onto — the world model

A model of abstractions kept separate from the narrative: TBox versus ABox from Smart annotator — layers, context, memory, section "World model — separate from the narrative". The guiding idea: build the world model before the mathematics; it still has to be taught, much like in fairy tales, that an apple is a fruit.

• World model: concepts, their hierarchies, parts, typical actions and roles, default properties with exceptions. Shared by all texts.
• Narrative (world) — the individuals of a single text. Crow1 refers to the concept "crow" rather than copying it.

The core does not depend on en: lemmas and parts of speech come from OEWN itself. Only the hs module (goods classifier) takes parses from en and exact numbers from math.

In brief (2026-09-26)

• Graph: 133 213 concepts — 120 564 OEWN 2025+ synsets, 1 221 frames and 11 428 FrameNet 1.7 roles. 154 608 words, 212 659 senses, 13 390 lexical units. 138 125 typed edges.
• Seeds:
  • 66 properties with a source on every row: 61 on the real layer, 3 in genres, 1 in a collection, 1 in a tale;
  • 2 extra edges — the Nixon diamond;
  • 26 layers: the real world, 3 genres, 7 collections, 15 tales.
• Frozen set, 74 questions, recorded and hashed before the query code: 72/74, 0 wrong, 2 "unknown" with the core and WSD; the blind round on the first query code (first sense only) — 70/74 (95%), 2 wrong, both from OEWN sense order.
• Negative controls: 10 of 10 green on real data and red on an injected defect. A shuffled hierarchy drops IsA from 95% to 53%, i.e. to the level of the "always yes" baseline.
• Bridge to the narrative world:
  • 302 of 305 fable and tale classes found (99%);
  • 14 fable questions via inheritance: blind round — 12/14, second — 14/14.
• Speed:
  • build from XML — 1.7–3.1 s;
  • the 19.1 MB snapshot loads in 10–20 ms (median 15); together with seeds and gates — also ~13–15 ms;
  • the graph takes 20 MB of arrays;
  • a query with proof — ~0.01 ms.
• Wikidata:
  • lexemes — statistics and a trial alignment: 13 067 synsets got a Q; a manual sample of "by gloss" matches — 17/20;
  • full dump — one pass of 2 h 05 min into wikidata.duckdb; a walker that collects seed candidates for the core.

How it works

Model (model.rs)

• A concept is a u32. Struct of arrays: source (Src: OEWN synset, frame, role, Wikidata Q), key (synset offset, frame ID), part of speech (Pos), lexicographer file (index in LEXFILES), ILI, gloss, member lemmas.
• A word is a u32. Sorted dictionary, binary search. A word's senses are in OEWN frequency order, i.e. the order of <Sense> in the entry.
• Relations — enum Rel, direction as it reads:
  • from OEWN: IsA, InstanceOf, PartOf (part → whole), MemberOf, MadeOf (whole → substance), Entails, Causes;
  • from FrameNet: HasRole, Inherits, Uses, SubframeOf, PerspectiveOn, Precedes, CausativeOf, InchoativeOf.
• Edges — CSR in both directions. A lexical unit is a word with a part of speech → frame. Within a word, units with more annotated sentences come first.
• Properties — enum Prop: Living Animate Human CanSpeak CanFly CanSwim CanWalk Edible LaysEggs Fast Exists Pacifist; character for fables: Cunning Vain Greedy Industrious Wise. Strength is Default or Strict; the layer is a LayerKind: real world, genre, collection, tale.
• Strings only at input and output: XML, TSV, query, answer.

Build (build.rs, zip.rs) and snapshot (store.rs)

• OEWN is taken from english-wordnet-2025-plus.xml.gz.
  • The base 2025 edition has no instance_hypernym at all: proper names and taxa were moved out. It has only 723 mero_member.
  • -plus is a superset: it contains all 107 519 base synsets, with 13 045 added. Such concepts are marked with a "plus only" bit.
• FrameNet is read directly from the zip by our own minimal reader (central directory + deflate): the zip crate is not available offline, and there is no need to unpack to disk. From the archive we take frame/*.xml — roles with coreType, lexical units and frame relations.
• Build gates (violation — refusal):
  • OEWN inverse pairs are consistent: hypernym/hyponym, holo_*/mero_*, entails/is_entailed_by and others — 0 mismatches;
  • the edge endpoint exists;
  • a frame relation points to an existing frame;
  • an IsA cycle — refusal.
• Snapshot — our own binary format, like en/src/store.rs but without serde: the graph already lies in flat arrays.
  • Header: magic MOVA-ONT, version, sizes of the enums in the code, length and checksum.
  • A corrupted byte, a truncated file, extra bytes or foreign enums give an error, not a silently different graph.

Seeds (seeds/) and layers

• layers.tsv — the layer tree. Root real; genres fable, fairy, detective; collections aesop, grimm, perrault, japanese, scottish, nights, holmes; tales tale:2451 etc. — following the world command files.
• defaults.tsv — property seeds: id, layer, lemma, synset, property, value, strength, source. Every row has a source:
  • gloss — the OEWN synset gloss says so directly: "flightless", "egg-laying", "swift";
  • manual — general knowledge, written down by hand;
  • a paper or text — by name: Reiter 1980; Reiter & Criscuolo 1981; Aesop, Perry 226; FairytaleQA ^s65–66.
• relations.tsv — extra IsA/InstanceOf. Currently only Nixon — a Quaker and a Republican.
• Seed gates:
  • unknown layer;
  • lemma not a member of the synset — catches a typo in the id;
  • duplicate id;
  • contradiction on the same concept and layer;
  • a descendant against a Strict ancestor on the same layer;
  • an IsA cycle through a new edge.

Inheritance (kb.rs)

A query on layer L sees only the seeds of L and its ancestors. An exception applies on its own layer and below: the fable CanSpeak does not make real crows speakers.

1. Layer: the most specific layer where at least one visible seed concerns the concept or its ancestor decides. A fable overrides the real world only where it has said something of its own.
2. Concept: within that layer the most specific concept wins. An ancestor's seed is overridden by a descendant's seed, and the proof shows it.
3. Nixon diamond: two incomparable seeds with different values give "undetermined" with both proofs, not a random choice.
4. Strength: Strict is not overridden on its own layer. A lower layer may explicitly override anything — that is what layers are for.

