Lingua Mundi: Building a Language-Agnostic Map of Meaning from Open Lexical Data

1 · The intuition: why dictionaries are the wrong shape

Every language learner has felt the friction. You look up cat in an English–Spanish dictionary and find gato. You look up gato in a Spanish–Japanese dictionary and find 猫. Three dictionaries later you have spent an afternoon confirming what you already suspected: the same idea — a small domesticated carnivore that purrs — is being described three times, in three separate files, by three different organizations, with three different identifiers, and never once connected to the other two.

The intuition behind Lingua Mundi is that this is the wrong shape for linguistic data. The meaning exists once. The words exist many times. A database that stores meanings as first-class objects, and treats words as labels that different languages attach to those meanings, would let a person (or a program) move from any language to any other language through a single shared map — and would let a learner see, at a glance, that cat, gato, chat, Katze, and 猫 are not five facts to memorize but one fact wearing five coats.

This intuition did not arrive as a thunderbolt. It grew, over roughly a year, out of a particular person's daily work — and the earliest documentary trace of it is remarkably plain.

1.1 The first message

On 3 December 2025, the author opened a conversation with ChatGPT and typed the following (quoted verbatim; the conversation remained open for four and a half months and was last edited on 21 April 2026) [H]:

"I want to find free dictionaries similar to "I want to find free dictionaries similar to JMDICT and Kanjidict2 for the following languages: Spanish, English, French, Italian, Portuguese, Chinese, Korean and Vietnamese. My intent is to create a unified multilingual database for various purposes. I'm also interested in knowing if similar databases exist for indigenous Mexican languages that I can freely access." and Kanjidict2 for the following languages: Spanish, English, French, Italian, Portuguese, Chinese, Korean and Vietnamese. My intent is to create a unified multilingual database for various purposes. I'm also interested in knowing if similar databases exist for indigenous Mexican languages that I can freely access."

(Session log, ChatGPT, conversation "Free multilingual dictionaries", 2025-12-03; source file cited in Appendix A.)

Three things in that short message foreshadow everything the project later became. First, the languages named — Spanish, English, French, Italian, Portuguese, and, later, Japanese — are exactly the languages the author speaks and works with as a translator and language teacher (a set he would later formalize as the "Core 6"), plus Chinese, Korean, and Vietnamese, which he planned to add next. Second, the phrase "a unified multilingual database" already names the goal: not six dictionaries, but one database holding all of them together. Third, the author was thinking about indigenous Mexican languages from the very first day — a thread that returns repeatedly (a Maya-language integration for Lingua Mundi was explored in August 2026 [H], and Mayan-language resources appear throughout his research notes).

The context behind that message matters too. The author is not a computer scientist by training. He is a translator and language tutor who was at the time also learning Japanese seriously, building small tools to make that learning easier: The context behind that message matters too. The author is not a computer scientist by training. He is a translator and language tutor who was at the time also learning Japanese seriously, building small tools to make that learning easier: kanji and kana learning games and a Japanese dictionary pipeline (a furigana editor for Japanese e-books would join this toolbox later, in 2026, as the project that became the and kana learning games and a Japanese dictionary The context behind that message matters too. The author is not a computer scientist by training. He is a translator and language tutor who was at the time also learning Japanese seriously, building small tools to make that learning easier: kanji and kana learning games and a Japanese dictionary pipeline (a furigana editor for Japanese e-books would join this toolbox later, in 2026, as the project that became the (a furigana editor for Japanese e-books would join this toolbox later, in 2026, as the project that became the Shikibu app). His home directory in late 2025 shows the archaeology of that learning: a copy of the KANJIDIC2 kanji dictionary file downloaded on 11 November 2025, a Joyo kanji list in December, and then — prompted by that 3 December conversation — the Open Multilingual Wordnet data downloaded on 5 December [H] (file timestamps, /home/rex/; see Appendix A). The ChatGPT conversation of 3 December, in other words, is the hinge: the moment when "I need Japanese data for my tools" widened into "I want to build a unified multilingual database, for free, for many languages, including my own country's languages."

1.2 From dictionaries to meanings: the December 2025 pipeline

What the author did next is visible in the files he created. Between 5 December and 16 December 2025 he wrote, with ChatGPT as a programming partner, a small pipeline of Python scripts — extract_omw.py, merge_dicts.py, unified_dictionary_pipeline.py, crossref_deterministic.py — that downloaded, parsed, and merged lexical data from several open sources into progressively more structured forms [H] (scripts and dates in /home/rex/). The most important design decision of this period is documented in the files themselves: the pipeline's final stage, dated 16 December, splits every entry into four kinds of objects — concept, sense, lexeme, and grapheme (a (a grapheme being a written character such as a kanji) — and writes them to disk as JSON Lines documents. The schema is telling: a being a written character such as a kanji) — and writes them to disk as (a grapheme being a written character such as a kanji) — and writes them to disk as JSON Lines documents. The schema is telling: a documents. The (a grapheme being a written character such as a kanji) — and writes them to disk as JSON Lines documents. The schema is telling: a is telling: a sense is a particular meaning of a particular word in a particular dictionary; a concept is the shared idea that several is the shared idea that several senses across several languages all point to; and the across several languages all point to; and the concept_id field of a sense is allowed to be empty — meaning "this word sense has not yet been linked to a shared — meaning "this word sense has not yet been linked to a shared concept." The author's notes in that code describe the method as "zero heuristics": no fuzzy guessing, no statistical similarity tricks to decide that two words mean the same thing; only explicit, documented links. ." The author's notes in that code describe the method as "zero heuristics": no fuzzy guessing, no statistical similarity tricks to decide that two words mean the same thing; only explicit, documented links. [H] (pipeline code and annotations, /home/rex/, December 2025; details in Appendix A.)

That 16-December schema is, in miniature, the entire architecture of the project that would later be called Lingua Mundi: words live in languages; meanings live in a shared space; and the hard, honest problem — which word in which language means which concept — is never faked. A sense that has not been linked to a concept is kept in the database, visibly unlinked, rather than being silently guessed into place. Ten months later, when a bug in an import pipeline made the English word cat invisible to its concept, the author would describe the principle in exactly these terms: better to show an orphan than to hide one ([H]; see Section 7).

1.3 The gap, and the widening

Then the paper trail goes quiet for roughly five months — from mid-December 2025 until late May 2026, the available logs contain no dictionary-pipeline work [NV]. This is not unusual for a working translator with students to teach, but it is worth stating plainly: the project's growth was not continuous. It advanced in bursts, separated by weeks or months of other work — a rhythm that would continue throughout 2026.

When the thread resumes, the ambition has widened. On 1 May 2026 the author is studying "OMW v2.0 data coverage" (the the author is studying "OMW v2.0 data coverage" (the Open Multilingual Wordnet release) release) [H] (conversation title and content, ChatGPT, 2026-05-01). On 27 May 2026 he opens a conversation titled "Open Source Language Resources" and asks, in effect, for a systematic map of every free lexical, grammatical, and semantic resource that could feed a multilingual knowledge base [H] (ChatGPT, 2026-05-27). By now the project has a name and a shape; the evidence for both appears in the next chapter.

1.4 Why this matters beyond one person's project

A reader may wonder why a paper should document a personal project's intuition at such length. The answer is that the intuition is the contribution. The idea that meanings — not spellings — are the natural unit of multilingual data has a long academic pedigree (it is the oldest dream of machine translation's " — not spellings — are the natural unit of multilingual data has a long academic pedigree (it is the oldest dream of machine translation's "interlingua" tradition and the organizing principle of Princeton WordNet's synsets; see Section 2), but it is rarely acted upon by an individual, with no budget, as a public good. Most multilingual resources that follow this principle — the Open Multilingual Wordnet, BabelNet, the " tradition and the organizing principle of — not spellings — are the natural unit of multilingual data has a long academic pedigree (it is the oldest dream of machine translation's "interlingua" tradition and the organizing principle of Princeton WordNet's synsets; see Section 2), but it is rarely acted upon by an individual, with no budget, as a public good. Most multilingual resources that follow this principle — the Open Multilingual Wordnet, BabelNet, the 's — not spellings — are the natural unit of multilingual data has a long academic pedigree (it is the oldest dream of machine translation's "interlingua" tradition and the organizing principle of Princeton WordNet's synsets; see Section 2), but it is rarely acted upon by an individual, with no budget, as a public good. Most multilingual resources that follow this principle — the Open Multilingual Wordnet, BabelNet, the ; see Section 2), but it is rarely acted upon by an individual, with no budget, as a public good. Most multilingual resources that follow this principle — the Open Multilingual Wordnet, BabelNet, the Universal Dependencies project — are maintained by university consortia with grant funding. Lingua Mundi is an attempt by one independent researcher, using only open data and free tools, to build a piece of that infrastructure himself, for his own translation and teaching work, and to give it away. Whether that attempt succeeds is a question the reader can judge from the evidence in the rest of this paper — and the author's own journals show him asking the same question, with the same honesty, all along (see Section 4.3).

2 · Background: what "a map of meaning" means

To understand what Lingua Mundi is trying to build, and what is genuinely new about it, three strands of prior work matter: the linguistic idea that meaning can be separated from spelling; the computational idea that such meaning can be stored as data and linked across languages; and the modern observation that a structured map of meaning complements — rather than competes with — neural language models.

2.1 The sign and its two faces

Modern linguistics begins with a deceptively simple observation, usually credited to Ferdinand de Saussure's early twentieth-century lectures: a word is a pairing of a sound/shape (the " (the "signifier") and a ") and a concept (the "signified") (de Saussure, 1916/1983). Two languages, on this view, are not two sets of words that happen to translate; they are two sets of signifiers that partially overlap on a shared space of signifieds. The English signifier cat and the Spanish signifier gato are different shapes with the same signified. If a machine could store signifieds once and attach signifiers to them, translation would become a lookup through a shared middle — an idea known in machine translation as the interlingua approach: instead of translating English→Spanish directly, translate English→meaning→Spanish.

The interlingua dream has a long history in machine translation research and repeatedly proved too hard in full generality (languages do not carve meaning into identical pieces, and every translator knows that cat the animal and cat the jazz musician are different signifieds wearing one signifier). But the dream never died; it was just narrowed into something buildable: the lexicon, the part of language that maps best onto stable, shared meanings.

