Canonical verdict vocabulary shared across all grounding/citation pipelines
in the AgentOS ecosystem (CitationVerifier core, the Grounding
Guard extension, future implementations).
'supported' — the claim is entailed by at least one source.
'weak' — confidence sits between the support and unverifiable
thresholds. Some implementations (e.g. the NLI-based
Grounding Guard) collapse this into 'unverifiable';
cosine-similarity-based pipelines surface it as its
own bucket.
'unverifiable' — no source supports or contradicts the claim with
sufficient confidence.
'contradicted' — at least one source contradicts the claim.
Canonical verdict vocabulary shared across all grounding/citation pipelines in the AgentOS ecosystem (CitationVerifier core, the Grounding Guard extension, future implementations).
'supported'— the claim is entailed by at least one source.'weak'— confidence sits between the support and unverifiable thresholds. Some implementations (e.g. the NLI-based Grounding Guard) collapse this into'unverifiable'; cosine-similarity-based pipelines surface it as its own bucket.'unverifiable'— no source supports or contradicts the claim with sufficient confidence.'contradicted'— at least one source contradicts the claim.