Word sense selection (query.rs)

Without a sentence, the context comes from the query itself (for isa and haspart the candidates are then ranked by the WSD score, section "Word sense selection by context"):

• isa X Y: the sense of X in the lexicographer file of Y's main sense plus the first sense of X; Y — any of its senses. "whale" with "fish" is the animal, not the first sense "a very large person".
• Property — selectional preference. Bodily abilities (CanFly, CanSwim, CanWalk, LaysEggs, Fast) look for an animal sense: the first sense of "ostrich" is a person who hides their head, and "emu" is a unit of measure. Edible looks for a food sense. Other properties take the first sense.
• The bridge to the world takes context from the object type: type=animal → noun.animal.
• An explicit sense is a synset id instead of a word: onto can oewn-01913388-n walk.

Every answer says which sense was taken, why, and at what frequency rank.

Queries

An unknown word gives "unknown", not a guess.

Examples (examples.txt):

$ onto can crow speak --layer fable
yes
  sense: crow → oewn-01581669-n «crow» (1st of 6, noun.animal; first sense) — «black birds having a raucous call»
  inherited: «crow» —IsA→ «corvine bird» —IsA→ «oscine» —IsA→ «passerine» —IsA→ «bird» —IsA→ «vertebrate» —IsA→ «chordate» —IsA→ «animal»
  seed d62 [fable — genre, Default]: oewn-00015568-n «animal» CanSpeak=yes — source: fable genre: animals talk (Aesop; …)
  exception of layer fable (applies only on it and below)
$ onto can crow speak          → no: seed d10 [real] animal CanSpeak=no
$ onto can penguin fly         → no: seed d14 penguin CanFly=no (gloss: «flightless»); overrides d13 bird CanFly=yes and d26 object CanFly=no
$ onto isa whale fish          → no: whale (2nd sense, noun.animal) ⊂ cetacean ⊂ aquatic mammal ⊂ placental ⊂ mammal ⊂ vertebrate;
                                  fish ⊂ aquatic vertebrate ⊂ vertebrate; diverge under «vertebrate»
$ onto frame buy               → yes: Commerce_buy ← buy.v (LU 2966); Core: Buyer, Goods; other: Money, Seller, …; Inherits Getting
$ onto prop nixon pacifist     → undetermined: d60 Quaker Pacifist=yes ⇐ Nixon —InstanceOf[r01]→ Friend;
                                  d61 Republican Pacifist=no ⇐ Nixon —InstanceOf[r02]→ Republican
$ onto isa blorft animal       → unknown: the word «blorft» is not in the world model (OEWN 2025+) — not guessing

Bridge to the narrative world (bridge.rs, read-only)

• Input: 18 world command files: world/data/cmds-*.txt and cmds-*.txt. Content duplicates are read once — 16 remain.
• Entity class → concept. Order of attempts:
  1. as is;
  2. spelling variant;
  3. head word.
• Sense — from the lexicographer file of the entity type.
• Questions in English (Can the crow fly?, Is the meat food?, Does the crow have a beak?): the entity is taken from the tale's world, and the query runs on that tale's layer.
• The bridge does not modify world/ files.

Sources and licenses

The papers Reiter 1980 ("A logic for default reasoning") and Reiter & Criscuolo 1981 (Nixon diamond) are only sources of ideas, not of data.

Numbers

Frozen set (data/frozen.tsv)

• 74 questions: IsA with traps, properties with exceptions, layers, parts, frames and roles, verbs.
• Frozen at 15:43Z before the query code: sha256 1affbff3…, commit 1072558b. The hash is checked before every run.
• The gold standard is common-sense truth, not what the data says.

With the core and WSD (section "Check"): 72/74, 0 wrong, 2 "unknown".

Wrong answers of the first-sense rounds (both answered right with WSD):

• isa dolphin fish → "yes". The first sense of "dolphin" in OEWN is dolphinfish (mahi-mahi), which really is a fish. The question meant the mammal.
• isa oak tree → "no", only in the blind round. The first sense of "oak" is the wood, and OEWN puts it in noun.plant.
  • The second round takes any sense of X in the category's lexicographer file, not only the first. The rule is general, but it was introduced after the blind run.

"Unknown":

• haspart crow feather: in OEWN 2025 "feather" is a new synset 89570581-n, not linked as a part of a bird. "plume" is recorded as the bird's part.
• haspart fish wing: the gold answer is "no", but we never say "no" for parts.

Reports: frozen-blind.md and frozen-second.md — every question with its proof.

Negative controls (onto controls)

Bridge to the narrative world

• Entities: 587, of which 434 with a class, 153 with a type only.
• Classes: 305 distinct, 302 found (99%):
  • as is — 278 (91%);
  • spelling variant — 11;
  • by head word — 13: demon_chief → chief, love_potion → potion etc.;
  • not found — 3: bonze, children (irregular plural), vizir (spelling variant of vizier).
• Type context gave a sense in its own lexfile for 217 of the 243 classes that have context.
• Fable questions: 14 (data/fable-questions.tsv, hashed before the bridge code).
  • Blind round — 12/14. "Is the meat food?" was parsed as the property "food" — a parsing defect. "Can the jellyfish walk?": the first animal sense of "jellyfish" is the Portuguese man-of-war, while the tale's seed sits on the medusa.
  • Second, NON-blind round — 14/14. Fixes:
    • a word without an article — IsA, unless it is a property;
    • "one sense per text": if a seed of this tale's layers sits on another sense of the same word in the same lexfile, take that one, and the proof says so.

Speed (onto bench)

• Build: 1.7–3.1 s — OEWN 0.8–1.5 s, FrameNet 0.2–0.5 s. Full cycle check — 0.01–0.03 s.
• Snapshot 19.1 MB, load 15 ms (median of 5; 10–20 ms). Together with seeds and gates — ~13–15 ms: there are few seeds, and the gates are cheap.
• Memory: 20 MB of graph arrays.
• Query with proof: ~0.01 ms. The whole set of 74 questions — ~1 ms.

Wikidata

Lexemes (CC0, onto wd-lexemes, 213 s on one core)

• All languages: 1 675 319 lexemes, 934 793 senses. With P5137 — 334 733, with P8814 (WordNet 3.1 synset) — only 5.
• English: 75 278 lexemes, 60 293 senses. With P5137 — 33 387 (30 231 distinct Q).

Trial alignment of lexemes with OEWN (onto wd-map)

English senses with P5137 → OEWN synsets. P8814 gives 1; other methods:

• 13 067 synsets got a Q.
• Manual check of 20 "by gloss" matches (every N-th): 17/20. Errors occur when the needed sense is missing from OEWN or is different: graph as a mathematical structure, a maritime fairway, etymology as an account rather than the science.
• Keyword traps:
  • apple Q89 ("a fruit of a tree of the genus Malus") → apple tree: the words "tree" and "genus" pulled it over;
  • tortoise → a lexeme for a Han dynasty sculpture: "unambiguous" trusts a lexeme's single sense.
• bird Q5113, cheese Q10943, meat Q10990, penguin, mouse, hare, genie Q3465 — correct.