2.2 WordNet and the synset

The decisive narrowing came from psychology and lexicography. In the 1980s and 1990s, George Miller and colleagues at Princeton built WordNet, a lexical database for English organized not alphabetically but semantically (Miller, 1995; Fellbaum, 1998). WordNet's core invention is the synset (a contraction of "synonym set"): a group of words that can stand for the same meaning in some context. The synset {cat, true cat, Felis catus, domestic cat} is one meaning; the synset {cat, kat, guy, hombre} — meaning "an informal term for a youth or man" — is another, quite different meaning that happens to share the signifier cat. WordNet also records relations between synsets — most importantly the is-a hierarchy ( hierarchy (hypernymy: a domestic cat : a domestic cat is an feline, a feline is an carnivore, a carnivore is an* mammal…) — turning the dictionary into a is an* mammal…) — turning the dictionary into a graph, with meanings as nodes and typed relationships as edges., with meanings as is an* mammal…) — turning the dictionary into a graph, with meanings as nodes and typed relationships as edges. and typed relationships as edges.

WordNet matters to Lingua Mundi for three reasons. First, it proved that a machine-readable lexicon organized by meaning is feasible and enormously useful — modern search engines, spell checkers, and WordNet matters to Lingua Mundi for three reasons. First, it proved that a machine-readable lexicon organized by meaning is feasible and enormously useful — modern search engines, spell checkers, and natural-language libraries still lean on it. Second, it gave the world a libraries still lean on it. Second, it gave the world a stable identifier system for meanings: Princeton WordNet numbers its synsets, so 02121620-n unambiguously means "domestic cat" and can be cited across databases. Third, it became the anchor of a multilingual project: the Open Multilingual Wordnet (OMW).

2.3 The Open Multilingual Wordnet: one meaning, many languages

The OMW, launched by Francis Bond and Ryan Foster (Bond & Foster, 2013), collects wordnets built in many languages and aligns them to the Princeton English wordnet: every participating language's synsets carry pointers to the English synset they correspond to. The result is the resource that made Lingua Mundi possible: a set of explicit, human-curated statements of the form "Spanish gato (sense 1) belongs to the same meaning as English domestic-cat synset 02121620-n." When the Lingua Mundi author wrote in his July 2026 engineering specification that the project would be "synset-centric" and would align languages through the OMW [H] (specification document, 24 July 2026), he was standing on Bond and Foster's shoulders — as his December 2025 pipeline scripts, which parsed the OMW's data files directly, already had been.

2.4 The open-data ecosystem around the wordnets

The OMW is one node in a large, openly licensed ecosystem that an individual can legally and technically combine:

• Wiktionary — the collaborative dictionary hosted by the Wikimedia

Foundation — provides very wide lexical coverage in hundreds of languages, including definitions, translations, Foundation — provides very wide lexical coverage in hundreds of languages, including definitions, translations, inflections, and pronunciation, under share-alike licenses (, and pronunciation, under share-alike Foundation — provides very wide lexical coverage in hundreds of languages, including definitions, translations, inflections, and pronunciation, under share-alike licenses ( (Wikimedia Foundation; see also Meyer & Gurevych, 2012, on its value as an expert-lexicon rival).

• Wikidata — Wikimedia's structured knowledge base — stores statements

with machine-readable identifiers and is released under CC0 (public domain) (Vrandečić & Krötzsch, 2014).

• Universal Dependencies (UD) — a community project that publishes

syntactically annotated sentence syntactically annotated sentence treebanks in more than a hundred languages under open licenses ( in more than a hundred languages under open licenses (Nivre et al., 2020). ). UD contributes real sentences and part-of-speech information — evidence of how words are actually used. contributes real sentences and ). UD contributes real sentences and part-of-speech information — evidence of how words are actually used. information — evidence of how words are actually used.

• [JMdict and KANJIDIC2](https://www.edrdg.org/jmdict/j_jmdict.html) — the Japanese dictionary and kanji files

maintained by Jim Breen's EDRDG project, covering Japanese extensively and with translations into several languages (Breen, 2004). These were the author's starting points (his very first dictionary file, downloaded in November 2025, was KANJIDIC2 [H]).

• UniMorph — a multilingual repository of inflectional — a multilingual repository of inflectional morphology (word (word

forms such as run/ran/running organized by grammatical feature) (Kirov et al., 2018; McCarthy et al., 2020).

• OPUS — a large collection of parallel (translated) sentences extracted

from the web and from movie subtitles, openly licensed (Tiedemann, 2012).

• Word frequency lists such as [wordfreq](https://github.com/rspeer/wordfreq) (Speer, 2022), which rank

words by how common they are in real words by how common they are in real corpora..

• The The ISO 639-3 standard standard, which assigns every human language a stable

three-letter code (e.g., spa for Spanish, jpn for Japanese), maintained by SIL International — the Lingua Mundi database uses these codes as the keys of its language table [H].

Two properties of this ecosystem deserve emphasis because they structure the whole project. First, the licenses are genuinely open — permissive or share-alike — so combining them into a derived database is legal as long as attribution and license terms are respected (Lingua Mundi records the source and license of every imported Two properties of this ecosystem deserve emphasis because they structure the whole project. First, the licenses are genuinely open — permissive or share-alike — so combining them into a derived database is legal as long as attribution and license terms are respected (Lingua Mundi records the source and license of every imported dataset in its own in its own source table; see Section 6). Second, the ecosystem is complementary: : Wiktionary is broad but shallow; the wordnets are deep but narrow; UD is about usage; UniMorph is about word forms; frequency lists are about prominence. The engineering problem of Lingua Mundi is not "find data" — it is "merge data that was never designed to be merged, without lying about the joins." is broad but shallow; the wordnets are deep but narrow; UD is about usage; : Wiktionary is broad but shallow; the wordnets are deep but narrow; UD is about usage; UniMorph is about word forms; frequency lists are about prominence. The engineering problem of Lingua Mundi is not "find data" — it is "merge data that was never designed to be merged, without lying about the joins." is about : Wiktionary is broad but shallow; the wordnets are deep but narrow; UD is about usage; UniMorph is about word forms; frequency lists are about prominence. The engineering problem of Lingua Mundi is not "find data" — it is "merge data that was never designed to be merged, without lying about the joins."; frequency lists are about prominence. The engineering problem of Lingua Mundi is not "find data" — it is "merge data that was never designed to be merged, without lying about the joins."

2.5 Why a map of meaning still matters in the age of large models

A fair objection must be stated plainly: if large language models can already translate, define, and reason about words, why build a structured lexicon at all? The Lingua Mundi author's answer, developed in his June 2026 research paper (Section 3) and refined through the summer, has three parts. First, verifiability: a structured database can show its work — every link has a source and a license and a date — whereas a model's knowledge is an opaque statistical mixture that cannot be audited, corrected, or cited. Second, cost and access: querying a local database costs nothing and works offline, whereas every model query costs energy, money, or both; the author's standing design rule is "zero-: querying a local database costs nothing and works offline, whereas every model query costs energy, money, or both; the author's standing design rule is "zero-token research queries" — the local API research queries" — the local : querying a local database costs nothing and works offline, whereas every model query costs energy, money, or both; the author's standing design rule is "zero-token research queries" — the local API is the research surface [H] (project note, 29 August 2026). Third, complementarity: the June 2026 paper argued — as a hypothesis, with the audits of three models converging on the same verdict — that a compact model operating over concept identifiers rather than raw tokens could be far more efficient precisely because the concept map does the semantic work outside the network (see Section 3). Whether or not that hypothesis survives experimentation, the map itself remains useful to the models: as an external memory that can be queried, updated, and cited without retraining.

The intellectual ancestry of Lingua Mundi can thus be summarized in one line: it is the interlingua dream, narrowed to the lexicon by WordNet, made multilingual by the OMW, made legal by the open-data movement, and made personal by one translator's need for a dictionary that treats meaning, not language, as the thing being filed.

3 · The name, the thesis, and the seven-graph architecture (May–June 2026)

3.1 Naming

By late May 2026 the author's notes and specifications used the working title "Multilingual Lexical Platform" [H] (markdown specification master dated 23 May 2026, /home/rex/Descargas/Phone backup/Download/). The name Lingua Mundi — "language of the world" in Latin — first appears in the available record on 1 June 2026, as the title of a research manuscript (see below) [H] (conversation logs, 1 June 2026; earliest verified occurrence in the author's archives). The name is apt: the project's claim is not that it contains many languages but that it is about the one space of meaning those languages share.

3.2 A research thesis: meaning, not tokens

Between roughly March and June 2026 the author developed — in conversation with ChatGPT, Claude, and DeepSeek — a research thesis about why a concept-level map of language could make language technology radically more efficient and accessible. The thesis, in the working framing the author brought to the conversation (quoted verbatim from his feasibility request of 18 July 2026, where he pasted the description for review) [H]:

"Create a multilingual, open, machine-readable map of human language that separates: 1. Meaning · 2. Form · 3. Structure · 4. History · 5. Usage — into interoperable graphs. The long-term objective is to make language computable at the concept level rather than the token level. […] Modern NLP largely models: token → token. Lingua Mundi models: concept ↔ language."

The underlying argument, stated in the June manuscript, is that a language model spends most of its capacity on surface variation — treating "dog", "perro", and "犬" as three unrelated things when they are one thing in three coats [H] (manuscript V4, quoted in full in the audit conversation of 1 June 2026). If a model could instead read and write concept sequences — with the concept map doing the multilingual work outside the network — the model could be smaller, cheaper, and runnable on ordinary computers.

3.3 Seven versions in one morning: the adversarial-collaboration method

The manuscript, titled Lingua Mundi: A Semantic Infrastructure for Concept-Grounded Language Models, proposed a seven-layer architecture: a concept graph (language-independent meanings, target size one billion nodes), a synset graph (mapping words of every language onto concepts), and graphs for morphology, architecture: a concept graph (language-independent meanings, target size one billion nodes), a synset graph (mapping words of every language onto concepts), and graphs for morphology, derivation, etymology, cross-language correspondence, and usage — all joined by a common identifier system; plus a concept-grounded model architecture (a Mamba-style sequence model reading concept identifiers), hyperbolic (Poincaré-ball) embeddings for the concept graph, explicit graph-retrieval of facts, and an audited budget showing the whole thing fitting on ordinary laptops , etymology, cross-language correspondence, and usage — all joined by a common identifier system; plus a concept-grounded model architecture (a Mamba-style sequence model reading concept identifiers), hyperbolic (Poincaré-ball) embeddings for the concept graph, explicit graph-retrieval of facts, and an audited budget showing the whole thing fitting on ordinary laptops [H] (manuscript, 1 June 2026).

