The Veracity Engine operates a two-phase assessment pipeline to evaluate how accurately an AI model represents its reference material. First, it captures the generated text, extracts the cited claims, and maps them directly to resolved source URLs using the provider's grounding metadata. In the background, it runs parallel evaluation queries using Gemini's url_context tool to compare each webpage against its specific attributed text segments, sorting the extracted claims into three categories:

• Survived: Accurate facts preserved from the original webpage.
• Lost: Relevant source claims omitted from the AI's final answer.
• Distortions: Source claims altered, taken out of context, or misrepresented.

Each analyzed page receives a score calculated as survived ÷ (survived + lost + distortions) * 100, which is then averaged across all cited URLs to produce the final aggregate fidelity score.