In Phase 2 of the Veracity Engine, assess_url invokes Gemini with the url_context tool enabled at a zero-temperature setting to fetch and analyze the web page in a single pass. Rather than testing the entire grounded answer, the prompt targets only the specific segments previously mapped to that source URL via groundingSupports metadata.

The model parses the live document and returns a structured JSON payload categorizing content into: * Survived: Claims from the page faithfully represented in the AI's output. * Lost: Relevant facts present in the source that the output omitted. * Distortions: Source claims that the model altered or misrepresented.

A deterministic fidelity score is then computed from the ratio of survived claims to total extracted claims, while prompt and candidate token counts are saved to meter the assessment run.