Ontology hosting & lifecycle¶

epistemic-graph is an ontology server: it hosts OWL/RDFS ontologies, maps them onto the property graph, and reasons over them in-engine with the OWL 2 EL⁺/RL reasoner (eg-rdf/owl). Reasoning is pure-Rust (no horned-owl, no whelk-rs) and adds two things most triple-stores don't: confidence weighting and Ebbinghaus time-decay of derived facts.

Status snapshot: EL⁺ + RL classification, consistency, incremental materialization, confidence weighting, time-decay, distributed reasoning, the OWL-DL tableau (cardinality, allValuesFrom, owl-dl feature), SWRL user rules (swrlb: built-ins) and rdfs:range completion are all supported (EG-KG.ontology.owl-reasoning/059/060) — the EL/RL fast path stays the default. See the capability matrix.

The lifecycle¶

flowchart LR
    LOAD["1. Load<br/>OWL/RDFS as RDF"] --> MAP["2. Map<br/>onto the property graph"]
    MAP --> REASON["3. Reason<br/>EL⁺/RL fixpoint closure"]
    REASON --> QUERY["4. Query<br/>SPARQL / REASON / SQL"]
    QUERY --> EVOLVE["5. Evolve<br/>incremental add_axioms"]
    EVOLVE --> REASON

1–2. Load & map¶

Load triples (AddTriples) or a Turtle/RDF document; the engine maps subjects/objects to nodes, resource triples to typed edges, and literals to properties (see the SPARQL/RDF guide).

3. Reason¶

The reasoner saturates the ontology to a monotone fixpoint:

OWL 2 EL⁺ completion — the named completion rules (CR-sub, CR-conj, CR-some±, CR-chain, CR-subrole, CR-bot, CR-disjoint) compute subsumers S(A) and role relations R(r). This reaches concept-level entailments (e.g. HumanHeart ⊑ HumanComponent through ∃partOf.Body) that a plain RL/Datalog pass cannot.
OWL 2 RL property rules — subPropertyOf, property chains, transitive/symmetric/inverse properties, rdfs:domain (lifted to EL ∃r.⊤ ⊑ D).
Consistency — any satisfiable class forced to subsume owl:Nothing flags the ontology inconsistent.
Confidence weighting (EG-KG.ontology.concept-13) — a per-axiom eg:confidence annotation propagates as a noisy-OR (MAX over derivations of axiom_conf × ∏ premise_conf).
Ebbinghaus time-decay — derived facts decay by conf · exp(−ln2 · age / half_life) against now − last_access (GRAPH_SERVICE_DECAY_HALF_LIFE, default 7 days).

Materialization is incremental: add_axioms keeps the prior S/R and resumes the fixpoint, provably equal to a from-scratch run. Reasoning runs single-graph (Method::OwlReason) or distributed across shards (Method::OwlReasonDistributed), unioning TBox + ABox into one weighted closure.

4. Query¶

SPARQL over the materialized graph (see the SPARQL guide).
REASON <Class> as a UQL source — the Op::Reason op (under owl-plan) seeds a RowSet with the inferred members of a class, which then composes with traversal, vector rank, etc.:

REASON HumanComponent
  |> TRAVERSE -[:partOf]->{1,2}
  |> LIMIT 50

- SQL / Cypher / GraphQL see the inferred type edges and properties like any other graph data.

5. Evolve¶

Add or retract axioms and re-run the incremental closure; confidence and decay keep the derived set honest about how strong and how fresh each inference is.

Scope¶

Default envelope: OWL 2 EL⁺ + RL forward-chaining (the fast path). rdfs:range is enforced in completion (EG-KG.ontology.owl-reasoning).
Full DL: the owl-dl feature adds a pure-Rust DL tableau (cardinality, allValuesFrom, nominals, negation) run consistency → classification → instance, plus SWRL user rules with the swrlb: built-in library (EG-KG.ontology.concept-2/060). These are opt-in beyond the EL/RL default.

Tier note¶

The lean pi tier carries SPARQL SELECT + OWL reasoning via the Method::Owl* RPCs, but the SQL-backed Op::Reason planner op needs owl-plan (which pulls DataFusion) and lands in the node tier. See tiers & binaries.