Full dump: probe

Measurement (onto wd-probe, nice, 1 core, read-only, nothing written to the geography project's folder):

• first 1.07 GB (1.04% of the file) → 11.1 GB of JSON (×10.4);
• 697 311 entities, with P31 — 660 538, with P279 — 33 595 (4.8%), with an English label — 615 835;
• 570 s, of which bzip2 — 548 s (96%), JSON parsing — 22 s;
• 1.88 MB/s compressed → a full pass on one core ≈ 15 h. The bottleneck is bzip2, so the full pass runs lbzip2 with several threads (below).

Wikidata: pass and database (2026-09-26)

Wikidata is loaded into DuckDB; it also carries translations — how an article, or an apple, is rendered in different languages. One streaming pass over the full dump → derived TSV (zstd) wikidata-derived → a separate database wikidata.duckdb (the main mova.duckdb — via ATTACH … (READ_ONLY)). Runs and reports — wikidata-2026-09-26.

How: dump copier (counts compressed bytes) → lbzip2 -dc -n 8 → 8 MiB chunks → 2 parser threads (Cow, labels first as RawValue) → each writes its own <table>-w<N>.tsv.zst; 10 threads, nice -n 10, a progress log every minute. The bottleneck is bzip2. The budget was estimated with the wd-sample sample: bzip2 blocks from an arbitrary bit offset, 48 chunks spread evenly over the file (the first GB of the dump are the most popular items, so a pilot on them overestimates).

Pass: 2 h 05 min, 13.8 MB/s compressed, 1.84 TB of JSON. 121 710 188 entities (121 696 284 items, 13 904 properties), 0 unparsed, [ and ] in place — exactly as many as the geography project counted on 17.09 (the wd-probe measurement gave ~67 M — a poor extrapolation from the heavy start of the file).

Macros: wd_label(q, lang) (lang → mul → en), wd_ancestors(q), wd_descendants(q), wd_isa(x, y).

Decisions and why:

• Wiki service pages (categories, templates, disambiguations, modules, portals — 8.3 M) — without labels in all languages and without sitelinks: they are about the wiki, not the world; in the sample they are 15% of labels and 36% of article titles. The en label stays in items.
• Scholarly articles (46 M) — in the database without titles and descriptions: the world model and translation don't need them, and with them the database is 16.7 GB (budget 15). The texts remain in the derived files.
• Label languages were not restricted: with these two decisions all 637 fit.
• Closure — 613.5 M pairs over 8.46 M nodes (72.5 on average, max depth 43) — exceeds the 400 M limit, so only nodes with a Wikipedia article are materialized (271.7 M); the rest — via the recursive wd_ancestors.

Sizes: derived files 6.6 GB (budget 10), database 13.3 GB (budget 15). Loading — ~3 min; large tables are not sorted (a full sort of 121 M wide rows — Out of Memory at 20 GB).

Check queries (sql/wikidata-checks.sql):

• apple Q89 ⊂ fruit (Q3314483, step 1; Q1364, step 2) ⊂ food; uk label «yabluko» (aliases «yabluka», «yabko»); articles: Apple / Yabluko / Tafelapfel / Pomme / リンゴ.
• whale Q1865281 ⊂ Mammalia — yes; but ⊂ fish — also yes: cladistics, Cetacea —P171 ×24→ Osteichthyes (Q27207) —P279→ fish. For the everyday "is a whale a fish", the walker's folk rule stops P171 at the class rank (section "A whale is not a fish").
• Corvus corone (Q26198) — a taxon, 42 common names in 35 languages (uk «vorona chorna»).
• P4733 "sound" — only 95 statements on 84 items; sounds are not collected.

P279/P171 cycles: 7 components, 14 nodes (cosmology ⊂ astrophysics ⊂ … ⊂ cosmology; corruption ⇄ abuse of power; two P171 loops). The gate lists them; the edges are removed from the closure and recorded; after breaking — a DAG. Negative control: edge fruit ⊂ apple — the gate goes red, a cycle of 7 nodes.

Synset ↔ Q (wd-align): 912 382 candidates from lexemes (P5137), P5063 ILI, P8814 and English labels; 53 392 synsets accepted (62.9% of nouns), 46 405 distinct Q. Silver — direct P5063/P8814 versus choosing from lexemes and labels only: 77.9% (21 231 of 27 239). Shuffled labels: 7 965 accepted — a 6.7× drop, while the pre-set threshold was 10×: support via general ancestors (up to 5% of nouns) lets random pairs through. This loose rule is the synset_q table of the database; the walker and the core use the strict rule (section "Strict synset ↔ Q alignment").

How to update: new dump → wd-extract (~2 h) → wd-lex (48 s) → wd-senses + wd-map → wd-closure → wd-align → sql/wikidata-load.sql and sql/wikidata-closure.sql → sql/wikidata-checks.sql. Commands — section "Commands".

License: Wikidata — CC0 1.0, both the full dump and lexemes: commercial use without restrictions, attribution not required, but the wikidata row is recorded in en/data/train-licenses.tsv. The dump lives in the geography project — only read, nothing written to its folder.

Wikidata walker (2026-09-26)

A small walker goes through Wikidata and collects seed candidates for the core. Sounds (P4733) are not collected. Numbers are kept as distributions, not as one specific animal: a lion's mass comes from a normal distribution. Runs — onto2-2026-09-26; database and derived files — read-only.

A whale is not a fish (src/wdwalk.rs)

Wikidata cladistics leads the whale into fish: whale —P279→ Cetacea —P171×16→ mammal [class] —P171×22→ Sarcopterygii [class] → Osteichthyes [superclass] —P279→ fish. Folk "is a kind of":

• a P171 chain does not climb from a taxon of rank "class" (P105 = Q37517) or higher (superclass, phylum, kingdom…);
• exception — one step class → superclass: Actinopterygii → Osteichthyes —P279→ fish, otherwise a pike is not a fish;
• bypassing clades (P171 past a class, to a higher class) are discarded: from the P171 region only nodes from which the lowest class is reachable remain;
• P279 — unrestricted; shortest distance by relaxation (whale —P279→ Cetacea —P279→ mammal — 2 steps, although via the P171 chain it's 17).

• Crow: Q388450 "crow" is a common name (P31 Q55983715) with only P279 omnivore; there is no path to birds in Wikidata. For animals without a class among their folk ancestors the walker adds an anchor taxon — the best taxon among alignment candidates with ancestor support (genus better than species): Corvus Q43365.
• Cycle gate: graph of 8.46 M nodes, 9.22 M edges — acyclic; an injected fruit ⊂ apple goes red.
• Tests on a synthetic graph: bypassing clade, grade, a shark does not climb out of Chondrichthyes, without the rule a whale is a fish.

