Sovereign AI starts with semantics, not infrastructure.
Sovereign AI starts with semantics, not infrastructure. An organization can own and run its model and still outsource meaning to an external foundation model's conceptual world — unless it decides, in a machine-readable Lexical Graph, what the model is allowed to mean. Sovereignty is who owns the categories and distinctions: embeddings may propose neighborhoods, but SKOS (since 2009) decides related versus exactMatch. Agent-Oriented Ontology Engineering (AOOE) binds three graphs — Lexical Graph (meaning), Entity Graph (memory), Action Graph (action) — so language becomes meaning, meaning becomes entities, and entities become permitted or blocked actions. Without that verification layer, scale without verification is the Hollow Economy's Measurability Gap accumulating as semantic debt.
View this analysis as a KG entity“You shall know a word by the company it keeps.” — J. R. Firth
Sovereign AI does not begin when you host the weights. It begins when you decide what the model is allowed to mean. An organization can run its own model and still live inside an external foundation model's conceptual world — its categories, its collapses, its marketing-language gravity. Sovereignty is ownership of distinctions.
Sovereignty is not GPU ownership. It is authority over categories and distinctions: customer versus account holder versus beneficiary; refund versus chargeback; marketing claim versus approved claim.
Owning or running the model can still outsource meaning if the conceptual world — the ontology the model acts in — is borrowed from a general foundation model rather than governed locally.
Agent-Oriented Ontology Engineering (AOOE) is the operating pattern: a Lexical Graph for meaning, an Entity Graph for memory, and an Action Graph for action. Language maps to meaning, meaning to entities, entities to permitted actions. The travel example — group lodging around Lungau and Obertauern — is the same path in a hospitality setting.
Machine-readable organization language: taxonomy backbone plus synonyms, preferred terms, broader/narrower, regional variants, deprecated terms, and customer language. Saussure's signifier versus signified becomes operational here.
Memory: the products, pages, places, and offers the organization actually has. Distinct lexical concepts bind to distinct entities so Debt Settlement Product is not Debt Consolidation Product.
Action: retrieve versus recommend (and book, file, pay). When the utterance is related but not exactMatch, retrieve may be allowed as context; recommend must be blocked.
The Lexical Graph is not a prompt. It is SKOS: prefLabel, altLabel, hidden/deprecated labels, broader/narrower, related, and exactMatch. Embeddings propose a neighborhood; the taxonomy decides whether that neighborhood is identity. SKOS has been able to say related versus exactMatch since 2009.
Proximity is not identity. Firth's 'company it keeps' describes collocation, not legal equivalence. Debt settlement, debt consolidation, loan refinancing, and payment restructuring sit in a tight embedding neighborhood with the customer utterance 'debt relief' — and must not collapse. Refund is not chargeback. A marketing claim is not an approved claim. Premium (brand) is not luxury.
Very high proximity; legally distinct roles. No exactMatch.
Illustrative cosine-like proximity in 0-1. High neighborhood score; taxonomy keeps them related, not exactMatch.
Illustrative proximity. Related, not equivalent.
Compact debt-* neighborhood. Still not identity.
Illustrative proximity. Related, not equivalent.
Brand neighborhood. Related, not exactMatch.
Homograph neighborhood. Related, not exactMatch.
Illustrative proximity. Related, not equivalent.
Catalini, Hui, and Wu (MIT Sloan, February 2026; H/T Ethan Hays) argue that scale without verification is accumulating debt: a Measurability Gap between what agents can execute and what humans can verify. The Lexical Graph is verification infrastructure for meaning. Wittgenstein's meaning-is-use lands the same point: enterprise semantics are local language games, not a global embedding.
Do not compress brand books into a prompt. A prompt is not a governed semantic space. Preferred terms, deprecated terms, and customer language have to live in the Lexical Graph if agents are to be auditable.
Showcase ABox: Northwind Bank encodes customer/account holder/beneficiary as distinct, keeps the debt-* cluster related-not-equivalent, binds two products in the Entity Graph, and blocks recommend of settlement when the utterance is only related via 'debt relief'.
Illustrative proximity. Related, not equivalent.
Fictional regulated bank used only as illustrative sample data for this meshup. Not a real WordLift customer and not production data.