What makes this episode notable as method — and unusually well documented — is that the author did not write the manuscript alone. On the morning of 1 June 2026 he ran the draft through seven successive revisions (V1–V7) in under half an hour, using two models adversarially: DeepSeek drafted each revision while ChatGPT audited each draft with the rigor of a hostile peer reviewer [H] (conversation logs, DeepSeek "Revisión y corrección del proyecto Lingua Mundi" and ChatGPT "Audit Feedback Lingua Mundi", both 1 June 2026). The audit transcripts are preserved verbatim and are unusually candid about what each round fixed:

• the arithmetic of INT8 embedding storage was corrected (an off-by-a-factor-

of-two error in the storage budget);

• a fabricated empirical claim ("empirically, we find…") was removed and

replaced with an explicitly labeled hypothesis grounded in prior work;

• a comparison against a hypothetical future model was replaced with a

comparison against a real one (DeepSeek-V3);

• the abstract's performance claims were progressively narrowed;
• unsupported scaling claims were re-labeled as open research problems; and
• the manuscript was reframed, by V7, from an empirical claim into an honest

position paper — a research agenda with an explicit falsifiable evaluation plan (word-sense-disambiguation robustness, ablations, baselines).

The audits' final verdicts, preserved in the logs, rated the manuscript's vision highly while consistently grading empirical support low — and the closing recommendation of the last audit was blunt [H]:

"If I were advising the project, I would now pivot from paper-writing to implementation. The next valuable artifact is not V8. It is Lingua Mundi Prototype 0.1 … with: 100k–1M concepts, English only, concept tokenizer, concept reconstruction, TinyStories training run, direct comparison against token-Mamba. That experiment would answer the central question: does semantic compression preserve enough information to justify the bottleneck?"

The author followed that advice. Three observations about this episode matter for understanding everything that followed. First, the idea was never the bottleneck — the author had, in his own words, a vision the audits rated A−; the bottleneck was evidence, and the audits said so seven times. Second, the author's working method — using multiple AI models as adversarial collaborators, keeping every draft and every audit, and preferring honest qualification over confident overclaim — became the studio's permanent standard, and it is the same method visible in the [H]/[O]/[NV] verification conventions of the present paper (Appendix A). Third, the seven-layer architecture of that June manuscript did not vanish when the author pivoted to implementation: it became the project's roadmap. The Lingua Mundi that exists today (September 2026) implements the concept graph, the synset graph, morphology import, corpus layers, and an API; the derivation, etymology, correspondence, and usage layers are explicitly queued phases of the same plan (Sections 5–6, 10.3) [H] (project roadmap, 29 August 2026).

4 · The platform takes shape (July–August 2026)

4.1 Feasibility: "can this be done with open-source components?"

On 18 July 2026 the author asked ChatGPT, point-blank, whether the vision could be built from existing free components and what he would need to develop it [H] (conversation "Lingua Mundi Project Feasibility", 2026-07-18, 00:44 UTC — the evening of 17 July in the author's Central Time zone). The conversation — and the weeks that followed — record the pivot from research paper to engineering program. The project adopted the formal mission statement he pasted into that conversation for review (quoted in Section 3.2), and the author began scaffolding actual code: schema files for the data model exist on disk dated 17 July 2026 [H] (code snapshots, July 2026; see Appendix A).

4.2 The engineering specification (24 July 2026)

On 24 July 2026 the author consolidated the design into a 104-page engineering specification, Multilingual Lexical Platform — Engineering Specification, version 2.0, subtitled "Revised & Production-Ready" [H] (PDF, /home/rex/, 24 July 2026; text extracted for this paper). The specification is the project's constitution, and its contents explain most of what later got built:

• The synset-centric principle. The document states that every lexical

object must ultimately attach to a synset (a meaning), and that the database is organized around meanings rather than spellings. Its worked example is the word bank, which must be one surface form with several distinct synsets ("river bank", "financial institution", "to bank on something") — never one blob of undifferentiated meanings [H].

• Language tiers. The specification planned languages in two tiers:

Tier 1 for launch — English, Spanish, French, Portuguese, Japanese, and Mandarin Chinese (simplified and traditional) — and Tier 2 — Italian, Korean, and Vietnamese [H]. (The set that actually shipped in the database differs in two interesting ways, documented in Section 10: the live six are English, Spanish, French, German, Portuguese, Japanese — Mandarin and Italian are absent, German present. The substitution is nowhere explained in the surviving records [NV], a candid admission of an unresolved provenance question.)

• The data model. The specification's chapter 5 sketches the schema

broadly, and the model that shipped in the repository's first migration is an eleven-table core matching the December 2025 prototype's four-entity intuition, now formalized: languages, broadly, and the model that shipped in the repository's first migration is an eleven-table core matching the December 2025 prototype's four-entity intuition, now formalized: languages, lexemes, senses, concepts (synsets), relations, sources, synsets, and supporting tables (see Section 5)., senses, concepts (synsets), relations, sources, synsets, and supporting tables (see Section 5).

• A five-level A five-level alignment hierarchy. hierarchy. Because automatic and manual linking

of words to concepts differ in reliability, the specification proposed five levels of alignment confidence, from explicit source-stated links (e.g., OMW mappings) down to heuristic guesses — with the rule that each stored link carries its confidence level. This is the mature form of the December 2025 "zero heuristics" rule: not no automatic linking, but labeled, honest linking [H].

• A phase roadmap with gates. Phases 0–5, each with an exit criterion,

from schema and from schema and ETL to API to applications. to API to applications.

• A risk register — including a self-critique annex proposing etymology

and and CJKV/Unihan layers and warning about Japanese–Chinese "false friends" (e.g., 娘 "daughter" in Japanese vs. "mother" in Chinese; 手紙 "letter" vs. "toilet paper") /Unihan layers and warning about Japanese–Chinese "false friends" (e.g., 娘 "daughter" in Japanese vs. "mother" in Chinese; 手紙 "letter" vs. "toilet paper") [H].

The specification also fixed the stack that the project would later simplify: FastAPI, , PostgreSQL, a search engine (, a search engine (Meilisearch in the spec), and a web frontend (Next.js). Notably, by 13 August the author was already questioning that stack himself (Section 4.4).

4.3 The July journal: thinking about the semantic operating system

Around the same days, the author's private journal (knowledge project.txt, 85 KB, read in full for this paper; file timestamp 18 July 2026, a single working session) records a related and more personal thread: his attempt to think through a "universal semantic graph" — a notation-independent way to represent meaning that would work across programming languages, music theory, chemistry, and natural language (a separate project he called USG) [H] (journal, July 2026). The journal's most Lingua-Mundi-relevant passage (lines 2563– 2587) is his reflection on what it would mean to render one semantic object in three languages at once — the same intuition as cat · gato · 猫, written as an exercise in thinking about interfaces rather than languages [H]. Tellingly, the journal's own worked example of a "language" to be explained is the Lingua Mundi codebase's LanguageCreate schema — proof that by mid-July the two threads (the research vision and the actual code) were already woven together in his mind [H].

4.4 August: from schema to running system

August 2026 is when Lingua Mundi became a system. The surviving logs show a dense sequence of decisions, each leaving its trace:

• 1–8 August. The author explored the graph and knowledge-representation

options ("JSON as a Graph", 29 July; API monetization ideas, 7–12 August) and — critically for the final architecture — asked whether the API "really needs" a heavy stack (Section 4.5) [H].

• 11 August. He explored integrating Maya languages into Lingua

Mundi ("Lingua Mundi Maya integration"), continuing the interest in indigenous Mexican languages first voiced in the December 2025 message [H].

• 13 August. Two design conversations settled the shape of the

product: "Linguistic API Tech Stack" (see 4.5) and "Semantic Interaction Layer", which sketched the future interface between the knowledge layer and applications [H].

• 19–25 August. He built the Japanese demonstration: a compact

offline database (SQLite) bundling Japanese dictionary data so that the concept-oriented design could be touched, queried, and shown to others ("Build Japanese Demo"; the resulting lingua-mundi-jpn.sqlite files are preserved on disk) [H]. The sprint had a concrete deadline and audience: representatives of the Japan Foundation were due at his university the following week, and the demo (together with the furigana editor that was being renamed Shikibu in the same period) was prepared as a leave-behind product dossier for that visit [H] (conversation log, 19–25 August 2026).

• 23 August. Asked for a one-paragraph description of the concept, he

heard back a summary that matches what the logs show him building: "a machine-readable linguistic infrastructure for human languages… rather than being just a Japanese dictionary or translation app, Lingua Mundi is intended to become a general linguistic intelligence layer" [H] — and, in the same conversation, a technical description naming the then-planned stack (PostgreSQL, Apache AGE for graph queries, OpenSearch for retrieval, for retrieval, FastAPI as the API) that the following month's measurements would render unnecessary (Section 4.5, Section 6) as the API) that the following month's measurements would render unnecessary (Section 4.5, Section 6) [H].

• 24 August onward. The codebase moved into its permanent home as a

proper repository ("LinguaMundi") with the daily GitHub-backup snapshot trail beginning 27 August [H] (repository snapshots; see Appendix A).

4.5 The lean-stack decision that shaped everything

The most consequential technical decision of August is preserved verbatim. On 13 August 2026 the author pasted his own planned technology stack — FastAPI, Polars, DuckDB, PostgreSQL, Apache AGE (a graph extension), OpenSearch, Redis, Celery, React, and Docker — and asked: "Does the linguistic API really need this stack if it already runs on postgres?" [H] (conversation "Linguistic API Tech Stack", 13 August 2026).

The answer he received — and adopted — was that almost none of it was needed yet: PostgreSQL alone could serve relational dictionary data, kanji, readings, conjugations, translations, The answer he received — and adopted — was that almost none of it was needed yet: PostgreSQL alone could serve relational dictionary data, kanji, readings, conjugations, translations, full-text search, and even graph-style recursive queries, and the correct rule was "PostgreSQL limitation → add technology," not "Phase 3 says graph database → install graph database" , and even graph-style recursive queries, and the correct rule was "PostgreSQL limitation → add technology," not "Phase 3 says graph database → install graph database" [H] (same conversation). The advice concluded: "Your API's competitive advantage isn't infrastructure. It's the linguistic data model and transformations you've built on top of the data" [H].

This decision — one relational database, one API framework, and nothing else until measured need — is the direct ancestor of the performance architecture described in Sections 6–7. When, in September 2026, the database grew to 1.6 million word senses and naive queries slowed to minutes, the author did not reach for a graph database or a search engine: he reached for PostgreSQL materialized views — the " — the "bake, don't re-derive" pattern — and reduced the slowest queries from over a minute to about one second on hardware from 2014 (Section 7). The August decision to stay lean is what made that September optimization both necessary and possible., don't re-derive" pattern — and reduced the slowest queries from over a minute to about one second on hardware from 2014 (Section 7). The August decision to stay lean is what made that September optimization both necessary and possible.