Strict synset ↔ Q alignment (wd-align)

The loose rule above plus: a generality threshold for the common ancestor of 2% (loose: 5%), two supporting sources — label, lexeme, direct link (ILI and P8814 together count as one support, they are set together), folk ancestors. Ancestor support but a single source — status single.

• Manual check of the loose rule: label only — 17/27 (63%), one direct — 1/3, two or more supports — 19/20 (95%). Errors: a frog species instead of the toad family, a pit viper instead of colubrids, a heraldic juniper, lecturer instead of reader, Tagalog instead of Filipino. The strict rule rejects 12 of the 13 wrong ones; shipwreck remains (event ↔ wreckage).
• Strict-rule error: football player (American) ↔ association football player.
• The price of strictness: general concepts (bird, mammal, fish, crow) lack support — their OEWN ancestors are too general (vertebrate, animal). The walker takes them with lower confidence: two supports without ancestors — 0.7.
• Gold — data/align-gold.tsv (100 pairs, +/− with explanations of errors); wd-align measures every run against it.
• Output — synset_q-v2.tsv in the run; the derived synset_q.tsv and the database keep the loose rule.

Core walker (onto wd-walk)

• Core — data/core.tsv, 294 concepts: animals from fables and tales (104), food (50), household (83), family (16), roles (24), magical (7), plants (10). Synset — by domain (noun.animal, noun.food…), 291 in their own lexfile.
• Anchors — for 286 concepts: strict alignment 209, two supports without consistent ancestors 35, ancestor support with one source 42, anchor taxon 14. No anchor: lamb, fish (animal), lamp, key, shield, cloak, shepherd, giant.
• Pass: statements of the anchor and folk ancestors — up to 3 steps, plus a class at any depth (bird, mammal, fish, insect).
  • Overly general ancestors (≥ 20 k descendants: product, physical object, food, organism, animal) — no inheritance.
  • A "part" that is a taxon or the concept itself (fox P527 red fox, Cetacea P527 whale) — 44 discarded.
  • Confidence < 0.4 — 151 discarded.
• Numbers as distributions: the anchor's own values, otherwise over descendants (P279/P171 down to 3 steps, ≤ 400): mean, σ, min, max, n, SI unit. A single value — "point estimate, σ unknown". Birth weight (P3831 = Q4128476) is not about adults: 159 values discarded. Lion: ≈ 157 ± 44 kg (126–188, female and male; excluding the 1.65 kg newborn).
• Output — seeds/wd-candidates.tsv: 2 171 candidates for 232 concepts. Columns: concept, domain, synset, property, value, source (Q, P, rank, anchor and steps), confidence, mean, sd, min, max, n, unit. These are candidates, not truth: they enter the core only after confirmation.

Confidence = anchor (strict 0.9 · two supports 0.7 · one support or anchor taxon 0.6) × statement rank (preferred 1 · normal 0.9) × 0.85^steps to the ancestor. Numbers: own — ×0.95, distribution over descendants — ×0.7.

Examples: cat — part tail, claw, paw, tooth, snout; knife — part blade, handle, tang, use cutting; bread — material flour, water, salt; lion — food ungulate; horse — height ≈ 1.47 ± 0.19 m (n = 174, over 108 breeds).

LLM: a second opinion (onto wd-confirm)

claude -p (claude-opus-5-5, effort high via --settings env, API key from local.md, no tools), with a cap, log vmm-calls.jsonl. The answer is "yes / no / refine + why". The criterion is the typical instance, default knowledge.

• Confidence separates: 57% "yes" versus 17%.
• Systematic candidate errors revealed by the LLM:
  • no_part "human" for mammals (P3113 of a general class);
  • material "human food" / "food ingredient" for fruit and fish;
  • part "livestock", "predation", "she-bear" (P527 not about parts);
  • P1034 — a list of all possible prey rather than typical food: wolf — apple, grapes;
  • the color of water from "colors of water";
  • spade — black, from the card suit (a different sense).
• Among the "no" answers for confident candidates, most are "not typical" rather than "wrong": a tiger rarely eats fish, a therapy dog.

Character — only in fables

At the cost of 5 Prop values (Cunning Vain Greedy Industrious Wise, preference noun.animal) — 7 seeds on layer fable (d67–d73): the fox is cunning, the crow vain (Perry 124), the wolf greedy (Perry 156), the ant industrious, the grasshopper not (Perry 373), the owl wise (Perry 437), the donkey not (Perry 188). The real layer is silent about character: prop fox cunning → "unknown", with --layer fable → "yes".

Check queries and the cycle gate — green (walk.md).

Compiled core, context-based sense selection, description (2026-09-26)

Well-known standard animals live a compiled life, not as a dynamic class (Smart annotator — layers, context, memory, "The core is compiled"). The test: how well the model describes the animals, trees, etc. of the tales. Runs — onto3-2026-09-26.

Core in code (build.rs, src/core.rs, src/coresel.rs, src/coregate.rs)

• Selection (onto core-select → data/core-props.tsv): the core gets what the LLM confirmed, and confident (≥ 0.7) items without an LLM "no"; the rest stay candidates.
  • The LLM (onto core-vmm confirm, Opus, effort high) went through all 2 071 candidates without a verdict — both disputed and confident, because among the confident ones the second-opinion sample saw 32% "no". Color and "no part" were not checked — they are not in the core. Verdicts: yes 876, no 1 009, refine 190.
  • Risks (onto core-vmm risks): action → harm → probability per contact, order of magnitude from the LLM, for all 294 concepts. Lion: pet → bite p ≈ 0.3, approach → bite p ≈ 0.01.
• Explicit synsets for 27 concepts (third column of data/core.tsv, gate: the lemma is a member of the synset, the synset is in the domain's lexfile). The first sense in the domain's lexfile was wrong: dolphin — dolphinfish, oak and birch — wood, net — the internet, bow — a bow knot, peacock — a butterfly, hedgehog — porcupine, castle — palace, hut — a military barracks… The walker was rerun: 2 280 candidates, the rows of the other concepts — byte for byte the same.
• build.rs (the build script in the crate root; src/build.rs is the graph build, a different thing) generates enum CoreId (294 variants, a duplicate lemma gets its domain: FishAnimal, FishFood) and static CORE: [CoreRec; 294]:
  • food, habitat, parts, material, use — sets of Val: Q, label, OEWN synset of the value (strict alignment), a link to a CoreId if the value is also a core concept;
  • mass, height, length, lifespan, speed — Dist {mean, sd, min, max, n, unit};
  • risks — Risk {act, harm, p};
  • every value has a Source: Q, P, rank, steps from the anchor, confidence, confirmation (LLM "yes", refinement with a note, LLM estimate).
• Completeness gate. Every "concept × field" cell is explicit: a value, "no risks", "unknown" with a reason, or "not applicable" by domain (an animal has no material, a table does not eat). A deleted cell makes the build fail: completeness gate: concept «lion (animal)» has no field mass — neither values nor an explicit «unknown» (runs/…/build-gate-red.txt). The same code (src/coregate.rs) is checked by onto controls. The first build without core-props.tsv — ONTO_CORE_BOOTSTRAP=1.

