Claims
json
{
"kind": "claim",
"text": "The bridge opened in 1894.",
"sources": ["https://example.org/history", "https://archive.example.com/1894"]
}sources are the URLs the agent is relying on. They are never checked against the agent's own output.
The check: claim.corroboration
A heuristic check. It does not ask a language model whether the claim sounds plausible: that would be the generator grading its own work.
Today it counts the distinct domains among the sources (www. is ignored), and fetches nothing:
- Fewer than two: fail, with a
mediumfinding such asOnly 1 distinct source domain(s). - Two or more:
skipped, because the sources are not fetched or matched against the claim yet, so the verdict is unknown.
A claim can fail, but is never verified as pass: a heuristic can't pass a target on its own.
Not checked
- Interpretation: what the claim means to its audience
- Whether the cited sources are themselves reliable