When the utterance is 'debt relief' (related, not exactMatch), retrieve of the settlement page MAY be allowed as context.
Illustrative product bound to the debt-settlement concept. Not WordLift production data.
Brand neighborhood. Related, not exactMatch.
Illustrative proximity. Related, not equivalent.
Illustrative Entity Graph offer bound to the group-lodging concept and the Obertauern/Lungau places.
Alpine region used in the article's travel example. Illustrative entity, not a live booking system.
Explainer page for the settlement concept. Retrieve MAY be allowed as context for a related utterance.
When the utterance is 'debt relief' (related, not exactMatch), recommend of the settlement product MUST be blocked.
Homograph neighborhood. Related, not exactMatch.
Illustrative product bound to the debt-consolidation concept. Distinct from settlement.
High proximity, distinct legal instruments.
Compact debt-* neighborhood. Still not identity.
Ski area in the Lungau region used in the article's group-stay example.
Illustrative proximity. Related, not equivalent.
Explainer page for the consolidation concept.
Recommend MUST be blocked: related, not exactMatch.
Retrieve MAY be allowed as context for the related utterance 'debt relief'.
Illustrative cosine-like proximity in 0-1. High neighborhood score; taxonomy keeps them related, not exactMatch.
Same-concept altLabel may permit book; contrast with debt-relief where recommend is blocked.
Governed SKOS concept scheme for the Northwind Bank example. Illustrative sample data, not WordLift production data.
Utterance 'group stay' is an altLabel of group lodging (same concept), so retrieve is allowed.
Very high proximity; legally distinct roles. No exactMatch.
Jason Gilbert asked whether embeddings suggest a relationship and the organization taxonomy then decides equivalence. That is exactly the governed-space rule: embeddings propose; SKOS decides related versus exactMatch.
Modeled from the article's comment thread. The commenter's public profile URL was not confirmed; the Person IRI is document-local.
Write down who decides meaning. List the categories that must not collapse (customer / account holder / beneficiary; refund / chargeback; marketing claim / approved claim). Sovereignty is this list, not the cluster.
Publish prefLabel, altLabel (including customer language), hidden/deprecated labels, broader/narrower, and regional variants. Do not compress the brand book into a prompt.
Record embedding neighborhoods as observations with numeric proximity. Never treat high cosine as exactMatch. Firth's company-keeping is a neighborhood signal, not identity.
For every proposed pair, assert skos:related or skos:exactMatch — or neither. SKOS has carried that distinction since 2009. Absence of exactMatch is a decision.
Attach each product, page, and place to its lexical concept. Debt Settlement Product and Debt Consolidation Product are distinct entities because their concepts are distinct.
For each utterance, retrieve MAY be allowed on a related concept (context). Recommend, book, file, or pay MUST require exactMatch. Model permission as an explicit boolean on the action.
Against the Measurability Gap, meaning has to be auditable. Query related-not-equivalent pairs, blocked recommends, and exactMatch counts the way you would query financial controls.
Yes. Sovereign AI starts with semantics, not infrastructure. Hosting or running the weights does not decide what the model is allowed to mean. If categories and distinctions are borrowed from an external foundation model's conceptual world, meaning has been outsourced.
Sovereignty is who decides what the model is allowed to mean — who owns the categories and distinctions. Not who owns the GPUs.
A machine-readable representation of the organization's language: a taxonomy backbone plus synonyms, preferred terms, broader/narrower, regional variants, deprecated terms, and customer language. It is the operational form of Saussure's signifier versus signified: language to meaning to entities.
Because proximity is not identity. Tokens that keep company in vector space (Firth) can still be legally, commercially, and operationally distinct. Collapsing debt settlement with consolidation, refund with chargeback, or marketing claim with approved claim is a semantic attack, not a retrieval success.
Related versus exactMatch. Embeddings propose a neighborhood; the taxonomy decides whether members of that neighborhood are equivalent. SKOS has carried that distinction since the 2009 W3C Recommendation.
Agent-Oriented Ontology Engineering: three graphs — Entity Graph (memory), Lexical Graph (meaning), Action Graph (action) — so language becomes meaning, meaning becomes entities, and entities become permitted or blocked actions.