All 937 values from Wikidata were confirmed by the LLM: "yes" — 896, "refine" — 41. No unchecked confident values remain. Numbers are the bottleneck: mass is known for 21 of 287 concepts, because Wikidata has numbers mostly for species, not for everyday concepts.

• "Core ↔ graph" gate (core::check, in load_kb): the synset is in the graph and the lemma is its member; graph seeds on the real layer do not contradict the domain (an animal is not Animate=no, food is not Edible=no, an object is not Living=yes); a core part or material is neither the concept itself nor its ancestor. A contradiction refuses the load and shows both sides, rather than silently picking one.
  • On real data the gate found 6 contradictions. Walnut, nut and acorn — food versus seed d55 "an object is not food": resolved by seeds d74–d76 (nut and acorn are edible). Lexfile not of the domain (nut, acorn — noun.plant, coin — noun.possession) — only a warning.
• Query order — core first, then graph:
  • isa, haspart — sense candidates: the query context's lexfile, the first sense, the core concept. They are ranked by a sentence-free WSD score: core 0.5 + corpus 1.5 × share + query lexfile 1; those within 0.3 of the best remain;
  • prop — in the preference lexfile, core concepts first;
  • describe — core fields, then the graph.

Word sense selection by context — WSD (src/wsd.rs)

A sense's score is a sum of terms, each visible in the explanation (onto wsd mouse "The cat caught a mouse"):

1. Extended Lesk (Banerjee & Pedersen 2003): words shared between the sentence and the glosses and lemmas of the synset (weight 1) and of its neighbours — hypernyms up to 2 steps, hyponyms, parts, whole, substances (weight ½); IDF weights over OEWN glosses.
2. Predominant sense from a corpus — McCarthy et al. 2004 (ACL P04-1036), onto wsd-prior:
   • a word's distributional neighbours — from within-sentence co-occurrences in the Leipzig corpora: eng_wikipedia_2016_1M, eng_news_2024_1M, co_s, significance, top-200 contexts, cosine;
   • neighbours vote for senses close in the graph: Wu–Palmer similarity to the 8th power (McCarthy used the sharp jcn);
   • output — 383 polysemous words in 18 s → sense-prior.tsv.
3. Core: +0.5 for a sense that is a core concept.
4. Domain: the share of sentence words whose senses (ranks 1, ½, ⅓) lie in the sense's lexfile.

Frozen set — data/wsd-frozen.tsv: 60 sentences, 14 words (dolphin, oak, bank, bass, apple, crane, seal, bat, jaguar, mouse, mole, pike, cricket, jellyfish). Sha256 2be56b43…, committed before the WSD code (f28a80d4).

• The weights were set in advance. Blind round — 43/60. The second (NON-blind) round changed the domain vote: previously the six senses of a word voted equally, and "cat" gave five artifacts (CAT scan, Caterpillar…) against one animal; now senses vote by rank. Together — the same 43/60; without corpus — 46 → 47.
• The corpus weight hurts here (72% versus 78% without it). News and Wikipedia pull mouse towards the device (neighbours goofy, clubhouse, keyboards), bass towards the instrument, cricket towards the game. In fairy-tale sentences this misfires; the core does not always outweigh it. Next — a fairy-tale corpus or the text's domain.
• Errors (all in runs/…/wsd.md with an explanation of each choice):
  • the core and corpus pull towards the tree where the context is about wood (oak ×2) and food (dolphin-mahi-mahi);
  • bank → "piggy bank";
  • apple-tree ×2.
• dolphin / oak (the errors of the first-sense rule) are answered right.
  • Frozen set: isa dolphin fish → "no" — the sense is the core mammal, dolphinfish rejected by the score; isa oak tree → "yes".
  • Sentences about the animal and the tree (w01, w02, w05, w06, w07, w10) — correct.
• Negative control — shuffled glosses (a sense's signature taken from a random synset): Lesk only 45 → 20/60 (baseline 21); everything together 43 → 28 (no sentence 29). Goes red. When only the gloss texts are shuffled while lemmas and neighbours are real, Lesk holds 36/60: neighbour lemmas (whale, cricket bat) carry part of the signal.

Concept description: onto describe and completeness on tale classes

onto describe fox --layer fable gives:

• the sense and why it was chosen;
• the IsA chain;
• parts, material, food, habitat, use (core, otherwise OEWN meronymy);
• sizes and mass as distributions;
• risks;
• seed properties, and on the fable layer — character.

Every row has a source.

onto describe-eval — 305 classes from the tale and fable worlds (302 described). "Before" — graph and seeds, without the core; "after" — the core, then the graph. The core covers 89 of 302 classes.

By field after: parts 96%, risks 37%, use 24%, material 20%, food 29%, habitat 8%, mass 4%.

Opus judge (onto describe-judge): all 604 descriptions — 302 classes "before" and "after". Blind: versions shuffled under random numbers. Fixed rubric: C — correctness, F — completeness, U — usefulness, E — errors, each 1–5. Classes split by name hash: training — 194, held out — 108.

• By category (before → after):
  • animals 2.95 → 3.31, food 2.83 → 3.34, objects 2.83 → 2.95, "other" 2.50 → 2.58;
  • people 2.89 → 2.80 and plants 2.96 → 2.88 — slightly worse: the core has almost no fields for them, and the judge counts "unknown" rows as gaps.
• Distribution of the mean score, rounded to an integer:
  • before: 15 / 51 / 164 / 72 / 0;
  • after: 18 / 54 / 141 / 77 / 12 — fives appeared.
• Best after: lion, wolf, cake — 4.5; the judge wants habitat and fuller risks.
• Worst after — 1.0, all due to a wrong word sense in classes outside the core:
  • flint — a river in Georgia;
  • shoes — "situation": the class uses the plural;
  • shell — an artillery shell.
• Typical errors:
  • animals — a taxon as a part (eagle "bird", hare "rabbit"), no habitat;
  • objects — no material, odd inherited parts;
  • plants — "hood" from plant;
  • people — taxonomy only.
• Judge control: 20 descriptions scored a second time with a differently worded rubric. Mean difference 0.35, within ±0.5 — 17 of 20, Spearman ρ = 0.75. The scores agree.

