - CONFIRMED | [html-first-sense-30] keep only the first sense trimmed to 30 words | research/the-dictionary-that-eats-itself/lexicon.json | Max 30 whitespace tokens (498 glosses at 30). My rebuild from matthewreagan dictionary_compact.json by the README recipe reproduces 1,046 of 1,200 glosses exactly. "First sense" is the first numbered sense, often obsolete: result = "To leap back; to rebound ..." (Webster marks it [Obs.]). - CONFIRMED | [html-closure-739] The reachable set stops growing at 739 words | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | Recomputed: 739 = all 724 Core words + 15 out-degree-0 words. - CONFIRMED | [html-no-exit] There is no exit. | research/the-dictionary-that-eats-itself/lexicon.json | True of a closure by construction: no arrow leaves the 739. - CONFIRMED | [html-cant-reach-bottom] Try to reach a word that is defined in terms of nothing you have to look up. You can't. | research/the-dictionary-that-eats-itself/lexicon.json | Every one of the 47 sink glosses still contains words a reader would have to look up (e.g. think = "To seem or appear;"), so it holds of the dictionary, though the instrument does show such words with nothing underlined. - CONFIRMED | [html-stay-98] Of every lookup you can make starting from a word inside the Core, 98.6% lands you on another word inside the Core. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | 3,748 of 3,800 = 98.63%. - CONFIRMED | [html-724-3] One component of 724 words; the next-biggest is 3. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | Recomputed: 724, then {saturday, thursday, friday}. - CONFIRMED | [html-tarjan-1972] In 1972 Robert Tarjan gave a way to find every such component in one linear pass. | https://en.wikipedia.org/wiki/Tarjan%27s_strongly_connected_components_algorithm | "It runs in linear time ... named for its inventor, Robert Tarjan." - CONFIRMED | [html-tarjan-1972-date] In 1972 Robert Tarjan gave a way to find every such component in one linear pass. | https://api.crossref.org/works/10.1137/0201010 | SIAM J. Comput. 1(2):146-160, issued June 1972; Wikipedia SCC: "published by Robert Tarjan in 1972, performs a single pass of depth-first search." - CONFIRMED | [html-one-giant-knot] one giant knot, and everything else a scatter of tiny fragments hanging off it | https://arxiv.org/pdf/1411.0129 | Full text: "contain an SCC much larger than all the others" and "The remaining 25% of the Kernel surrounding the Core consists of many tiny strongly connected subsets". The N1 note that the paper does not describe this shape read only the abstract. - CONFIRMED | [html-false-floor-47] 47 of the 1,200 words look grounded | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | 47 recomputed. - CONFIRMED | [html-false-floor-list] data, user, computer, server, photo... | research/the-dictionary-that-eats-itself/lexicon.json | All five are sinks: "See Datum.", "One who uses. Shak.", "One who computes.", "One who serves.", "A contraction of Photograph." - CONFIRMED | [html-holes] They are holes in the sample. | research/the-dictionary-that-eats-itself/lexicon.json | True of four of the five named (datum, compute, serve, photograph are not headwords); user is a stoplist artefact ("uses" lemmatises to the stopword use). See html-widen. - CONFIRMED | [html-kernel-755] What survives is the Kernel: 755 words here | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | 755 recomputed; the Core is inside it. - CONFIRMED | [html-core-path] component where every word has a path to and from every other | research/the-dictionary-that-eats-itself/lexicon.json | Checked all pairs (forward and backward BFS inside the Core from one word reach all 724), not just the verifier's one probe pair. - CONFIRMED | [html-fvs-def] That set is a minimum feedback vertex set: the fewest words whose removal makes the whole tangle acyclic. | https://en.wikipedia.org/wiki/Feedback_vertex_set | "a set of vertices whose removal leaves a graph without cycles ... the size of a smallest FVS." - CONFIRMED | [html-minset-vl] That