In the initial phase, the engine executes a grounded search to capture raw response text, grounding metadata chunks, and specific cited segments, resolving intermediate redirect URLs to their final destinations. In the second phase, a background worker pool evaluates each resolved URL using the url_context tool, extracting source claims into survived, lost, and distorted categories to compute individual and aggregate fidelity scores via the formula:

\[ \text{Fidelity Score} = \frac{\text{survived}}{\text{survived} + \text{lost} + \text{distortions}} \times 100 \]