The "student — architect" loop — only the first round was done: the judge gave scores, the architect (describe-judge architect, saw only the training part) proposed three changes (runs/…/architect-r1.md):

1. Word sense by score, not the first one:
   • lemmatization of plurals (shoes → shoe);
   • fairy-tale lexfiles — artifact, animal, plant, food, person — up; state, communication — down;
   • a threshold below which the LLM breaks ties.
2. Parts and material only from physical ancestors and no further than 3 steps; general ancestors (whole, structure) give no parts; for magical creatures — a bridge to creature.
3. WordNet glosses into fields via templates: "made of", "used for", "having a … and a …", "feeds on"; such values — with source gloss and lower confidence.

The changes were not implemented, and there was no second round. To measure rounds, the judge must use stable ids (class + version): currently the numbers depend on the shuffle of a single run.

Pronoun API (src/pron.rs) — for the upcoming coreference

• pron::info(kb, lemma, &Ctx {sentence, layer}) -> PronInfo {gender, animate, human, persona, why}; wrappers — gender, animate, human; PronInfo::pronouns() — the allowed he / she / it.
• Sense: WSD by sentence, otherwise the core concept, otherwise the first sense.
• Human and animate — seeds Human / Animate on the layer.
• Gender — hints from the gloss and ancestor names: female / woman / queen… → fem., male / king / stallion… → masc.; both together — unknown.
• Character: an animal that talks on this layer (the fox in a fable) allows he / she / it.
• Only the interface and the pronoun_api test: crow — it, in a fable — he / she / it; Nixon — he or she; stone — it; unknown word — None.
• For a batch of sentences, Wsd should be built once (IDF — ~1 s); currently sense() builds it on every call with a sentence.

Check (2026-09-26)

• Frozen set (74) — 72/74, 0 wrong, 2 "unknown"; without the core and WSD (first-sense rule, non-blind) — 71/74, 1 wrong (isa dolphin fish).
• Bridge: 302/305 classes, fable questions — 14/14.
• Negative controls — 10/10, two added:
  • core completeness: cell fox × habitat deleted → an error naming the concept and field;
  • "core ↔ graph": an injected fox seed Animate=no → gate.
• WSD control (shuffled glosses) — goes red.
• cargo test --release — 31/31, including the completeness gate, pronouns, WSD basics, LLM risk parsing.
• API calls: core LLM — confirmation pass, risks, probe; judge — 604 descriptions and control, architect.

Independent tests (2026-09-26)

More tests for the common-sense builder. The sets are in tests/data/vmm/ (LLM) and in …/ext/, and the same files run on any build: onto vmm <tag> [--haspart rows.json] (the hash of every file is checked before the run; report — vmm-<tag>.md and .tsv in ONTO_RUNS).

LLM questions

• Author — Opus (effort high, saw neither the code nor the data; prompts — tests/data/vmm/prompts/), gold standard — common sense and biology, every question has an explanation. 400 questions, frozen by hashes before the first run (commit 66711cecf): isa with traps (80) and made-up words (10); parts, materials, members (80); abilities, properties and fable/fairy layers (80); measure X mass|height|length|speed|lifespan — ranges (70) and made-up animals (5); sense W | sentence — lexicographer file and keyword (75).
• Run on the build before the core and WSD, which had no measure and sense: 150 "not supported".

Analysis of the 13 wrong answers (a third path — biology and definitions, not the data): all are our defects.

• sense selection (the first-sense rule): isa dolphin fish → yes (first sense — dolphinfish), isa bamboo grass → no (first sense — stems as material), isa king woman → yes (the instance Billie Jean King in lower case);
• OEWN data: isa moss plant → no (mosses are bryophytes outside "plant"), isa tomato fruit → no (culinary taxonomy: a vegetable; botanically — a fruit), isa doctor person → no — check the chain;
• part inheritance is too generous: haspart insect bone, haspart jellyfish brain → yes;
• seeds: can butterfly fly → no (no CanFly for insects), can human speak → no ("human" is a hominid under "animal", with no human seed); the fairy layer lacks "animals talk": can wolf|cat|frog speak --layer fairy → no.

"Unknown" instead of "yes" for parts (48) — incomplete OEWN meronymy, as intended: we never say "no" about parts.

From the internet: catalog and samples

HS: world model of goods (2026-09-26/27)

The goods classifier (HS, from the import-export project, about 6 thousand entries) is read into a word ontology, as for the tales. HS is a ready-made world model of goods (Smart annotator — layers, context, memory, "Thousands of domains"): a tree, "others", features, thresholds. Runs — hsonto-2026-09-26.

Data and license.

• Source — BACI/product_codes_HS22_V202601.csv (BACI, read only): 5 609 HS22 subheadings with English descriptions. BACI takes the descriptions from UN Comtrade.
• The description texts belong to the WCO, so they are not in the repo. data/hs/ carries only derived data: code → concept, features, labels. Verified: not a single description verbatim.
• Section and heading names are not available in our data: the trade data and the translation project have only HS6, and P5471 (HS 2022) was not extracted into wikidata.duckdb. So heading labels are ours: the most frequent head concepts of the heading.

How it works (src/hs.rs, src/hsmap.rs)

1. Tree: 97 chapters → 1 229 headings → 5 609 subheadings, with parents (data/hs/tree.tsv). "Others" are marked separately:
   • by code (…9, …90) — 1 375;
   • textual (n.e.c., other than) — 1 476;
   • both together — 906.
2. Parsing. en (model ud-ewt-eslspok) gives lemmas, parts of speech and the tree in 2.6 s for all descriptions. The HS telegraphic style ("Fruit, edible: apples, fresh") is split by colon into levels (0 — the heading top) and by commas into pieces. A piece is either features from closed enums (enum State, Process, Attr) or a noun phrase. Rules that emerged on real descriptions:
   • the list "of sheep, goats, horses" extends across commas;
   • "excluding …" and "not fillets, meat of 0304…" — exclusions to the end of the level, "(excluding X)" — only within the parentheses;
   • "men's or boys'" — a purpose, not a product;
   • Latin names in parentheses — taxa;
   • "Nuts, edible" → edible nut, "Lighters: cigarette" → cigarette lighter, if such a compound exists in OEWN.
3. Thresholds — via math:
   • "over 1000 but not over 1500cc" → Capacity ∈ (0.001, 0.0015] m³: exact rationals Rat, bounds with inclusivity;
   • units from math::units (kg, g, m, cm, mm, %), g/m² and cm³ — via Mono; t, l, kW, kVA, V, dtex, years — with an exact multiplier.
4. Alignment (data/hs/terms.tsv, codes.tsv), with the reason on every row. Steps:
   1. Phrase → the longest one present in OEWN; British spelling as a fallback.
   2. WSD with the heading context: Lesk, corpus, core, domain.
   3. Product state: live → animal or plant; fresh, frozen… → food (also accepts animal). If the longer phrase lacks such a sense, the shorter one is taken: fatty liver — a disease, liver — food.
   4. Chapter domain — in a second pass: the lexfile share of the chapter's head concepts, "one sense per text".
   5. Then — the core concept, and then a Q via strict alignment (accepted).
5. LLM confirmation (onto hs confirm → data/hs/sense-confirm.tsv).
   • Key — phrase, chapter and state class: live and frozen fish are not merged.
   • We take 240 disputed (margin < 0.15) and most frequent keys.
   • Final run: the LLM agreed in 112, chose another sense in 105, "none fits" — 23.
   • WSD errors in frequent words are the most costly: rubber in chapter 40 — a condom, iron — a clothes iron, steel — a knife sharpener, glass — a drinking glass, wood — not timber.

