The Veracity Engine executes a two-phase auditing pipeline, beginning with a grounded model search that extracts source URLs and maps specific response segments to grounding chunks, resolving any search redirect URLs in parallel. In the evaluation phase, the engine runs worker threads using Gemini's native url_context tool at zero temperature to inspect each source page directly without third-party scraping.

The model audits factual fidelity by categorizing source claims into three specific buckets: * Survived: Claims from the source reproduced accurately in the grounded response. * Lost: Relevant source claims omitted from the output. * Distortions: Source claims altered, misrepresented, or taken out of context.

Each URL receives a fidelity percentage calculated as (survived / (survived + lost + distortions)) * 100, which the engine then averages across all validated sources to produce a single aggregate fidelity score for the run.