4.6 The human-AI development method, in practice

A reader of this history will have noticed that nearly every milestone is accompanied by a conversation log: the author thinks in dialogue. By August 2026 his method had stabilized into something that can be described precisely from the record:

1. The human sets the direction and owns the data. Every request in the

logs is the author's own; every file is on his machine; every decision — including reversals like the lean-stack question — is his.

1. AI models are used as specialized collaborators. ChatGPT was used

heavily for research, architecture critique, and audit; DeepSeek for drafting and revision cycles and (later) as the harness underlying the studio's own automation; Claude for long-horizon memory and project context (the studio's knowledge base began as a Claude memory export).

1. The record is kept. Conversations are exported (the author's archives

contain thousands), code is snapshotted, and — from 26 August 2026 — every working session is logged in an Obsidian knowledge base with frontmatter, dates, and verification status (Section 9). This paper is possible only because of that record-keeping habit.

1. Adversarial review is built in. The June manuscript's V1–V7 audit

loop (Section 3.3) became a standing practice: drafts are audited by a different model than the one that wrote them, and the audits are kept on file.

Whether one calls this "vibe coding," "AI-assisted development," or simply "using the tools," the honest label matters for the record: Lingua Mundi is a human-authored project built in deliberate, documented collaboration with AI assistants, and this paper (itself an AI-generated draft under the author's direction, per the front-matter notice) is an artifact of the same method.

5 · Design: the linguistic and mathematical principles of the model

This section explains, in plain language, the ideas that determine how Lingua Mundi stores language. Readers who want the history may skip to Section 6; readers who want the formulas and schema will find them here and in Appendix A.

5.1 Four kinds of objects, one space of meanings

The data model descends directly from the December 2025 pipeline's four entity types (Section 1.2) and from the July 2026 specification's eleven tables [H]. In its mature form it stores five core kinds of things (verified against the live schema, September 2026 [H]):

• Language — one row per human language, keyed by its ISO 639-3

three-letter code (eng, spa, jpn, …), with its name, , …), with its name, autonym (the name speakers use), language family, and writing script. (the name speakers use), language family, and , …), with its name, autonym (the name speakers use), language family, and writing script. . Adding a language to the whole system is, by design, adding one row plus datasets: there are no per-language tables anywhere [H] (code and project note, 29 August 2026). This "language-agnostic" property is the schema-level expression of the interlingua idea (Section 2.1).

• Lexeme — a word form in a specific language: the string cat, the

string gato, the string 猫. A lexeme belongs to exactly one language and carries a part of speech.

• Sense — one meaning of one lexeme as recorded by one dictionary:

"cat" the animal sense from WordNet, "cat" the jazz-musician sense from WordNet, "cat" the yule-log sense from Wiktionary. A lexeme typically has many senses — this is how the model represents polysemy (one word, many meanings) without ever collapsing meanings together. A sense may or may not yet be linked to a concept; the link is the field concept_id, and it is nullable by design — an unlinked sense is kept and shown as an orphan rather than guessed into place [H] (schema; pipeline notes, December 2025).

• Concept — a meaning, shared across languages: the concept

"domestic cat" identified by the Princeton WordNet synset "domestic cat" identified by the Princeton WordNet synset URI princeton-wordnet:02121620-n [H] (live database). Concepts are where the multilingual alignment happens: Spanish gato (one of its senses), English cat, French chat, German Katze, and Japanese 猫 can all point at the same concept — they are "one star" with many names (Section 8). The concept table is the heart of the system and, in the current database, holds 96,434 rows [H] (measurement, 6–7 September 2026).

• Relation — a typed, directed connection between two concepts,

e.g., domestic-cat IS_A feline, or feline IS_A carnivore. Relations carry a predicate, a confidence, and a source. The current database holds 171,672 of them [H] (measurement, 6–7 September 2026).

Around these five sit provenance machinery — a source table recording where every dataset came from, its license, and its URL (so that no claim in the database is ever unattributed), synset tables bridging to the OMW format, and supporting tables for kanji and readings [H] (schema, September 2026).

5.2 The alignment problem, and why it is not faked

The central intellectual problem of a multilingual map of meaning is alignment: deciding that this sense of this word in this language is the same meaning as that sense of that word in that language. Lingua Mundi's answer, inherited from the OMW design, is to not decide it fresh: the OMW already states, per language, which of its synsets correspond to Princeton WordNet synsets, so the database imports those explicit mappings as its primary alignment layer [H] (Bond & Foster, 2013; import code). Where no explicit mapping exists, the system applies its five-level confidence hierarchy (from the July 2026 specification): explicit source-stated links rank highest; inferred links rank lowest and are stored with their level — never silently promoted to "fact" [H] (specification, 24 July 2026). This is the mature form of the December rule the author scribbled in his pipeline code: "zero heuristics." It is not that the system never guesses; it is that guesses are labeled as guesses, which is what makes the database honest enough to grow: a wrong guess that is marked as a guess can be found and fixed; a wrong guess stored as fact poisons everything downstream.

The same philosophy governs the granularity of languages. ISO 639-3 distinguishes real languages, not political boundaries or writing-system variants. The author has flagged one genuinely hard case for the future: Korean, whose single ISO code kor covers both the standard of the Republic of Korea and the standard of the Democratic People's Republic of Korea. The surviving records show this flagged as an open modeling decision [H] (session log, 7 September 2026) — an example of the schema's principle that it stores languages, not locales, and a reminder that real-world language data resists neat one-row-per-language boxes.

5.3 The graph: taxonomies, neighborhoods, and the long tail

Once senses are aligned to concepts and relations connect concepts, the database is a graph: 96,434 meaning-nodes joined by 171,672 typed edges, the largest component being the is-a taxonomy inherited from WordNet (mammal → carnivore → feline → domestic cat), enriched by taxonomy inherited from WordNet (mammal → carnivore → feline → domestic cat), enriched by instance-of, part-of, and cross-language relations , part-of, and cross-language relations [H] (measurement, September 2026; schema). Three mathematical facts about such graphs shape the user interface and the engineering:

1. Taxonomies are trees of meaning. Because every concept's parents are

known, the system can answer "what is this a kind of?" (climb up) and "what are the kinds of this?" (climb down) — the operations behind the interface's taxonomy lens (Section 8).

1. Degree distributions are Zipfian. Real language graphs are

scale-free-ish: a few concepts (like "animal", "person", "thing") have enormous neighborhoods while most concepts have a handful of neighbors. This ": a few concepts (like "animal", "person", "thing") have enormous neighborhoods while most concepts have a handful of neighbors. This "long tail" is the same statistical shape Zipf famously measured in word frequencies (" is the same statistical shape : a few concepts (like "animal", "person", "thing") have enormous neighborhoods while most concepts have a handful of neighbors. This "long tail" is the same statistical shape Zipf famously measured in word frequencies ( famously measured in word frequencies (Zipf, 1949): a small number of words account for most usage. Two design consequences follow: ranking must exist (you cannot show 30,000 meanings of one query), and the ranking must be meaningful — the interface therefore ranks concepts by their sky brightness, a proxy for how "important" or well-connected a meaning is (Section 5.5).

1. Shortest paths are the natural translation route. Given two concepts,

the graph can be searched for a connecting path — the operation behind the interface's sextant tool (Section 8.4), implemented as a bounded bidirectional the graph can be searched for a connecting path — the operation behind the interface's sextant tool (Section 8.4), implemented as a bounded bidirectional breadth-first search over the taxonomy (≤ 10 hops, 25 neighbors per node, answering in milliseconds) over the taxonomy (≤ 10 hops, 25 neighbors per node, answering in milliseconds) [H] (code, live measurement, September 2026).

5.4 Zipf and the shape of dictionaries

Zipf's law deserves one explicit mention because it quietly explains the whole performance architecture. If word frequencies follow a power law, then any dictionary is dominated by a short head of common words and an astronomically long tail of rare ones. The June 2026 manuscript even derived a Zipf-based coverage calculation for concept vocabularies [H] (manuscript V7 audit, 1 June 2026). For the database, the long tail means: most rows are rare words, most queries hit common words, and any query that must scan the whole table — including all the rare words — is catastrophically slower than a query that can start from a precomputed ranking. The materialized-view pattern of Section 7 is, at bottom, a way of turning Zipf's law from a performance enemy into a design tool: bake the head (the ranking, the top words per concept) once, and let the tail stay in the relational tables for the rare lookups that actually need it.

5.5 Sky brightness: one ranking to rule the map

The interface's organizing metaphor is a star map (Section 8), and the metaphor is backed by an actual computation. Each concept is assigned a brightness — a numeric magnitude used to sort the sky — derived from how it stands in the multilingual graph: how many languages have words attached, how rich those word sets are, and how prominent the concept is in the taxonomy. The database materializes this as concept_sky_rank, a precomputed ordering of all 96,000+ concepts, so that "give me the sky" never requires sorting a million rows at request time [H] (scripts and measurement, 6 September 2026). Two companion views — concept_lang_rank (per-language rankings) and concept_word_rank (the top four words per concept per language, "baked" so that enriching a star with its words is a lookup, not a join over 1.6 million senses) — complete the ranking system [H] (scripts, 6 September 2026; row counts in Appendix A).

5.6 Bounded work: the game-designer's discipline applied to a database

The final design principle is borrowed from video games. A game renders a frame in 16 milliseconds no matter how large the world is: it budgets work per frame. Lingua Mundi applies the same discipline — bounded work per query on an unbounded data space — because its hardware demands it (a 2014-era laptop; Section 9). No request may scan the whole corpus; every request must start from an — because its hardware demands it (a 2014-era laptop; Section 9). No request may scan the whole corpus; every request must start from an index or a baked view; the interface must render only when the user does something (idle CPU ≈ 0%); and results are paginated with stable windows. Sections 6 and 7 show the three performance walls this discipline hit and how each was solved with the same pattern: or a baked view; the interface must render only when the user does something (idle CPU ≈ 0%); and results are paginated with stable windows. Sections 6 and 7 show the three performance walls this discipline hit and how each was solved with the same pattern: bake, don't re-derive.

