grounding-verification

LLM hallucination detection compares scrubbed versus full evidence passes to flag confabulation. This grounding verification calculates a delta between the two passes to output a clear verdict, such as VERIFIED or UNSUPPORTED, for your claims.

Two-pass grounding verification is a deterministic method comparing an anonymized evidence pass against a full detailed pass. It calculates a delta score to determine if LLM claims are reliable and outputs a structured, auditable verdict.

You audit LLM claims against API schemas by running them through a profile-driven verification process. Using domain-specific thresholds for code or data schemas, the process outputs a structured report indicating whether the claims are VERIFIED or SUSPICIOUS.

Yes, grounding verification works across domains like code, documentation, data schemas, and security. It applies profile-based thresholds tailored to each specific domain to evaluate the reliability of LLM claims.

To run a grounding verification check, you need to provide frontmatter with a name and a verification profile. You must also prepare two passes of evidence, one scrubbed and one full, to compare against your LLM claims.