set is a minimum feedback vertex set | https://arxiv.org/abs/1411.0129 | Abstract: "the smallest set of words that can define all the rest (the graph's Minimum Feedback Vertex Set or MinSet)". - CONFIRMED | [html-nphard-core] First: finding the true minimum on the full Core is NP-hard. | https://arxiv.org/pdf/1411.0129 | True of the general problem (Karp 1972, quoted in the paper). The inference the page draws from it is not: see the added "out of reach" line. - CONFIRMED | [html-not-unique] even where you can compute it, it is not unique. | https://arxiv.org/abs/1411.0129 | "But every dictionary has a huge number of MinSets." - CONFIRMED | [html-six-words] Six words from the Core and every definition-arrow that runs between them | research/the-dictionary-that-eats-itself/lexicon.json | 11 arrows, each target in the source's shipped gloss: cause -> effect, result, power, act, make; act -> power, effect, cause; power -> act, effect; make -> cause. - CONFIRMED | [html-each-pair] each of which, if you already knew it, would let you define the other four without any circle. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | Each pair leaves the other four acyclic. - CONFIRMED | [html-greedy-upper] the real minimum is smaller, and finding it is the NP-hard problem above | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs (recreated in scratch: its greedy set, then each member dropped while the rest stays acyclic) | Dropping redundant members leaves a valid feedback vertex set of 174 Core words (the greedy run removes 465, 416 of them in the Core); an SCC-restricted greedy gives 159, and with one word from each of the five small SCCs a 164-word set leaves the whole 1,200-word graph acyclic (checked). The real minimum is far below 465. - CONFIRMED | [html-harnad-1990] In 1990 Stevan Harnad posed the symbol grounding problem: ... like trying to learn Chinese from a Chinese-only dictionary | https://arxiv.org/abs/cs/9906002 | "The problem is analogous to trying to learn Chinese from a Chinese/Chinese dictionary alone." Physica D 42, June 1990. - CONFIRMED | [html-26-later] Twenty-six years later, Harnad was a co-author of the paper that measured the trap he'd described | https://api.crossref.org/works/10.1111/tops.12211 | Topics in Cognitive Science 8(3), July 2016; Harnad is the sixth author. 1990 + 26 = 2016. - CONFIRMED | [html-llm-view] The distributional / large-language-model view holds the opposite | https://arxiv.org/abs/2208.02957 | Piantadosi & Hill 2022: "we argue that they likely capture important aspects of meaning ... in which meaning arises from conceptual role"; Harnad 2024 records "some of us ... conclude that ChatGPT actually understands." The view exists as stated. - CONFIRMED | [html-harnad-still-symbols] Harnad says that's still just symbols chasing symbols. | https://arxiv.org/abs/2402.02243 | Not a quotation; a fair paraphrase of "It is not true that it understands" and what it lacks, "direct sensorimotor grounding". - CONFIRMED | [html-liar-quote] the liar's "this sentence is false." | https://en.wikipedia.org/wiki/Liar_paradox | "Assume that "this sentence is false" is true ..." - CONFIRMED | [html-longman-2000] writes every definition using a fixed list of about 2,000 words | https://en.wikipedia.org/wiki/Defining_vocabulary | "Longman Dictionary of Contemporary English: approximately 2,000 words" (the LDOCE article says 2,200; the 2016 paper says the principle "was not faithfully followed"). Fine as a statement of Longman's policy. - CONFIRMED | [html-defining-vocab-quote] its "defining vocabulary" | https://en.wikipedia.org/wiki/Defining_vocabulary | The standard term. - CONFIRMED | [html-small-capitals] prints anything outside it in small capitals | https://lexikos.journals.ac.za/pub/article/view/1013/530 | Tool said UNREADABLE (PDF); I read it with pdftotext. Xu 2012, note 4, on LDOCE5: "almost all learner's dictionaries occasionally resort to some words outside a CDV list, and those words are printed in small capitals." - CONFIRMED | [html-edge-direction] The paper draws arrows from defining words to the words they define | https://arxiv.org/abs/1411.0129 | "directed graphs with links from defining words to defined words". - CONFIRMED | [html-reversal] A strongly-connected component is identical under reversal, so the Core is direction-agnostic. | research/the-dictionary-that-eats-itself/lexicon.json | Tarjan on the reversed graph: 724, 3, 2, 2, 2. - CONFIRMED | [html-kernel-direction] here it removes, recursively, every word with no incoming lookup | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | computeKernel deletes in-degree-0 words to a fixed point. - CONFIRMED | [html-grow-shrink] A stricter lemmatiser or a fuller definition would grow the Core; a harsher stopword list would shrink it. | research/the-dictionary-that-eats-itself/lexicon.json (recreated in scratch: Core size under parse variants) | No stemmer 653; my rebuild of the glosses at 30 tokens 743, at 60 tokens 840, at 100 tokens 852; 21 extra common stopwords 685; shorter trims 466 (15) and 316 (10). Every direction as stated. - CONFIRMED | [html-survives] a large SCC survives every reasonable choice | research/the-dictionary-that-eats-itself/lexicon.json (recreated in scratch: same parse variants) | Smallest Core seen 316 (10-token gloss), runner-up never above 4. "Every" cannot be tested in full. - CONFIRMED | [html-nonunique-exhibited] Non-uniqueness is exhibited, not merely claimed: the verifier enumerates all three minimum pairs. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | It brute-forces every subset and asserts the three. - CONFIRMED | [html-six-bruteforce] Exact only on the six-word subgraph, by brute force over all vertex subsets. | research/the-dictionary-that-eats-itself/verify-the-dictionary-that-eats-itself.mjs | minFeedbackVertexSet enumerates subsets by size. - CONFIRMED | [html-gcide] The dictionary: Webster's Revised Unabridged (1913) | https://gcide.gnu.org.ua/ | "derived from Webster's Revised Unabridged Dictionary Version published 1913 by the C. & G. Merriam Co." - CONFIRMED | [cite-vl2016] Vincent-Lamarre, Blondin Massé, Lopes, Lord, Marcotte & Harnad, "The Latent Structure of Dictionaries," Topics in Cognitive Science 8(3):625–659 (2016). | https://api.crossref.org/works/10.1111/tops.12211 | Title, six authors in order, 8(3) 625-659, 2016. The tool's DOI MISMATCH is Crossref's "Vincent‐Lamarre" with U+2010. - CONFIRMED | [cite-tarjan-title] R. E. Tarjan, "Depth-first search and linear graph algorithms," SIAM J. Comput. 1(2):146–160 (1972). | https://api.crossref.org/works/10.1137/0201010 | Matches. - CONFIRMED | [cite-harnad-title] Harnad, "The Symbol Grounding Problem," Physica D 42:335–346 (1990). | https://api.crossref.org/works/10.1016/0167-2789(90)90087-6 | Matches. - OBSERVED | [html-placard] Measured: with every file this layer ships emptied, it still passes. | git show 87ef051577:research/verifier-independence/depends-1.json | Its row: tier C0, exitCode 0, "depends": false. Consistent with the code: the verifier reads only lexicon.json. (census-pairs.tsv, N1's source, is the classification; depends-1.json is the measurement.) - CONFIRMED | [rel-godel] Gödel's sentence is a deliberate logical construction, a self-reference built on purpose. | src/content/strata/godel-incompleteness.md | "He built, inside arithmetic, a sentence that asserts its own unprovability". - CONFIRMED | [rel-nofix] which that page circles from Tarski and the map/territory gap | src/content/strata/no-fix-from-inside.md | It has Tarski 1936 and "the map is not the territory" and never names grounding or Harnad, which is what "circles" says. - CONFIRMED | [rel-map] there a self-study citation network with communities and hubs | src/content/strata/the-map-no-one-drew.md | "a citation network ... detected clusters ... concentrates into hubs", re-run live in the browser. - CONFIRMED | [rel-shakespeare] There, the words Shakespeare knew but never wrote, estimated by unseen-species statistics | src/content/strata/how-many-words-did-shakespeare-know.md | "Efron & Thisted's 1976 unseen-species estimator". - CONFIRMED | (added) Harnad, "The Symbol Grounding Problem," linked to sciencedirect.com/science/article/abs/pii/0167278990900876 | https://doi.org/10.1016/0167-2789(90)90087-6 | The DOI resolves to linkinghub.elsevier.com/retrieve/pii/0167278990900876, the same PII (ScienceDirect itself returns 403 to curl). - CONFIRMED | (added) word → sign, word → sound, word → idea | research/the-dictionary-that-eats-itself/lexicon.json | adj(word) = sign, idea, sound, human, voice, custom. (sound here is the adjective "Whole; unbroken".) - CONFIRMED | (added) Nearby layers: "The graph this page computes on, rebuilt by slicing that layer's own parse code out of that layer's own verifier and running it over that layer's own lexicon" | research/something-has-to-stop-the-loop/build-data.mjs | It reads this layer's verifier and lexicon.json and slices the parse block out (sliceOut(dictVerifierText, PARSE_START, PARSE_END ...)). ## Findings - [fixed] any-word-reaches-core (K1, MINOR): the dek, meta description and og description said any word leads into the 724-word knot; 48 words never reach it. Now 'Pick almost any word' / 'Follow almost any word's definition'. - [fixed] harnad-26-year-old (K2, WRONG): the dek called Harnad's question '26-year-old'; Physica D 42 was issued June 1990, 36 years before the layer's date. Now 'Harnad's 1990 open question'. Covered by the Corrected 2026-10-18 line. - [fixed] core-is-whole-dictionary (K3, MINOR): the markdown body said the whole dictionary collapses into the Core (it is 60.3%). Now 'most of the dictionary'. - [fixed] every-number-verified (K4, MINOR): 'Every number on screen is recomputed ... by' the verifier (markdown body), 'Every number shown is recomputed by' (footer), README 'recomputes every number shown' and the verifier header: now the headline figures/numbers, recomputed and asserted; the README names what it does not assert. - [fixed] proper-names-dropped (K5, MINOR): 'every proper name, archaic term and rare word is dropped' now reads 'so any word outside it (archaic, rare or a name) is dropped', which is what the parse does (london, paris, john are headwords and survive). - [fixed] handful-of-clicks (K6, MINOR): 'Within a handful of clicks the trail turns red' now reads 'Keep clicking and, unless you stop on one of the false floors below, the trail turns red'. - [fixed] paper-dictionaries-omits-webster (K7, MINOR): 'in learner dictionaries and WordNet' now 'in two learner dictionaries, Merriam-Webster and WordNet' (page and README honesty ledger). - [fixed] defining-words-outside-slice (K8, WRONG): the false-floor box said the sinks' real defining words (digital, machine, network) fell outside the slice; all three are headwords. It now quotes the 1913 glosses and names the words they lean on: datum, compute, serve, contraction, photograph (outside the slice) and use (a stopword). Covered by the Corrected 2026-10-18 line. - [fixed] widen-all-arrows-reappear (K9, MINOR): 'Widen the vocabulary and the arrows reappear, pointing right back into the Core' now 'most of the arrows reappear, pointing back into the Core' (34 of 47 along the whole frequency list). - [fixed] minset-nested-in-core (K10, WRONG): 'three nested objects' now 'three objects ... the Kernel, and inside it both the Core and the Minimum Grounding Set (which the paper finds is part Core, part in the small knots around it)'. Covered by the Corrected 2026-10-18 line. - [fixed] learner-figures-misattributed (K11, MINOR): the 10% / 75% / 1% figures were attributed to curated learner dictionaries; they are the abstract's figures over the paper's four dictionaries (Cambridge's Core is 90% of its Kernel). Now 'across the four dictionaries the paper analysed (two learner dictionaries, Merriam-Webster and WordNet)'. - [fixed] parasitic-sense-1990 (K12, MINOR): 'Meaning, he argued, is parasitic: the words in your head feel meaningful only because...' now puts 'parasitic' where the 1990 paper does: such a system's meaning is