6 · Architecture: importing the world, serving the map

6.1 The pipeline: one registry, twenty-plus importers

Every dataset enters Lingua Mundi through an importer — a program that knows how to read one external format and write it into the canonical schema. The single authoritative list of importers lives in one file, the importer — a program that knows how to read one external format and write it into the canonical schema. The single authoritative list of importers lives in one file, the importer registry; by September 2026 it contained more than twenty registered plugins covering: Princeton WordNet and the English WordNet, the Open Multilingual Wordnet (OMW), Wiktionary and its machine-readable derivative ; by September 2026 it contained more than twenty registered plugins covering: Princeton WordNet and the English WordNet, the Open Multilingual Wordnet (OMW), Wiktionary and its machine-readable derivative kaikki, , Wikidata, UniMorph and , UniMorph and J-UniMorph (morphology), Universal Dependencies treebanks (UD), OpenThesaurus (German synonyms), the CMU pronouncing dictionary, IDS character decompositions, CLICS cross-linguistic data, word-frequency lists, and the Japanese family — JMdict, EJDict, KANJIDIC2 [H] (importer registry and project notes, accessed 6–7 September 2026). (One caveat for precision: a folder for (importer registry and project notes, accessed 6–7 September 2026). (One caveat for precision: a folder for OPUS parallel corpora exists in the importers directory, but it is an empty, unregistered stub awaiting the roadmap's sentence layer — it is listed here for honesty, not as a shipped source parallel corpora exists in the importers directory, but it is an empty, unregistered stub awaiting the roadmap's sentence layer — it is listed here for honesty, not as a shipped source [H] (code inventory, 6–7 September 2026).) Each importer registers its target languages against the same ISO codes, which is what keeps the schema language-agnostic in practice, not just in theory [H] (code).

Imports are supervised by a pipeline that runs each source, records the import job, and — crucially for Section 7 — reports success. After each import it refreshes the ranking views (Section 5.5) so the "sky" always reflects the freshest data [H] (scripts, September 2026).

6.2 The corpus today (measured)

Direct Direct SQL measurement of the live database on 6 September 2026 at 23:02 Central Time, re-verified 7 September 2026 (both counts identical) measurement of the live database on 6 September 2026 at 23:02 Central Time, re-verified 7 September 2026 (both counts identical) [H]:

Per-language lexical weight (enrichment audit, 6 September 2026 [H]):

Two honest caveats belong next to these numbers. First, per-language figures are approximate (rounded from the audit); the table-level SQL counts are exact at their measurement instant. Second, the corpus is layered: some datasets (: some datasets (kaikki Wiktionary and word-frequency lists) were deliberately capped at 50,000 rows per import run, with deeper coverage supplied by other sources, and that cap is documented rather than hidden Wiktionary and word-frequency lists) were deliberately capped at 50,000 rows per import run, with deeper coverage supplied by other sources, and that cap is documented rather than hidden [H] (audit, 6 September 2026). Section 7 shows what happens when a similar cap was not documented as a risk.

6.3 The API surface

The system exposes a small, deliberately frozen The system exposes a small, deliberately frozen HTTP API (FastAPI), whose Phase-0 surface is documented as never breaking as new phases arrive API (FastAPI), whose Phase-0 surface is documented as never breaking as new phases arrive [H] (project note, 29 August 2026). The endpoints, verified live in September 2026 [H]:

• /languages — list the live languages and their metadata;
• /lexemes and /concepts — paginated list queries (limit/offset, capped

at 1,000 per page after an August–September audit found them unbounded);

• /analyze — morphological/lexical analysis of a word;
• /kanji and /lookup — the Japanese kanji and dictionary lookups;
• /graph/sky — the star-map feed: concepts ranked by sky brightness, with

parameters for search (query, lang), depth (skip/tier), whether to attach member words (words), and a taxonomy lens (anchor, lens=up|down) that climbs or descends the is-a tree from a given concept;

• /graph/star/{id} — one star in full: its member words per language and

its neighbor concepts, both paginated;

• /graph/path — the bounded graph search (Section 5.3) between two

concepts;

• /graph/word — the "word lens": given a word and a language, the concepts

its senses currently link to.

All responses are compressed (All responses are compressed (gzip), and every query is designed to run from an index or a materialized view — the "bounded work" rule of Section 5.6 ), and every query is designed to run from an index or a materialized view — the "bounded work" rule of Section 5.6 [H] (code and live headers, September 2026).

6.4 Why the design is language-agnostic in the database, not just on paper

The reader can now see the whole loop. A new language arrives as one language row (say, Mandarin, code cmn) plus registered datasets (the OMW Mandarin wordnet, a UD Mandarin treebank). Importers write its lexemes, senses, and — where the OMW or other explicit mappings exist — its sense→concept links. The ranking views recompute; the sky gains a new language; no table changes; no migration; no special case in the API. The next family planned along this exact path — Mandarin, Cantonese, Korean, and Vietnamese — is described in Section 10.3, where the data-availability reconnaissance that will decide its schedule is recorded [H] (session log, 7 September 2026). The design's promise — "add languages indefinitely" — is thus not marketing: it is the direct consequence of the schema.

7 · Two infrastructure lessons: bake-don't-re-derive, and fail-loud

Every database project eventually meets its performance walls. Lingua Mundi met three of them in one week — and then discovered that the most interesting problem was not speed but silence. This section tells both stories, because together they are the project's most transferable engineering lessons.

7.1 The pattern: bake, don't re-derive

The rule that solved all three performance walls is simple enough to state in a sentence: when many queries need the same expensive computation, compute it once, store the result, and keep it fresh — do not recompute it for every query. In database terms, Lingua Mundi uses PostgreSQL materialized views: precomputed tables that act like indexes full of answers.

• Wall 1 — ranking the sky. Originally, asking for the default star map

forced the database to re-rank the entire corpus per request: about 8.6 seconds each time on the project's 2014-era laptop. The fix baked the ranking into concept_sky_rank and concept_lang_rank (94,498 and 218,120 precomputed rows), after which the default sky answered in ~0.25 s and deeper requests in tens of milliseconds [H] (measured, 6 September 2026).

• Wall 2 — naming the stars. Asking for the sky with its words

attached (top words per concept per language) required joining across 1.6 million senses on every request; at 1,000 stars the query took more than a minute — a pathological case. The fix baked the words themselves: concept_word_rank precomputes the top four words per concept per language (338,378 rows at creation), turning a minute-plus join into a ~1.1 s lookup, and the default view into ~0.25 s [H] (measured, 6 September 2026).

• Wall 3 — hierarchy and Wall 3 — hierarchy and pagination.. Climbing and descending the

taxonomy (lens up/down) and paging through stable windows of stars were solved with canonical predicates, index ordering, and stable pagination windows (DISTINCT ON-style queries) rather than a graph database — the payoff of the August "lean stack" decision (Section 4.5) [H] (code and measurements, 6 September 2026).

After each import the views refresh automatically, so "baked" never means "stale" [H] (scripts). The whole pattern is a database-native instance of what game developers call budgeting: the expensive work happens once, off the critical path, and the user's request only ever reads precomputed answers.

7.2 The silent truncation that hid cat

In late August 2026 the English layer of the corpus was imported with a row cap that no one intended to be a cap at all. The supervising import script limited the English Open Multilingual Wordnet import to 50,000 candidates — and then reported success [H] (session log, 6 September 2026, root-cause entry). English cat is candidate number 52,195 in the file's alphabetical order. It was cut off. Roughly 21,830 of the 117,659 English synsets were never created; for those meanings, English senses existed as raw dictionary entries but were never linked to concepts. The word cat was not missing from the database — it was invisible to the map: a search for its meaning could not find it, because the meaning-node its senses should have pointed to had never been built [H] (same source).

For about six days, nobody knew. The demo the author cared about — cat · gato · 猫 on one star — worked in Spanish and Japanese but came up empty in English, and the empty result looked like a data gap — as if English lacked the word. (A timing note from the record: the multi-language sky itself only arrived when integrate_core6.py finished linking the imported languages near the end of the August–September import run — before that the fallback sky was a small single-language view — which is why the English hole surfaced only in early September, once the shared-concept view was actually live [H] (session log, 6 September 2026).) The author's own insistence, recorded in the session log, was that this was "not a data gap — infrastructure" [H] (session log, 6 September 2026). He was right. The failure was not in the data but in the pipeline's honesty: a truncated import had reported success, and nothing in the system was designed to notice that "success" with a 50,000-row cap meant failure for candidate 52,195.

7.3 The fix: an uncapped, idempotent repair chain

The repair — designed and launched on 6–7 September 2026 and verified complete on 8 September 2026 (see Section 10.1) — is a model of how to fix a data pipeline without losing data or trust [H] (scripts and session log, 6–7 September 2026):

1. Re-import the English WordNet without the cap — idempotently, so an

interrupted run can simply be relaunched without duplicating rows;

1. re-run the synset-to-synset relations for the newly created synsets;
2. integrate: promote new synsets to concepts, link member senses'

concept_id fields, and re-express concept relations;

1. refresh the materialized views so the map reflects the repaired

corpus; and

1. verify end-to-end with an acceptance script that proves, from the

database and through the live API, that English cat reaches the domestic-cat concept princeton-wordnet:02121620-n together with its multilingual members — gato, chat, Katze, 猫 — and that the concept has taxonomy parents. The acceptance check is designed to be re-run at any time and to exit with a loud PASS or FAIL.

Because every step was written to be idempotent (get-or-create, skip-if-exists), a machine crash mid-repair costs nothing but time: the chain can be relaunched and will finish the job. The same property protects the live database from partial states.

7.4 The lesson: fail-loud vs. fail-silent

The cat episode is a textbook case of a problem every data engineer fears: a pipeline that fails silently and reports success. The methodological lesson — that a row cap is not a success criterion, and that importers must verify completeness, not just completion — is worth stating as a general principle for open-data infrastructure, where datasets arrive with no guarantees and pipelines are assembled by hand. Lingua Mundi's response was institutionalized in two forms: the acceptance script that any future import chain can run, and the studio's standing rule that unverified claims are labeled unverified — in the database ( — in the database (orphan senses kept visible, confidence levels on links) and in every written record (the kept visible, confidence levels on links) and in every written record (the [H]/[O]/[NV] convention of this paper, Appendix A). A system that can show its orphans is a system whose gaps can be found. A system that hides them is a system whose gaps will be found — eventually, by the worst possible person: a user looking for cat.

8 · The interface: a star map of meaning

Data alone is not a product. The question that shaped Lingua Mundi's interface was simple: what should looking at a map of meaning feel like? The author's answer, developed through the first week of September 2026, was that it should feel like looking at a sky — and, specifically, like the chart of a fictional treasure planet: a navigable paper map of a universe where each star is one idea, and languages are the telescopes you choose to view it through [H] (session logs, 1–7 September 2026). This section describes the interface that resulted: its pedagogical metaphor, its visual language, its instruments, and the engineering discipline that lets all of it run on a 2014 laptop.

8.1 From dictionary page to sky: naming the product

The interface went through three identities in two weeks, and the names track the product's evolution [H] (session logs, 31 August – 6 September 2026). It began as an ordinary dictionary page ("/gui/", called the Observatory on 31 August), became the Planetarium on 1 September as the data imports finished and a star-field visualization appeared, and — after the author looked at the result and pronounced the stars "scaffolding, not the product" — pivoted on the night of 6 September into a semantic-graph explorer: the same sky, but now the point was the meaning, with nodes that are concepts labelled in the languages you enable, edges colored by relation family, and search that disambiguates senses. The older sky view was preserved but unlinked; the explorer became the default face of Lingua Mundi [H] (session log, 6 September 2026).

