Grounded fidelity scoring calculates the percentage of factual claims from cited source pages that survive accurately in an AI model's output relative to the total volume of extracted facts.