Ontology
The formal vocabulary of a domain: what things exist and what relationships are allowed between them. Defines that a "medicinal product" has a substance, a dose form, a route. Doesn't say what any specific instance is, just what's possible.
Semantics
The meaning layer that keeps a concept interpreted the same way everywhere it's used. Two systems can share an ontology and still disagree on semantics if "10mg" or "adverse event" gets coded or scoped differently between them.
Knowledge Graph
The ontology populated with real, connected instances: nodes are actual entities (this drug, this device, this claim), edges are their real relationships. What lets a machine traverse and reason over the data, not just store it.
Ontology, plain version
The rulebook says animals have a kind (duck or bunny), a home (pond or burrow), and a food (bugs or carrots). It doesn't name any real animal yet, just what's allowed to exist.
Semantics, plain version
Everyone has to agree "duck" means the bird that quacks and swims, not someone dodging a question. If one farm's records use "duck" for the bird and another farm's use it as a verb, their animal counts will never match even though they typed the same word.
Knowledge graph, plain version
Now put real animals into the rulebook. Daffy is a duck, lives in Pond 2, eats bugs, and is friends with Thumper, a bunny who lives in Burrow 5 and eats carrots. Now you can ask a question and trace the answer: "who are Daffy's friends, and what do they eat?"
Two real examples of all three working together
FHIR + SNOMED CT knowledge graphs (healthcare interoperability)
Shared ontology (SNOMED CT clinical concepts) plus a semantic layer (FHIR resources carrying standardized codes) plus an actual knowledge graph (Neo4j-based semantic engines linking patient, encounter, diagnosis, and procedure nodes) lets EHR systems from different vendors exchange and reason about clinical meaning safely, not just pass strings back and forth.
IDMP-O (pharmaceutical product identification, EU-wide)
ISO's IDMP standard defines the ontology (substances, pharmaceutical products, dose forms, routes, units of measurement, five sub-standards total). The Pistoia Alliance and EDM Council are building the actual knowledge graphs and shared vocabularies on top of it, engineered for continuous release through Accurids, so pharma companies and the EMA can interoperate on medicinal product data ahead of the mandatory June 2026 submission deadline to EMA's Product Management System.
Three real examples of what happens when they're misaligned
Healthcare: CMS billing
ICD-10-CM is a billing-purpose code set; SNOMED CT is a clinical, computable ontology. The two don't map cleanly. A diagnosis code that doesn't logically support the billed procedure (a respiratory diagnosis paired with an orthopedic procedure) fails payer semantic edits outright.
CMS FY2025: 6.55% improper payment rate, $28.83B. Medical necessity mismatches: 15.3% of that.
Pharma: formulations
IDMP-O is a real, shared ontology effort, but adoption has run in silos: different companies implement the same standard with different interpretations of the same fields.
Direct finding, 2026: "today's IDMP implementations are built in silos and have already created inconsistencies of interpretation."
MedTech: regulatory (EUDAMED)
Basic UDI-DI is meant to be a semantic grouping key: same intended purpose, risk class, and design characteristics. When a manufacturer's grouping doesn't actually align with those three criteria, or the parent-child hierarchy to unit-of-use UDI-DIs breaks, EUDAMED's automated business rules reject the submission outright.
Mandatory from May 28, 2026. AI is starting to predict likely rejections from historical patterns.
Why this matters for this specific interview
This isn't trivia. Cortex is described in its own launch materials as a "knowledge engineering and multi-agent orchestration platform" that codifies 25+ years of expertise into knowledge graphs. The three failure cases above are exactly the shape of problem a platform like that has to prevent in regulated content: a shared ontology alone doesn't guarantee semantic alignment, and semantic drift between systems is what produces denials, inconsistent interpretation, and automated rejections. Say this directly if asked what ontology/KG work has to do with medical writing: "The failure mode isn't usually a missing ontology, it's semantic drift between systems that technically share one, exactly what's happening in IDMP right now and what CMS's own improper-payment numbers show for ICD-to-CPT alignment."