8.2 The pedagogical metaphor: an instrument, not a dashboard

Every element of the interface maps one-to-one onto a concept from astronomy and from the language project itself — a deliberate pedagogical design (the metaphor table is the product's internal documentation) [H] (session logs, 6–7 September 2026):

None of this is decoration; the author's design brief for the tool is quoted in the records and worth repeating: the interface should be "a real discovery tool, not a toy" — a map of all languages that a learner or researcher can genuinely explore — and its sources should behave like a navigable paper map, not a dashboard [H] (session logs, 6 September 2026). In the same spirit the instrument panels are deliberately skeuomorphic: brass-knob exposure dials, sextant measurements, charted courses, an epoch stamp — the aesthetic of an old observatory, chosen to teach by feel rather than by menu labels.

8.3 What is on screen

The explorer, verified end-to-end in September 2026 [H], provides:

• Search with sense disambiguation. Typing gato (Spanish) returns

meaning-stars; each result can show which words in which languages share the meaning — the panel for the domestic-cat star displays its multilingual synset members, e.g. English cat and grimalkin, German Katze, Japanese 猫 (Section 8.5).

• Travel. Click a star (or shift-click / press the measure key) to make

it the focus; the sky re-centers on it and its neighborhood appears.

• The star panel. The focused star's members per language, its

definition, and its neighbors — with pagination for large neighborhoods.

• The taxonomy lens. Climb to a star's parents ("what is this a kind

of?") or descend to its children ("what kinds of this exist?") along the canonical is-a-family predicates.

• The instruments (all live, verified): the exposure dial widens the

focused star's field in steps (90 → 180 → 360 → 720 → 1,000 stars), merging the wider sky without re-travelling; the sextant measures a bounded graph path between two stars (the /graph/path search of Section 5.3), drawing a dashed gold course through the sky and logging each leg; skymarks pin stars (persisted locally, so your own mini-constellation survives reloads); the word compass rings the focused star with word chips — one per enabled language, clickable to switch the search language; and the course log, home bearing, epoch stamp, and reticle keep you oriented [H] (session log, 7 September 2026).

• The support affordance. A "✦ support" control opens the donation

modal (Section 10).

8.4 Engineering the sky on 2014 hardware

The interface's beauty is subordinate to one hard constraint: it must run smoothly on a machine with a 2014-era processor, 8 GB of RAM, and a spinning disk — the same machine that hosts the database. The engineering that makes this possible is worth documenting because it is the same bounded work discipline as Section 7, applied to pixels [H] (audit and session logs, 6–7 September 2026):

• The page renders only on input. Idle CPU usage is ~0%; nothing

animates unless the user does something. There is no perpetual animation — a deliberate choice under the "potato budget" (Section 9).

• The nebula is baked. The decorative background (purple/teal/rose

blooms and dust) is painted once per resize and then cached, with camera parallax on redraw, instead of being recomputed per frame.

• Labels are rationed. Only ~24 star labels are drawn per frame — a

label budget — keeping redraws cheap and text legible [H] (the interface code bounds placements with MAX_LABELS = 24; gui/planetarium-v1.html).

• Requests are protected. Stale-response tokens and abort controls

cancel in-flight queries when the user moves on; double-clicks defer single-click actions so no wasted fetch fires; resizes are debounced and the canvas pixel ratio is capped (shrinking a resize's memory from ~24 MB to ~11 MB at 1080p).

• Accessibility is explicit: canvas roles and aria labels, reduced-motion

support, and keyboard travel were added in the same pass as the visual polish [H] (session log, 7 September 2026).

8.5 The demo that ties it together

The project's running demonstration is the word cat — or rather the concept it belongs to. Search Spanish gato in the explorer, travel to the star, and the panel shows the synset princeton-wordnet:02121620-n with its members across the live languages. The subset actually verified end-to-end in the headless checks of 6–7 September 2026 is recorded in the test log — English cat, grimalkin and felis silvestris catus (the scientific name WordNet stores), German Katze, Spanish gato, Japanese 猫 — and the wider member lists the database stores for that synset (e.g. English housecat, mouser, puss, tomcat; French chat, chatte; German Kater; Japanese ネコ; Portuguese gato, bichano) come from the imported wordnet membership rather than from the end-to-end log; a reader who wants the complete current enumeration can re-run the acceptance query of Section 7.3 [H with [NV] on the untested tail] (headless E2E log and live database, 6–7 September 2026). One star, six languages, one meaning: the December 2025 intuition, visible on screen. The English side of this demonstration is exactly what the Section 7.2 bug briefly broke: a Spanish-driven gato search reached the star all along, while an English-driven cat search only works after the coverage repair, and the acceptance script of Section 7.3 exists precisely to prove, on demand, that cat · gato · 猫 share their star.

9 · The workbench: tools, workflow, and the "potato" machine

A project history is incomplete without the workbench it was built on. Lingua Mundi's workbench is unusual, and the constraints it imposed shaped nearly every engineering decision in this paper.

9.1 The potato

The machine that hosts the database, the API, the interface, and the development environment is a single consumer laptop: an ASUS X751MA with an Intel Pentium N3540 (Bay-Trail, four cores, no AVX2), 8 GB of RAM (7.6 GiB usable), a 5400 rpm hard disk, and an internet uplink of roughly 47 KB/s — captured live for this paper on 7 September 2026 ([H], lscpu//proc/meminfo on the machine itself). The project's notes describe this same machine variously as an Atom Z36xx/Z37xx or a Pentium N3540 — the same Bay-Trail silicon family under different names [O]; earlier 2026 chat records also reference a different profile (an Intel Jasper Lake "N5105" machine), which Section 10.2 lists as an unresolved hardware-record conflict. The studio calls the live box, with affection, the potato — and it set a rule that appears throughout the records: one heavy operation at a time, memory-guarded wrappers, no perpetual animation, no premature infrastructure (Section 4.5), and every heavy job run as a self-alerting detached script that reports when it is done rather than being polled [H] (session logs and tooling, September 2026). The "bake, don't re-derive" pattern (Section 7), the input-only interface (Section 8.4), and the zero-LLM-cost research surface (Section 9.4) are all, at bottom, consequences of the potato — proof that a tight budget can be a design virtue rather than a limitation.

9.2 The software stack

The running system (verified live, September 2026 [H]): PostgreSQL 16 as the single datastore (Section 4.5), FastAPI (Python) as the API layer, SQLAlchemy + Alembic for models and migrations, Python scripts for the importers and materialized-view management, and a small dependency-free web interface (HTML/CSS/JavaScript) served from the same origin. Auxiliary machinery includes systemd user services that keep the API and the memory/thermal guard alive (Section 9.5), a daily GitHub backup of the codebase, an Obsidian knowledge base as the project memory (Section 9.3), headless-Chromium end-to-end checks for the interface, and the studio's detach-run + notify tooling for self-alerting background jobs [H] (services and tooling inventory, September 2026).

9.3 The knowledge base: a project memory that can be audited

From 26 August 2026 the author and his agents maintained a structured knowledge base (an Obsidian vault) in which every working session is logged: work done, decisions made, next actions, and the file where each fact came from, with frontmatter recording provenance and a verified: true/false flag [H] (vault conventions, August–September 2026). The vault is the reason this paper can cite dates and quotes at all: when this draft says "session log, 6 September 2026", it is pointing at a file that exists, was written that day, and carries its own source notes. Two conventions of that vault deserve mention as method: every claim is supposed to state its source, and history is never invented — unconfirmed material is kept but flagged unverified [H] (vault conventions). Those conventions are this paper's conventions (Appendix A).

9.4 Zero-token research and the local-first rule

The author's standing rule — visible in the project note of 29 August 2026 — is that the local API is the research surface: /lexemes, /concepts, /analyze answer questions about the corpus with zero LLM tokens spent [H]. The same local-first instinct governs data (everything self-hosted; nothing depends on a vendor), money (free tiers and open data only; Section 10), and privacy (secrets never leave the machine; the knowledge base is private) [H] (session logs, September 2026).

9.5 Guardrails on a fragile machine

Because the potato has no headroom, the studio built guardrails that are themselves part of the engineering story: a memory/thermal guard (tools/mem-manager.sh, v2, live 7 September 2026) that warns below 900 MiB available memory or 80 °C and sheds optional background units below 500 MiB or 88 °C, plus a system-level early-oom killer (earlyoom -m 8 -s 5, i.e. it acts when free memory collapses below 8%, with a 5-second cadence) [H] (live configuration: mem-manager.sh thresholds and /etc/default/earlyoom, read 7 September 2026). The studio's operating rules keep those guardrails from ever having to act: one heavy operation at a time, memory-guarded wrappers, and self-alerting detached jobs for everything long-running. Long repairs are written to be idempotent (get-or-create, skip-if-exists) so that any interruption — a reboot, a power loss, a guard trip — costs nothing but time: the repair can simply be relaunched and will finish the job (Section 7.3). The same discipline treats concurrency itself as a resource to be budgeted, like memory or disk.

9.6 Governance: licenses, attribution, and the "free forever" decision

Every dataset in Lingua Mundi entered with its license recorded (the repository's docs/DATA_LICENSES.md inventory; the live database's source table) [H]: WordNet (permissive), the Open Multilingual Wordnet (per-language licenses, predominantly CC BY-SA and CC BY), Wiktionary and kaikki (CC BY-SA, with GFDL legacy terms), Wikidata (CC0), UniMorph (CC BY-SA 3.0), Universal Dependencies (CC BY-SA 4.0), JMdict/KANJIDIC2 and the EDRDG family (permissive, with attribution), OPUS (per-dataset), CLICS and word-frequency lists (per-source), and so on [H] (license inventory, accessed September 2026). The project reciprocates: its own outputs are planned to be free as well.

The funding model evolved publicly across nine days in late August/early September 2026, and the record shows each step [H] (session logs and calendar, 29 August – 7 September 2026): an initial plan to sell an offline Core-6 dictionary app (Texupan, $29 or $9/month) on 29 August; a store listing with payment links on the store hub by 1 September; and — by 7 September — a decision to make Lingua Mundi free forever, funded by voluntary donations, with the Stripe product literally named "Lingua Mundi — keep it free forever", one-time prices of €5/€15/€50 and monthly €3/€9/€25, and an honest annual goal of €2,400 to cover infrastructure and the next language's data work [H] (session log, 7 September 2026). The account was deliberately set up dormant — donations disabled until the author completes the payment provider's identity verification — so that "free forever" is, in the author's framing, a verifiable claim, not a slogan [H] (session log, 7 September 2026). Public-facing text follows the studio-voice rule (front matter) and the honest-copy rule: no hype, no overclaim — the same discipline that governs the database and this paper.

