/semantic-reasoning — Inference Over a Knowledge Graph

Semantic reasoning turns an ontology from static schema into active inference. Given facts and rules, a reasoner derives new facts. Given questions, SPARQL retrieves them with provenance. This is the layer that makes M365 Work IQ output actionable, not just retrievable.

When to invoke

• An ontology exists (see /ontology) and a decision needs the graph to answer questions
• Work IQ / M365 semantic index returns candidates and you need to filter/rank by rules
• Athanasios is building a citation-heavy brief that requires provenance
• Atlas or Bob needs to verify a graph's consistency (SHACL constraints)

Steps

1. Load the ontology and instance data. Point the reasoner at the TTL/JSON-LD schema plus the instance graph. Confirm both parse cleanly (no undefined properties, no dangling references).
2. Identify the inference goal. Write down the question in plain English first. Example: "Which projects have both a delivery risk and no assigned owner?" Explicitness prevents scope creep in the query.
3. Design the SPARQL query. Draft the query with named graphs for provenance. Use OPTIONAL for missing-data checks, FILTER NOT EXISTS for negations, ?g graph variable to track source. Test on a small subset first.
4. Run the reasoner. Apache Jena (OWL RL / OWL Mini rules) is the default — active, stable, no license friction. HermiT and Openllet are OWL 2 DL alternatives when full DL reasoning is required. Verify the current stable release at author time. Capture inferred triples for audit.
5. Validate with SHACL. Write SHACL shapes for the constraints the domain requires (cardinality, value ranges, type membership). Run validation. Report violations with node paths and offending values.
6. Cite provenance. For every returned fact, include the source graph, the ontology term, and the reasoner rule that produced it (if inferred, not asserted). No answer ships without a citation trail.

Deliverable format

• query.sparql — the exact query executed
• results.json — bindings with ?source_graph, ?asserted_or_inferred, ?rule_applied
• shapes.ttl — SHACL shapes plus a validation report
• One-paragraph plain-English answer to the original question, citing 2-5 provenance links

Non-negotiables

• Never return an inferred fact without naming the rule that produced it
• Always validate the graph with SHACL before publishing results
• Reasoner and version pinned in the deliverable — reproducibility matters
• Escape RDF-star / named-graph metadata explicitly when relevant — never conflate assertion with observation

References

• Apache Jena — reasoners
• SPARQL 1.1 Query Language
• SHACL — W3C recommendation
• Openllet — OWL 2 DL reasoner (active Pellet fork)
• Superteam ADR — ontology/semantic skills