Numbers

Features (data/hs/feats.tsv): how many features, in how many subheadings.

• Most frequent values:
  • state — fresh 253, chilled 175, frozen 162, dried 134, edible 131, raw 116, live 70;
  • processing — knitted 292 (¬ 230), coated 232, woven 227, rolled 165.
• "of X": material 1 889, source organism 357, other 713, not aligned 162.
• Thresholds: share (%) 285, width 125, g/m² 121, thickness 67, mass 48, dtex 39, volume 31, power 28, voltage 21, age 14, Brix 12.

IsA gate. HS says "product ⊂ heading top" (apple ⊂ edible fruit). The negative control swaps products between headings.

• Strict verdicts — "yes", another sense, part, taxon, modifier: 1 068 (40.7%) versus 55 (2.1%) in the control. The gate discriminates.
• "Made of the material" and "source" — a type check via lexfiles, not a fact check. The control gives even more of them, so these are soft labels, not confirmation.
• "No" in a sample of 16:
  • about ⅔ are our defects. The phrase fell back to a more general one: dairy produce → produce, ferro vanadium → vanadium. Wrong sense: lead the object instead of the metal. Parsing garbage.
  • about ⅓ are genuine HS vs WordNet discrepancies: relay ⊄ electrical apparatus; starter motor ⊄ ignition equipment; safety seat belt sits under vehicle in HS, while for us seat belt is neither a kind nor a part of vehicle.

Feature versus core and seeds — 21 cases, and the gate found a gap in the seeds:

• HS "edible" contradicts seed d55 "an object is not food by default" on fruit (the fruit as a plant organ), snake, salmon;
• HS "live" contradicts d02 on bulb (a bulb is a living plant part);
• fruit and animals as food lack Edible;
• core-material (4) — a parsing defect: "snake of meat".

Neighbour thresholds. Subheadings of a heading with otherwise identical features must differ by a threshold.

• Pairs — 313, undistinguished — 33: parsing lost a feature or cut off a threshold. An unknown feature remains a label with negation ("filled / not filled" in chocolate).
• Control — copying a neighbour's threshold — goes red on 313 of 313.

New concepts (data/hs/new-concepts.tsv) — 494 phrases (894 occurrences) that are neither in the core nor in the graph:

• monofilament 35, cermet 14, peroxyacid 10, furskins 9, nonwoven 8;
• fish families: Merlucciidae, Moridae, Muraenolepididae;
• toothfish, epoxide;
• some are parsing fragments: mechanical, poly, non-electric.

Blind Opus judge (onto hs judge).

• Sample: 150 subheadings by code hash, across all chapters. Plus 20 decoys: the output of another code attached to a foreign description. Everything shuffled.
• Rubric (1–5):
  • M — head concept and its sense;
  • F — completeness of features;
  • P — precision: nothing wrong.
• Runs. The sample and decoys are the same in both runs; r1 — before the last parsing fixes.

• Worst in r2 — a wrong sense of a word or of the head concept:
  • portal (crane) — an entrance;
  • counterweight;
  • cocoon instead of silk waste;
  • "Reception" — an action, not apparatus.
• Decoys — exactly 1.00: the rubric does not praise foreign output.

Labels in other languages (onto hs labels, sql/hs-labels.sql). Q labels from Wikidata (CC0, data/hs/labels.tsv) were checked against the translation project's translations of HS names: does the stem of our concept's label occur in the subheading's translation.

Mismatches are mostly Q taxa with a Latin label (Bos taurus, Bubalina, Thunnini), families instead of species (Nephropidae), synonyms ("Innereien" versus "Schlachtnebenerzeugnisse"). Less often — our wrong concept: rose → rose hip.

API calls — within the cap: confirmation (a probe and two runs), judge (two runs).

onto hs describe examples (in full — runs/…/describe-examples.txt):

080810 — subheading of heading 0808 [apple, pear, quince] of chapter 08 [fruit, coconut, almond, apple]
head concept (goods): apple → oewn-07755101-n «apple» (noun.food; core: Apple; Q89)
  why: Lesk 0.77 [fruit, edible] + corpus 1.17 + core 0.50 + domain 0.32 + HS state 1.00 + chapter 0.94 = 4.70
  is a kind of (OEWN): apple → edible fruit → produce → solid food → food → …
HS: apple ⊂ edible fruit — ours: yes
features: state: edible; fresh
world model: parts pericarp; mass ≈ 0.167 kg ± 0.047; risks eat → allergy p ≈ 0.01 [core, with sources]

030213 — subheading of heading 0302 [fish, Alaska pollock, Atlantic salmon, Nile perch] of chapter 03
head concept (goods): Pacific salmon → oewn-02537365-n «salmon» (noun.animal; core: Salmon; Q2796766)
HS: Pacific salmon ⊂ fish — ours: yes (salmon → food fish → fish)
features: state fresh; chilled · except fillet; meat; fish offal · reference 0304 · taxon Oncorhynchus nerka, … (7)

870322 — subheading of heading 8703 [vehicle] of chapter 87
head concept (heading head): vehicle → oewn-04531608-n «vehicle» (noun.artifact)
features: threshold Capacity ∈ (0.001, 0.0015] m³ · composition: spark ignition internal combustion piston engine

Check.

• cargo test --release — 36/36: +5 HS tests — exact thresholds, bound intersection with a negative control, telegraphic-style features, CSV and tree, IsA gate on a synthetic graph.
• Frozen set — 72/74 (0 wrong).
• WSD — 43/60.
• Controls — all green.