10 · Where the project stands (7 September 2026), its honest limits, and its road ahead

10.1 Current state of the system

Measured directly against the live system on 6–7 September 2026 [H] (see Appendix A for the exact queries):

• Corpus (pre-repair snapshot, 6–7 September 2026): 96,434 concepts · 1,238,383 lexemes · 1,608,018 senses · 171,672 relations, in six live languages (English, Spanish, French, German, Portuguese, Japanese) — the SQL counts reproduced identically on both measurement dates.
• Corpus (post-repair, 8 September 2026): 125980 concepts / 125980 synsets / 352333 English lexemes, as recorded in the repair report (lm-en-coverage-REPORT-2026-09-08.txt, RESULT: PASS); the multilingual members of the cat star princeton-wordnet:02121620-n span English, Spanish, Japanese, German and Portuguese (cat, gato, 猫, Katze, Gato doméstico).
• Ranking views: concept_sky_rank 94,498 rows; concept_lang_rank

218,120 rows; concept_word_rank 338,378 rows at creation (a live re-count was denied to the standard database role; flagged [NV]).

• Performance: default sky ≈ 0.25 s warm (8.6 s before the ranking

views); deep 1,000-star query with words ≈ 1.1 s (>60 s before the words view); graph-path searches in milliseconds; search ≈ 0.5 s warm / ≈ 10 s cold on the hard disk's page cache.

• Interface: semantic-graph explorer live at the API's web origin, with

all instruments (exposure dial, sextant, skymarks, word compass, course log, epoch stamp) verified by 23/23 headless-browser end-to-end checks on 6–7 September 2026.

• English coverage repair: the uncapped English WordNet re-import, relations, integration, view refresh, and end-to-end verification chain (Section 7.3) was launched on the morning of 7 September 2026 and ran to completion on 8 September 2026 at 22:19 (2026-09-08 22:19:12). Its acceptance gate passed: the chain's own report (lm-en-coverage-REPORT-2026-09-08.txt, RESULT: PASS) records the concept-linked English cat sense on princeton-wordnet:02121620-n, whose multilingual members include cat, gato, 猫, Katze, Gato doméstico [H] (report, 8 September 2026).
• Governance: donation channel live but intentionally dormant pending

the author's payment-provider verification; codebase backed up daily; knowledge base current.

10.2 What is honestly not known, and what is unresolved

The studio's verification conventions require this paper to say plainly what it could not verify, and to record contradictions between sources. The important ones:

1. The "Core 6" discrepancy. The author's personal language set (recorded

29 August 2026) is Spanish, English, Japanese, French, Italian, and Portuguese. The live database's six languages are English, Spanish, French, German, Portuguese, and Japanese. German replaced Italian in the database with no surviving written explanation; the Italian language row and Italian datasets were registered earlier but the live table holds German [H] with the reason [NV] (see also Section 4.2 on the specification's tiers, which included Mandarin and Italian but not German — three different published language sets in three months).

1. Corpus-count drift within a single day. A technical dossier written on

6 September 2026 reports 1,146,066 lexemes / 1,595,537 senses, while direct SQL the same evening reports 1,238,383 / 1,608,018 (concepts and relations identical: 96,434 / 171,672); the session log of that same evening carries a third, intermediate figure (≈1.15 million lexemes), consistent with imports still landing through the day. Per-language sums reconcile with the SQL figure, so the dossier's numbers appear to be from a different moment or filter; the discrepancy is unexplained in the records [H] with the reconciliation [O]. This paper quotes the SQL figures with their date.

1. Hardware records conflict. The machine hosting the live system in

September 2026 was captured directly for this paper as an ASUS X751MA with an Intel Pentium N3540 (Bay-Trail), 8 GB RAM (Section 9.1) [H]. The written record is messier: project notes describe the box as an Intel Atom Bay-Trail (Z36xx/Z37xx) and as a Pentium N3540 — the same silicon generation under two marketing names [O] — while earlier 2026 conversation records reference a different machine, an Intel Jasper Lake "N5105" with 12 GB RAM (January 2026) and later as "a Jasper Lake 4 chip" (May 2026) and "Intel Jasper Lake N5105, 4 GB" (mid-2026 memory file). Jasper Lake is not Bay-Trail silicon, so the records describe either two different machines over the year or assistant mislabeling; the surviving evidence cannot decide which, and the point is flagged here rather than smoothed over [H] with the reconciliation [NV] (session logs and AI-conversation records, January–September 2026).

1. Data-quality warnings from the September audit: the corpus contains

178,576 duplicate 178,576 duplicate lemma pairs; 124,790 lexemes with no sense; 33.9% of definitions empty; and (before the English repair) 71.7% of senses not yet linked to concepts. The same audit found referential integrity held throughout (sums exact, zero dangling/orphan rows at the database level): the 71.7% figure describes senses pairs; 124,790 lexemes with no sense; 33.9% of definitions empty; and (before the English repair) 71.7% of senses not yet linked to concepts. The same audit found referential integrity held throughout (sums exact, zero dangling/orphan rows at the database level): the 71.7% figure describes senses not yet linked to a concept, the designed visible-orphan state, not broken references — an honest distinction the project deliberately keeps visible while alignment work continues [H] (audit, 5–6 September 2026).

1. Unmeasured or pending: performance on hardware other than the potato

has never been measured; the project has no public users yet, so there are no real-usage metrics; the studio's legal structure is not documented in the records; the "both Koreas" modeling question, Cantonese data scarcity, and the Taiwanese wordnet licensing review are open (Section 10.3).

10.3 The road ahead

The roadmap, user-approved on 29 August 2026, is the mature form of the June manuscript's seven graphs [H] (project note): phonetics/phonology, morphology (imported), syntax (imported), semantics (imported), lexicography (imported), corpora (imported), and — queued — discourse, etymology, and the correspondence and usage layers, plus the eventual public web interface. In the near term the documented plans are [H] (session logs, 7 September 2026):

• The CJKV family. The next languages planned for the schema are

Mandarin (cmn), Cantonese (yue), Korean (kor), and Vietnamese (vie). Their schedule depends on a data-availability reconnaissance that is honest about what the open-data landscape offers: OMW Wiktionary wordnets exist for Mandarin (19,079 rows), Cantonese (527), Korean (9,268), and Vietnamese (5,498); UD treebanks for Mandarin (123k tokens), Cantonese (14k), Korean (80k GSD plus 350k Kaist), Vietnamese (58k). Cantonese is the scarce one — no large wordnet, small treebank — and the Taiwanese wordnet licenses are under review [H] (session log and data inventory, 7 September 2026). No CJKV dataset has been imported into the live corpus as of this draft; the family's timing is an open decision pending that review.

• Research-v1 interface panes (word lens, lineage climb, field cards,

notebook, comparison table) were designed and backend-approved on 6 September 2026 and queued behind the instrument build [H].

• Publication. This paper is a preliminary AI-generated draft; the

human-authored publication is in development (front matter).

• Languages without end. The schema's promise — add a language as one

row plus datasets — is the project's stated long-term goal, and the donation pitch says it plainly: Lingua Mundi starts with six languages, built to add them indefinitely [H] (session log, 7 September 2026).

10.4 Conclusion

What began on 3 December 2025 as a request for free dictionary files is now a running, queryable, six-language map of meaning built on open data by one person with free tools and a 2014 laptop. The ideas that carried it — that meaning is the natural unit of multilingual data; that alignment must be explicit and labeled, never silently guessed; that expensive computation should be baked once and read many times; that a budget constraint is a design tool; that every claim, in a database or in a paper, should carry its source and its date — are not new, but they are true, and the project demonstrates that one person can act on them. Its most valuable artifact may not be the database at all, but the demonstration that a solo researcher can build honest infrastructure in public: measuring what he measures, dating what he dates, showing his orphans, and keeping his word — one star, six languages, and the whole sky still ahead.

References

APA 7th edition. Bibliographic details were verified against publisher pages and digital-object-identifier records on 7 September 2026. Software and data sources are cited per APA's software/data-set formats. (Only works cited in the text are listed.)

Bond, F., & Foster, R. (2013). Linking and extending an open multilingual wordnet. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (pp. 1352–1362). Association for Computational Linguistics. https://aclanthology.org/P13-1133/

Braille Institute of America. (2019). Atkinson Hyperlegible [Computer font]. https://brailleinstitute.org/freefont

Breen, J. W. (2004). JMdict: A Japanese-multilingual dictionary. In Proceedings of the Workshop on Multilingual Linguistic Resources (COLING 2004) (pp. 65–72). COLING. https://aclanthology.org/W04-2209/

Electronic Dictionary Research and Development Group. (n.d.). JMdict / EDICT Japanese–English dictionary project [Data files and documentation]. Retrieved September 7, 2026, from https://www.edrdg.org/jmdict/edict_doc.html

Fellbaum, C. (Ed.). (1998). WordNet: An electronic lexical database. MIT Press. https://mitpress.mit.edu/9780262561167/wordnet/

Gaia Collaboration, Brown, A. G. A., Vallenari, A., Prusti, T., de Bruijne, J. H. J., Babusiaux, C., … Zwitter, T. (2018). Gaia Data Release 2: Summary of the contents and survey properties. Astronomy & Astrophysics, 616, Article A1. https://doi.org/10.1051/0004-6361/201833051

Hernández, R. (2026, June 1). Lingua Mundi: A semantic infrastructure for concept-grounded language models (Versions 1–7) [Unpublished manuscript; revision and audit records preserved in the author's conversation-log archive].

Hernández, R. (2026, July 24). Multilingual Lexical Platform — Engineering specification (Version 2.0) [Unpublished specification document; text extraction on file].

