Provenance, lineage, and temporal claims

A graph becomes dependable when a consumer can answer not only “what does it say?” but also “who asserted this, from which material, through which transformation, under what interpretation, and for what time?” Provenance is therefore part of the claim model rather than a decorative source URL. The W3C PROV model gives a compact vocabulary: an entity is a thing with some fixed aspects, an activity occurs over time and uses or generates entities, and an agent bears responsibility. Relations such as wasDerivedFrom, wasGeneratedBy, used, and wasAttributedTo make lineage queryable.

Choose provenance granularity from the decisions the graph must support. Dataset-level provenance can say that release R came from source snapshot S through pipeline P. Record-level provenance can identify the source row or document. Claim-level provenance can distinguish two conflicting assertions extracted from the same document. Claim-level detail costs more storage and complicates updates, but it is necessary when answers, audits, or automated actions must cite exact support. A practical design often records every source unit and extraction activity, then assigns stable assertion identifiers only to high-consequence claim types.

RDF-star-style annotations may be convenient in stores that support them, but portable designs should not depend silently on vendor semantics. A first-principles pattern creates an assertion resource that has subject, predicate, object, evidence, extraction activity, status, and confidence. Another pattern places a coherent assertion set in a named graph and describes the graph. The two patterns answer different questions: named-graph provenance is excellent for batches; explicit assertion resources support per-claim review. Document the chosen level and test round trips through every serialization and store.

Keep observed time, valid time, and transaction time separate. Observed time says when a source or process encountered information. Valid time says when a claim applies in the modeled world. Transaction time says when the platform accepted or superseded it. If a contract signed on March 4 is ingested on March 8 but effective April 1, collapsing the dates destroys useful semantics. OWL-Time provides instants, intervals, and relations; applications may use simpler date properties while retaining these distinctions.

Corrections should be additive and inspectable. Do not erase an assertion merely because a later source contradicts it. Mark its lifecycle state, link a replacement with a supersession relation, and publish a new graph release. A current-state view can select the latest approved assertion, while an audit view reconstructs what was believed at an earlier transaction time. Physical deletion remains appropriate for retention or privacy obligations, but the deletion event and policy basis should be recorded outside the deleted content where permitted.

Confidence is not provenance. A model score says something about a process estimate; it does not identify the evidence. Source authority is also contextual: a regulator may be authoritative for a license but not for a corporate nickname. Store method, model and prompt version, thresholds, and review outcome with generated assertions. Avoid combining extraction confidence, identity confidence, and source trust into one opaque number.

Lineage must survive derived answers. When a rule infers that Supplier A is high risk from sanctions and ownership claims, represent the conclusion as derived, identify the rule or ontology release, and retain links to supporting assertion identifiers. A natural-language answer should cite source units rather than the derived triple alone. If a support claim is retracted, dependency tracking should invalidate affected conclusions, caches, and evaluation fixtures.

Design provenance queries before designing storage. Ask for all claims derived from a deleted document, all answers produced with an obsolete ontology, all assertions reviewed by a given policy version, and the belief state at a past transaction time. If those queries require log archaeology, provenance is not yet a product capability. Test completeness by sampling released claims and proving a path back to immutable evidence and responsible processing agents.