SOURCE-BOUND AI CLAIM AUDIT

Does the cited source actually support the AI claim?

SourceAudit evaluates the relationship between a claim and the source supplied to support it. It is designed to catch unsupported additions, contradictions and material mismatches before a source-backed answer is trusted.

Open GPT surface ↗Request integration

20,000HaluEval QA samplesdocumented strict-mode benchmark

93.96%accuracynamed dataset / mode only

93.99%F1named dataset / mode only

94.36%hallucination recallnamed dataset / mode only

SourceAudit

Conversational and API-facing source verification. A claim is evaluated only against the declared evidence boundary. It does not claim to be a universal fact-checker.

Example

Source:
"The company has 24 million customers."

Claim:
"The company has 34 million customers."

→ BLOCK / CONTRADICTED
numeric mismatch

Guardian Runtime

The same verification philosophy can be placed before an LLM response is shown to the end user. Guardian Cloud is the pre-output deployment mode: inspect a draft, then veto / request rewrite when the evidence boundary is not met.

Guardian is presented as a deployment mode / enterprise capability rather than a separate flagship SKU because the full client-side authenticated end-to-end path still requires deployment-specific validation.