Skip to content

Trust and Provenance

Placeholder — this page needs to be written. Content currently lives in fragments across the pipeline and storage docs.

Provenance is the philosophical core of this project. Every assertion in the graph must be traceable to a source document, a location within that document, and the extraction step that produced it. Transparency is not an afterthought — it is load-bearing.

Source attribution

Each entity and relationship carries a reference to:

  • The source document (title, DOI, or other identifier).
  • The chunk or section within that document where the mention appears.
  • The text span — the exact string the extractor saw.

This means a user or downstream system can always ask: "where did this come from?" and get a meaningful answer.

Confidence scores

Confidence is attached at extraction time. It reflects:

  • The LLM's expressed certainty (where available).
  • The quality of the source (peer-reviewed vs. grey literature).
  • Whether the entity resolved to a canonical ID or remained provisional.

Confidence scores are not hidden; they are surfaced in query results and available for filtering.

Conflicting claims

When two sources make contradictory assertions about the same entities, the graph records both claims — with their respective provenance and confidence — rather than silently picking one. See Conflicting Claims.

Audit trail

The full chain is preserved:

raw document
  → chunk
    → extracted mention
      → resolved entity / provisional entity
        → graph assertion (edge)

At any point in this chain, you can walk backward to the source.

See also