Provenance & Grounding

A triple without provenance is a rumor with good punctuation. Attach to every edge (or every statement, via RDF reification): source document, character span, extraction timestamp, extractor version, and confidence.

Now 'where did this claim come from' is a query, not an investigation — and when extraction model v2.3 turns out to have a bug, the cleanup is WHERE edge.extractor_version = 'v2.3' rather than 'rebuild the graph'. Provenance is cheap to write and impossible to retrofit.