#knowledgegraphs #owl #shacl #typetheory #datavalidation #ontology #ai #linkeddata | Volodymyr Pavlyshyn | 15 comments

You're building an enterprise knowledge graph. Someone queries: "Does Alice have parents?" Your OWL ontology says: "I don't have that information yet. Maybe she does, maybe she doesn't. I simply don't know." Your SHACL validator says: "I see zero parent relationships for Alice. She has no parents. Validation FAILED." Same database. Same query. Opposite conclusions. This isn't a bug—it's two different philosophical assumptions about reality. The Open World (OWL's View): The world is inherently incomplete. We're constantly discovering new information. Just because we haven't stated Alice's parents doesn't mean they don't exist—they're simply not in our system yet. This is perfect for integrating data from multiple sources, evolving schemas, and reasoning about distributed knowledge where no single source has the complete picture. I've seen this play out at LiveIntent with identity graphs—user data comes from dozens of sources, each with partial information. Concluding that missing data means "false" would be disastrous. The Closed World (SHACL's View): For validation and data quality, we need definitive answers. If your form requires two parent names and the user submits blank fields, that's not "unknown"—it's incomplete data that should fail validation. This assumption works beautifully for databases, compliance checking, and any scenario where you have complete information about a bounded domain. The breakthrough for me: Modeling both assumptions in dependent type theory (specifically Agda) made the distinction crystal clear. Open World Assumption maps beautifully to intuitionistic logic—where absence of proof is NOT proof of absence. You need three-valued logic: true, false, and unknown. Closed World Assumption requires classical logic with decidability—everything must be true or false. No middle ground. The type system itself can enforce which assumption you're operating under, preventing the silent philosophical mismatch that causes so many bugs in production systems. The hybrid approach that actually works: Modern systems need both. Here's the pattern I've found most effective: Use OWA for your domain ontology and schema reasoning. Let your knowledge evolve, integrate incomplete data sources, derive new facts through inference. Embrace the "unknown." Use CWA for concrete data validation. When a user fills out a form, when an API returns data, when you're checking compliance—demand completeness and fail on missing required fields. The bridge: Run OWA inference to enrich your data before CWA validation. Let the reasoner fill in derivable facts, then validate the enriched result. For AI agent memory systems: Agents learn from uncertain, partial observations (OWA), but must make decisions based on verified facts (CWA). In enterprise data integration: Merge data from silos that each have partial truth (OWA), then validate against regulatory requirements (CWA). #KnowledgeGraphs #OWL #SHACL #TypeTheory #DataValidation #Ontology #AI #LinkedData | 15 comments on LinkedIn