Commands

export PATH=$HOME/.cargo/bin:$PATH
cargo build --release
../target/release/onto build                        # OEWN 2025+ and FrameNet 1.7 → $MOVA_DATA/onto/onto-v1.bin (ONTO_SNAPSHOT)
../target/release/onto isa apple fruit
../target/release/onto can crow speak --layer fable
../target/release/onto haspart bird beak · parts bird · frame buy · role buy Seller · inherits Commerce_buy Getting
../target/release/onto entails snore sleep · word whale · up crow
../target/release/onto eval blind                   # frozen set → $MOVA_DATA/runs/onto-2026-09-26/frozen-blind.{md,tsv} (ONTO_RUNS)
../target/release/onto controls | bridge | bench | stats
../target/release/onto wd-lexemes $MOVA_DATA/raw/wikidata/20260923/wikidata-20260923-lexemes.json.bz2 <dir>
../target/release/onto wd-map <dir>
nice ../target/release/onto wd-probe $MOVA_DATA/raw/wikidata-all/latest-all.json.bz2 1 <dir>
# Wikidata: pass and database (section "Wikidata: pass and database"; runs — ONTO_WD_RUNS, default $MOVA_DATA/runs/wikidata-2026-09-26)
D=$MOVA_DATA/raw/wikidata-derived
nice -n 10 ../target/release/onto wd-sample $MOVA_DATA/raw/wikidata-all/latest-all.json.bz2 <dir> --k 48 --mb 12 --level 9 --skip-wm
nice -n 10 ../target/release/onto wd-extract $MOVA_DATA/raw/wikidata-all/latest-all.json.bz2 $D --bz 8 --workers 2 --level 9 --skip-wm [--gb 2]
nice -n 10 ../target/release/onto wd-lex $MOVA_DATA/raw/wikidata/20260923/wikidata-20260923-lexemes.json.bz2 $D
../target/release/onto wd-senses $D <run>/wd-senses-en.tsv && ../target/release/onto wd-map <run>
nice -n 10 ../target/release/onto wd-closure $D --threads 8
nice -n 10 ../target/release/onto wd-align $D --wd-map <run>/wd-map.tsv [--generic 0.02] [--min-src 2] [--no-folk] [--out synset_q.tsv]
# walker (section "Wikidata walker"; runs — ONTO_WD_RUNS=$MOVA_DATA/runs/onto2-2026-09-26)
nice -n 10 ../target/release/onto wd-walk $D [--align <run>/synset_q-v2.tsv] [--core data/core.tsv] [--out seeds/wd-candidates.tsv]
../target/release/onto wd-confirm [--n 60] [--batch 30] [--cap 10] [--lo 0.45 --hi 0.8] [--out seeds/wd-confirm.tsv]   # LLM, $ — API
../target/release/onto prop fox cunning --layer fable
(cd $D && ~/bin/duckdb $MOVA_DATA/db/wikidata.duckdb < <repo>/onto/sql/wikidata-load.sql)
~/bin/duckdb -readonly $MOVA_DATA/db/wikidata.duckdb < sql/wikidata-checks.sql
# core, WSD, description (section "Compiled core, …"; runs — ONTO_RUNS / ONTO_WD_RUNS=$MOVA_DATA/runs/onto3-2026-09-26)
../target/release/onto core-vmm confirm [--batch 50] [--cap 7]      # LLM goes through candidates without a verdict, $ — API
../target/release/onto core-vmm risks [--batch 50] [--cap 8]        # risks of core concepts
../target/release/onto core-select                                  # → data/core-props.tsv; then cargo build (completeness gate)
../target/release/onto wsd-prior                                    # predominant senses from the Leipzig corpora → $MOVA_DATA/onto/sense-prior.tsv
../target/release/onto wsd mouse "The cat caught a mouse in the kitchen"
../target/release/onto wsd-eval                                     # WSD frozen set + shuffled-gloss control
../target/release/onto describe fox --layer fable [--no-core]
../target/release/onto describe-eval                                # tale classes before/after the core, completeness
../target/release/onto describe-judge [--cap 15] | describe-judge report   # blind Opus judge, $ — API
# HS: world model of goods (section "HS"; runs — ONTO_RUNS, default $MOVA_DATA/runs/hsonto-2026-09-26)
../target/release/onto hs build [--no-confirm]                      # HS22 (BACI) → data/hs/*.tsv, report hs-report.md (~5 s)
../target/release/onto hs describe 080810 [--text]                  # product description; --text — WCO description from a local file
../target/release/onto hs parse 030213                              # parsing debug (tokens and features)
../target/release/onto hs confirm [--n 240] [--margin 0.15]         # LLM confirms disputed and frequent senses, $ — API
../target/release/onto hs judge [--n 150] [--decoys 20]             # blind Opus judge, decoys as control, $ — API
(cd $ONTO_RUNS && ~/bin/duckdb -readonly $MOVA_DATA/db/wikidata.duckdb < <repo>/onto/sql/hs-labels.sql) && ../target/release/onto hs labels
cargo test --release                               # 36: world model, Wikidata, folk rule, anchors, LLM, core, pronouns, HS

What's next

0. Core and WSD:
   • a fairy-tale corpus for predominant senses (news pulls mouse towards the device);
   • numbers for everyday concepts — mass is known for 21 of 287;
   • the "student — architect" loop: implement the round-1 proposals (word sense by score, physical ancestors for parts, glosses → fields) and measure on the held-out part;
   • one Wsd per batch for pronouns.
1. Word sense from the text. WSD chooses by sentence; by text, "one sense per text" exists only as a hook. A second frequency vote — the sense order of Wikidata lexemes and our corpora.
2. Wikidata into the core: human confirmation of what build.rs stitches into the core; filters for the systematic errors revealed by the LLM (P3113 of general classes, P186 "human food", P527 not about parts); P1034 — typical food, not all prey; general concepts (bird, mammal) — support via P8814/ILI without ancestors.
3. Seeds from OEWN glosses automatically ("flightless", "edible", "swift" → a property with gloss provenance) with a manual sample.
4. Selectional preferences from FrameNet roles (role semTypes) and IsA: who can be a Buyer, what can be eaten. Then — arc plausibility for the second parsing pass (Smart annotator — layers, context, memory).
5. World: the LLM writes an entity's class directly as a synset id. The bridge reads the world state, not just the commands.
6. Reasoning module: Horn clauses over the TBox (world model) and the ABox (the tale's world), every conclusion with a proof.
7. Size, quantity, comparison: a lion is big, a mouse is small, a hare is faster than a tortoise.

Files

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/onto/