Kirov, C., Cotterell, R., Sylak-Glassman, J., Walther, G., Vylomova, E., Xia, P., Faruqui, M., Mielke, S. J., McCarthy, A., Kübler, S., Yarowsky, D., Eisner, J., & Hulden, M. (2018). UniMorph 2.0: Universal morphology. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). European Language Resources Association. https://aclanthology.org/L18-1293/

McCarthy, A. D., Kirov, C., Grella, M., Nidhi, A., Xia, P., Gorman, K., Vylomova, E., Mielke, S. J., Nicolai, G., Silfverberg, M., Arkhangelskiy, T., Krizhanovsky, N., Krizhanovsky, A., Klyachko, E., Sorokin, A., Mansfield, J., Ernštreits, V., Pinter, Y., Jacobs, C. L., … Yarowsky, D. (2020). UniMorph 3.0: Universal morphology. In Proceedings of the Twelfth International Conference on Language Resources and Evaluation (LREC 2020) (pp. 3922–3931). European Language Resources Association. https://aclanthology.org/2020.lrec-1.483/

Meyer, C. M., & Gurevych, I. (2012). Wiktionary: A new rival for expert-built lexicons? Exploring the possibilities of collaborative lexicography. In S. Granger & M. Paquot (Eds.), Electronic lexicography (pp. 259–292). Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199654864.003.0013

Miller, G. A. (1995). WordNet: A lexical database for English. Communications of the ACM, 38(11), 39–41. https://doi.org/10.1145/219717.219748

Nivre, J., de Marneffe, M.-C., Ginter, F., Hajič, J., Manning, C. D., Pyysalo, S., Schuster, S., Tyers, F., & Zeman, D. (2020). Universal Dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of the Twelfth International Conference on Language Resources and Evaluation (LREC 2020) (pp. 4034–4043). European Language Resources Association. https://aclanthology.org/2020.lrec-1.497/

Princeton University. (n.d.). About WordNet. WordNet. Retrieved September 7, 2026, from https://wordnet.princeton.edu/

SIL International. (n.d.). ISO 639-3. Retrieved September 7, 2026, from https://iso639-3.sil.org/

de Saussure, F. (1983). Course in general linguistics (C. Bally, A. Sechehaye, & A. Riedlinger, Eds.; R. Harris, Trans.). Duckworth. https://www.bloomsbury.com/us/course-in-general-linguistics-9781472508829/ (Original work published 1916)

Speer, R. (2022). wordfreq: A database of word frequencies in many languages (Version 3.0) [Computer software]. GitHub. https://github.com/rspeer/wordfreq

Tiedemann, J. (2012). Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC 2012) (pp. 2214–2218). European Language Resources Association. https://aclanthology.org/L12-1246/

Vrandečić, D., & Krötzsch, M. (2014). Wikidata: A free collaborative knowledgebase. Communications of the ACM, 57(10), 78–85. https://doi.org/10.1145/2629489

Wikimedia Foundation. (n.d.). Wiktionary: The free dictionary. Retrieved September 7, 2026, from https://www.wiktionary.org/

Zipf, G. K. (1949). Human behavior and the principle of least effort. Addison-Wesley Press. https://archive.org/details/humanbehaviorpri0000zipf

Appendix A · Sources, method, and verification

A.1 The annotation convention

Following the studio's rigor standard (inherited from its investigative work), claims in this paper are tagged where a reader needs to know their epistemic status: [H] = verified fact — a measurement, code, or dated source consulted for this paper; [O] = attributed interpretation or opinion; [NV] = not verified / not located. Every figure in the abstract and body is either an [H] measurement with its date or is labeled.

A.2 Source families consulted (access 6–7 September 2026)

Representative verification queries (run 7 September 2026, role lingua): SELECT count(*) FROM concept|lexeme|sense|relation → 96,434 / 1,238,383 / 1,608,018 / 171,672; SELECT code FROM language → deu, eng, fra, jpn, por, spa; membership of synset 02121620-n per language (Section 8.5); the project's own acceptance script verify-lm-cat-demo.sh (Section 7).

A.3 Key numbers with dates (compressed timeline)

A.4 Contradictions and unresolved items

As documented in Section 10.2: the Italian/German Core-6 discrepancy; the same-day corpus-count drift (dossier 1,146,066 vs SQL 1,238,383 lexemes, with a third intermediate figure in the same evening's session log, reconciled per-language with the SQL figure); hardware records conflict (Atom Bay-Trail vs Pentium N3540 vs a Jasper Lake N5105 in earlier 2026 records — Section 10.2 item 3); the 71.7%-unlinked-senses and other data-quality warnings; and the unverified/undocumented items: exact repo founding date before 26 August, the reason for the language-set change, the model provenance of the 1-June audit texts (the audits survive in the author's ChatGPT conversation from the user side; which assistant model wrote which audit round is not independently attested in the export — see Section 3.3), the concept_word_rank live re-count (permission-denied), and the studio's legal structure.

A.5 Methodology notes

• The system under test is the production system; warm measurements carry

±0.3 s hard-disk noise; dates and times are given in the records' own timezone (Central Time unless noted; DeepSeek logs use +08:00 and were converted).

• Quotes are verbatim from the cited logs/files; ellipses mark cuts;

Spanish quotes are given with English glosses in the body where used.

• This paper is itself an AI-generated draft (front-matter notice); it was

written under the author's direction from the sources above and has not been peer-reviewed. Verification tooling: direct SQL, HTTP queries, headless-browser checks, and source-file reads; no claim was filled from memory when a source existed.

Appendix B · Glossary — plain-language terms used in this paper

How to use this glossary: the first time a technical term appears in the paper it is underlined with dots — hover it (or tap it) for a quick definition, or click it to jump here. This list is alphabetical.

Alignment

the act of deciding that a word in one language means the same thing as a word in another.

API

an interface that lets one program ask another for data (e.g., a dictionary app asking the Lingua Mundi database for a word).

Autonym

the name a language's own speakers call it (Spanish speakers call Spanish español).

Bake

the project's word for computing an expensive answer once, storing it, and reading it many times (the working rule is "bake, don't re-derive").

Breadth-first search

a way of exploring a graph outward in layers, like ripples from a stone, to find the shortest route between two points.

CJKV

a shorthand for the writing-system family spanning Chinese, Japanese, Korean, and Vietnamese (the letters stand for the script names).

Concept

a meaning, shared across languages (in this paper, "the idea of domestic cat," not any particular word).

Corpus

a large collection of language data (texts or dictionary entries) used as evidence.

Dataset

one packaged collection of data from one source, with its own format and license.

Derivation

derivation is how new words are built from existing ones (run → runner); etymology is a word's history across time and languages.

ETL

extract-transform-load: the process of reading data from an external source, reshaping it, and storing it in the database.

FastAPI

a free, open-source Python framework for building the kind of interface (API) this project serves.

Full-text search

searching a database by matching words inside stored text rather than by exact identifiers.

Grapheme

a written character: a letter of an alphabet, a syllable sign, or a kanji.

Graph

in this paper, a set of meanings (nodes) connected by typed relationships (edges).

Gzip

a standard compression method that makes files smaller before they travel over the internet.

HTTP

the protocol web browsers and APIs use to ask for and receive data over the internet.

Hypernym

hypernym = "a kind of": feline is a hypernym of domestic cat. A taxonomy is the resulting tree of kinds.

Importer

a program that knows how to read one external dataset and write it into Lingua Mundi's own schema.

Index

a database structure that lets queries find rows fast, like a book's index instead of reading every page.

Inflection

inflection is changing a word's form for grammar (run → ran); morphology is the study of those forms.

Instance-of

two relation types in the graph: "this is one example of that" (a cat is an instance of a mammal-kind) and "this is a piece of that" (a paw is part of a cat).

Interlingua

a shared middle representation between languages: instead of translating Spanish→English directly, go Spanish→meaning→English.

ISO 639-3

the international standard assigning every human language a unique three-letter code (spa = Spanish, jpn = Japanese).

JMdict

the JMdict is a large Japanese–multilingual dictionary; KANJIDIC2 is its companion file of kanji (Chinese characters used in Japanese) with readings and meanings.

JSON Lines

a simple text format where each line is one data record, used for the project's earliest pipeline files.

Kaikki

a machine-readable extraction of Wiktionary data, easier for programs to read than the original pages.

Kanji

the Chinese-origin characters used in written Japanese.

Lemma

the dictionary form of a word (the form you look up: cat, not cats; run, not ran).

Lexeme

a word form in a specific language (the string cat).

License

the legal terms stating how data may be used and shared; "open" licenses allow reuse with attribution (e.g., CC BY-SA).

Long tail

the great mass of rare items at the far end of a Zipfian distribution (see Zipf's law).

Materialized view

a stored, precomputed query result, refreshed on a schedule, so repeated questions don't redo the work (the "bake" pattern).

Morphology

the study of word forms and how they change for grammar (see inflection).

Natural language

human language as people actually speak and write it, as opposed to programming languages.

Node

in a graph, a node is one item (here, one concept) and an edge is a connection between two items (here, a typed relation).

OMW

the collection of wordnets aligned to the English WordNet.

OPUS

a large collection of translated sentences gathered from the web and from movie subtitles.

Orphan sense

a dictionary meaning not yet linked to a shared concept; kept visible rather than hidden.

Pagination

splitting a long list of results into numbered pages instead of sending everything at once.

Part of speech

the grammatical class of a word: noun, verb, adjective, and so on.

Pipeline

a chain of programs that moves data from a raw source through cleaning steps into its final home.

Plugin

a plugin is a small add-on program for one dataset; the registry is the single authoritative list of them.

Polysemy

one word with several meanings (bank: riverbank / bank of money).

PostgreSQL

a free, open-source relational database (the system Lingua Mundi stores everything in).

Relation

a typed, directed connection between two concepts (domestic cat is a feline).

Schema

the blueprint of a database: which tables exist, what each row holds, and how tables refer to each other.

Sense

one meaning of one word as recorded by one dictionary.

Signifier

Saussure's two faces of a word: its sound or shape (signifier) and the concept it points to (signified).

Sky brightness

the project's ranking of how "important" a meaning is, based on how many languages and words attach to it.

SQL

the standard language for asking questions of a relational database.

Synset

a set of words that can express the same meaning in some context; in Lingua Mundi, a synset's meaning is stored as a concept.

Token

in machine learning, the small chunk of text (often a word or word-part) a model reads or writes.

Treebank

a collection of real sentences annotated with their grammar (which word plays which role).

UD

a community project publishing grammatically annotated sentence data in many languages under open licenses.

UniMorph

a multilingual repository of word-form (inflectional) data.

URI

a stable, unique address for a thing, here used for concepts (e.g., princeton-wordnet:02121620-n for domestic cat).

Wikidata

Wikimedia's structured knowledge base of statements with machine-readable identifiers, released under CC0.

Wiktionary

the collaborative dictionary hosted by the Wikimedia Foundation.

Word form

one spelling of a word in one language; see lexeme.

WordNet

the original English meaning-organized lexical database (Princeton); the OMW aligns other languages to it.

Wordfreq

a database of word-frequency lists across languages, used to rank words by how common they are.

Writing script

the set of characters a language is written in (Latin, Arabic, Chinese characters, and so on).

Zipf's law

the observation that a few words account for most usage and the long tail of rare words accounts for the rest.