In the Veracity Engine, claim tracking begins during grounded search by parsing groundingSupports metadata to tie specific model response segments directly to resolved source URLs. A secondary evaluation call uses the url_context tool to inspect each source page and compare it against those attributed segments, categorizing every claim into survived, lost, or distorted arrays. A mathematical fidelity score is then derived per URL as the percentage of total claims that survived accurately, which is finally rolled up across all cited sources into an overall aggregate fidelity metric.