Skip to main content
Once a graph has approved content, Nora uses it as a safety net: the graph’s live knowledge joins the evidence that answers are checked against, so an answer that contradicts your causal model gets flagged.

​
Where verification runs

Verification lives on the Trace. Each trace detail carries a Verification block you can run on demand: it decomposes the answer into claims, searches for evidence, and returns a verdict (supported / partial / insufficient / unverified). See Grounding in a Trace.

​
What feeds verification

Only approved (live) nodes and edges participate. Proposals waiting for review don’t affect anything, which is exactly why the review step matters:
  • Approve items in their Review tab, or batch-approve with the toolbar’s Approve all button.
  • Demote a live item with To proposal and it immediately stops feeding verification.
The quality levers on each edge also shape checking: an edge’s sign, its when does it apply? condition, and its confidence all describe how strongly the graph commits to a claim.

​
No per-Agent setup

There is no verification toggle on the Agent. The graph reaches the verifier through the data sources an Agent already uses: if the Agent is bound to a folder the graph is connected to, the graph is in scope. See How agents pick up a graph.

​
Keeping the graph trustworthy

Verification is only as good as what you approved:
  • Review LLM-built proposals before approving: origin and confidence sit on each item’s Evidence tab.
  • Mark checked edges reviewed so LLM restructuring can’t silently change them.
  • Keep speculative edges as proposals (or dashed, non-causal links) until you have evidence.