Provenance, citations, and claim support

Hard-wiring auditability by mapping LLM assertions back to source text spans.

Why it matters

A sophisticated graph means nothing if users can't verify the final synthesis. By mapping every extracted node and edge back to its original chunk and text span, you create a direct paper trail. This ensures the generation layer can ground every single claim with bulletproof inline citations, transforming a black-box model into a trusted, auditable system.

In a regulated or high-trust setting, an answer is only as good as its evidence trail. Professional GraphRAG attaches support to each claim, not one citation block bolted onto the end.

{
  "answer": "Alice approved the exception under Policy P-12.",
  "claims": [
    {"text": "Alice approved the exception",
     "support": {"graph_edges": ["Alice-approved-Exception42"], "chunks": ["doc://approvals/42#L10-18"]}},
    {"text": "Exception42 is governed by P-12",
     "support": {"graph_edges": ["Exception42-governedBy-P12"], "chunks": ["doc://policy/p12#L3-11"]}}
  ]
}

Why per-claim, not per-answer. A single citation block lets one unsupported sentence hide among three supported ones. Per-claim support makes every sentence independently verifiable — and makes the unsupported claim stick out, so a critic node (or a human) can reject exactly it instead of the whole answer.

Store it as structured fields, not markdown. claim_id, source_id, edge_path, timestamp are queryable; a markdown footnote string is not. Structured provenance lets you audit ('show me every answer that cited the now-retracted doc'), prefer the newest valid evidence, and prove compliance.

Pair it with a refusal contract. If a claim has no support, the system must say so or refuse — provenance is what makes 'I don't have evidence for that' a deterministic behaviour rather than a hope.

Per-claim citations are receipts for each line item, not one lump total at the bottom of the bill. With line-item receipts an auditor can challenge a single charge; with only a total, one padded charge hides inside a plausible sum.

• One citation block per answer. Symptom: unsupported sentence hides. Fix: support per claim.
• Citations as markdown text. Symptom: not auditable/queryable. Fix: structured fields.
• No refusal when support is missing. Symptom: confident gaps. Fix: explicit 'insufficient evidence' path.
• Stale evidence preferred. Fix: attach timestamps; prefer newest valid source.