No. 'Debt relief' is customer language related to debt settlement, not exactMatch. Retrieve of a settlement explainer MAY be allowed as context. Recommend of a Debt Settlement Product MUST be blocked until the Lexical Graph says identity.
They are different legal roles. High embedding proximity does not make them interchangeable. A governed Lexical Graph keeps them as separate concepts with no exactMatch.
Catalini, Hui, and Wu (MIT Sloan, Feb 2026) describe a Hollow Economy in which scale without verification accumulates debt — a Measurability Gap. The Lexical Graph is verification infrastructure for meaning: it is what you can audit when agents act at scale.
A prompt is not a governed semantic space. Preferred terms, deprecated terms, and customer language have to be queryable, versionable, and binding on actions. Brand governance belongs in the Lexical Graph.
Meaning is use. Enterprise semantics are local language games. A global embedding space is not the organization's language; the Lexical Graph is.
Signifier (the words customers and staff actually use) maps through the Lexical Graph to signified (governed concepts), which bind to entities, which authorize actions. Language to meaning to entities to actions.
Whether embeddings suggest a relationship and the organization taxonomy then decides equivalence. Yes: embeddings propose; SKOS decides related versus exactMatch.
The same AOOE path applies outside banking: group lodging around Lungau and Obertauern is language (group stay, ski-week) to meaning (group lodging concept) to entities (properties) to actions (retrieve vs book).
No. They can share an embedding neighborhood with 'my card got charged twice' and with duplicate-transaction and dispute concepts. Related is allowed; exactMatch is not. The taxonomy, not cosine, decides.
Machine-readable organization language: taxonomy backbone plus synonyms, preferred terms, broader/narrower, regional variants, deprecated terms, and customer language. The verification infrastructure for meaning.
AOOE memory graph: products, pages, places, and offers bound to lexical concepts so distinct meanings stay distinct entities.
AOOE action graph: retrieve, recommend, book, file, pay. Permission is a function of the lexical relation (related vs exactMatch), not of embedding proximity.
Agent-Oriented Ontology Engineering: the three-graph pattern (Lexical, Entity, Action) that takes language to meaning to entities to actions.
In this thesis, sovereignty is who decides what the model is allowed to mean (categories and distinctions), not who owns the model infrastructure.
The set of near-meanings embeddings will treat as close enough. Un-governed proximity lets agents collapse distinct concepts (debt settlement/consolidation; refund/chargeback; marketing claim/approved claim).
Catalini, Hui, and Wu: an equilibrium of explosive measured output with decaying verification and human agency. Scale without verification is accumulating debt.
The widening gap between what agents can execute and what humans can afford to verify. The Lexical Graph is the verification layer for meaning inside that gap.
W3C Recommendation (18 August 2009) for thesauri and taxonomies. In this thesis SKOS decides related versus exactMatch after embeddings propose a neighborhood.
Vector representations of tokens. The article treats them as an expanded semantic attack surface: proximity is not identity, and Firth's company-keeping is not a license to collapse distinct SKOS concepts.
W3C RDF query language. Every recipe in this collection targets the URIBurner SPARQL endpoint and projects IRI-valued variables alongside labels.
Interactive graph visualization derived from the companion RDF. Click nodes to resolve, drag to explore. Graph data embedded from companion RDF at generation time.
Query this knowledge graph on URIBurner. The editor opens on the canonical SAMPLE entity-type summary (DAV named graph). Pick a recipe, edit freely, then run live or copy.
Reproduced verbatim from the companion RDF. Execute loads the query into the workbench below and runs it live.
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
PREFIX schema: <http://schema.org/>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX post: <https://www.linkedin.com/pulse/you-can-own-the-model-still-outsource-meaning-andrea-volpini-jgflf/#>
SELECT DISTINCT ?conceptIri ?concept ?relatedIri ?related
WHERE {
GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/you-can-own-the-model-still-outsource-meaning-grok-1.ttl> {
?conceptIri skos:related ?relatedIri .
FILTER NOT EXISTS { ?conceptIri skos:exactMatch ?relatedIri }
?conceptIri skos:prefLabel ?concept .
?relatedIri skos:prefLabel ?related .
FILTER (STR(?conceptIri) < STR(?relatedIri))
}
}Concepts the taxonomy treats as related, not equivalent. Proximity is not identity.