only parasitic on the meanings in our heads, and its symbols would have to be grounded bottom-up in something that is not symbols. - [fixed] undisclosed-eight-word-filter (K13, MINOR): the corpus description (page Assumptions and README) now discloses the eight words left out of the 1,200 (sex, god, nude, ass, anal, pussy, rape, incest); README step 1 says the recipe did not record that filter and the list runs to 'path'. - [fixed] gutenberg-661-673-range (K14, WRONG): the sources line cited 'Project Gutenberg #661–673'; the 1913 dictionary is #660 to #670, with #673 an XML file of the whole, and #671/#672 are other books. Now '#660 to #670, and the 2009 compilation #29765 that this page parses (via the matthewreagan/WebstersEnglishDictionary JSON)'; same fix in README and the verifier header. Covered by the Corrected 2026-10-18 line. - [fixed] gutenberg-661-673-range (K15, WRONG): same false range as K14 (#671 is Horatio Alger's 'Phil, the Fiddler'); answered by the same edit. - [fixed] gcide-public-domain (K16, MINOR): '/ GCIDE, public domain' removed from the sources line; it now says the 1913 text is public domain, and the README says GCIDE is distributed under the GNU GPL. - [fixed] define-nothing-not-quoted (K17, MINOR): the paraphrase in quotation marks, 'remove words that define nothing', is replaced by the paper's own words, 'Recursively removing all words that are reachable by definition but that do not define any further words' (page, README step 4, verifier header). - [fixed] core-is-whole-dictionary (K18, MINOR): 'almost the entire thing collapses into a single component' (page) and 'almost the whole thing collapses' (README) now 'most of it collapses'. - [fixed] exact-minset-out-of-reach (K19, WRONG): 'finding the true minimum on the full Core is NP-hard. Nobody can hand you the minimum grounding set of English by running an algorithm to completion' and 'On the full 724-word Core the exact minimum is out of reach' now say the problem is NP-hard, that the 2016 paper computed exact MinSets for its two learner dictionaries with an integer-programming solver and near-optimal ones for the two big ones, and that this page does not compute it on the full Core; the Assumptions line likewise. Covered by the Corrected 2026-10-18 line. - [fixed] no-word-defined-first (K20, WRONG): the six-word instrument's static line and status code said no word among those left can be defined first; effect and result can. Both now say the words left still contain a circle, so they cannot all be put in a define-first order. Covered by the Corrected 2026-10-18 line. - [fixed] minset-nested-in-core (K21, WRONG): the footer's 'Kernel ⊃ Core ⊃ Minimum-Grounding-Set decomposition' now 'Kernel ⊃ Core and Kernel ⊃ Minimum-Grounding-Set decomposition (each MinSet part Core, part Satellite)'; README primary sources likewise. Covered by the Corrected 2026-10-18 line. - [fixed] greedy-whole-core (K22, MINOR): the greedy button and sentence said it runs on the whole Core; the code starts from all 1,200 words. Button now 'run greedy on all 1,200 words', the sentence and Assumptions line say all 1,200 words, README table and the verifier comment likewise. - [fixed] true-dead-ends (K23, MINOR): the hand trail's status line 'true dead-ends reached' now 'apparent dead-ends reached', matching the false-floor box. - [fixed] larger-because-vast-vocabulary (K24, MINOR): 'which are larger because 1913's vocabulary is vast and unrestricted' now says they are larger and that this is a 1,200-word frequency slice with arrows only among itself, while the paper's full Merriam-Webster comes out close to the learner dictionaries (README ledger likewise). - [declined] stranger-cannot-recompute (K25, MINOR): the 'Recompute it yourself' line cannot be followed by a stranger (lexicon.json is not served under /checks/ and the repository is private); the placard already says running it needs the repository. Not a few words on this page: the fix is publishing the lexicon beside the verifier, shared infrastructure. Left as is.