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
PREFIX schema: <http://schema.org/>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX post: <https://www.linkedin.com/pulse/you-can-own-the-model-still-outsource-meaning-andrea-volpini-jgflf/#>
SELECT ?utterance ?conceptIri ?concept ?relatedIri ?related ?targetIri ?target ?actionIri ?action ?permitted
WHERE {
GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/you-can-own-the-model-still-outsource-meaning-grok-1.ttl> {
?conceptIri skos:altLabel ?utterance .
FILTER (LCASE(STR(?utterance)) = "debt relief")
?conceptIri skos:prefLabel ?concept .
?conceptIri skos:related ?relatedIri .
?relatedIri skos:prefLabel ?related .
?targetIri schema:about ?relatedIri ;
schema:name ?target .
?actionIri schema:object ?targetIri ;
schema:name ?action ;
post:actionPermitted ?permitted .
}
}Maps the customer utterance 'debt relief' through the Lexical Graph to related products and retrieve-vs-recommend actions.
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
PREFIX schema: <http://schema.org/>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX post: <https://www.linkedin.com/pulse/you-can-own-the-model-still-outsource-meaning-andrea-volpini-jgflf/#>
SELECT ?conceptIri ?concept ?definition
WHERE {
GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/you-can-own-the-model-still-outsource-meaning-grok-1.ttl> {
VALUES ?conceptIri {
post:conceptCustomer
post:conceptAccountHolder
post:conceptBeneficiary
}
?conceptIri skos:prefLabel ?concept .
OPTIONAL { ?conceptIri skos:definition ?definition }
}
}
ORDER BY ?conceptThree party roles that must not be treated as interchangeable.
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
PREFIX schema: <http://schema.org/>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX post: <https://www.linkedin.com/pulse/you-can-own-the-model-still-outsource-meaning-andrea-volpini-jgflf/#>
SELECT ?relationIri ?relation (COUNT(*) AS ?edgeCount)
WHERE {
GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/you-can-own-the-model-still-outsource-meaning-grok-1.ttl> {
{
?s skos:related ?o .
BIND(skos:related AS ?relationIri)
BIND("related" AS ?relation)
} UNION {
?s skos:exactMatch ?o .
BIND(skos:exactMatch AS ?relationIri)
BIND("exactMatch" AS ?relation)
}
}
}
GROUP BY ?relationIri ?relation
ORDER BY ?relationHow many SKOS related vs exactMatch edges the governed space actually asserts.
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
PREFIX schema: <http://schema.org/>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX post: <https://www.linkedin.com/pulse/you-can-own-the-model-still-outsource-meaning-andrea-volpini-jgflf/#>
SELECT ?productIri ?product ?conceptIri ?concept
WHERE {
GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/you-can-own-the-model-still-outsource-meaning-grok-1.ttl> {
?productIri a schema:Product ;
schema:name ?product ;
schema:about ?conceptIri .
?conceptIri skos:prefLabel ?concept .
}
}
ORDER BY ?productProducts in the Entity Graph bound to their Lexical Graph concepts.
PREFIX skos: <http://www.w3.org/2004/02/skos/core#>
PREFIX schema: <http://schema.org/>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX post: <https://www.linkedin.com/pulse/you-can-own-the-model-still-outsource-meaning-andrea-volpini-jgflf/#>
SELECT ?obsIri ?obs ?conceptAIri ?conceptA ?conceptBIri ?conceptB ?proximity
WHERE {
GRAPH <https://linkeddata.uriburner.com/DAV/demos/daas/you-can-own-the-model-still-outsource-meaning-grok-1.ttl> {
?obsIri a schema:Observation ;
schema:name ?obs ;
post:observedConceptA ?conceptAIri ;
post:observedConceptB ?conceptBIri ;
schema:value ?proximity .
?conceptAIri skos:prefLabel ?conceptA .
?conceptBIri skos:prefLabel ?conceptB .
FILTER NOT EXISTS { ?conceptAIri skos:exactMatch ?conceptBIri }
}
}
ORDER BY DESC(?proximity)Numeric embedding neighborhoods where high proximity coexists with skos:related and the absence of skos:exactMatch.