Factual hallucination gaps represent the measurable divergence between a language model's ungrounded parametric memory and verified, real-world source data. Rather than treating hallucinations as random errors, automated audit pipelines expose systemic blind spots where an AI's internal model weights distort or erase verifiable reality.

+-------------------------------------------------------------+
|               Divergence Diagnostic Workflow                |
|                                                             |
|   Live Source Content  <=========>  Parametric Output   |
|            |                                    |           |
|            +-----------------+------------------+           |
|                              |                              |
|                              v                              |
|               Extraction & Comparison Phase               |
|                              |                              |
|          +-------------------+-------------------+          |
|          |                   |                   |          |
|          v                   v                   v          |
|     Omissions        Distortions      Fabrications    |
|   (Excluded Data)    (Skewed Context)   (Invented Claims)   |
|          |                   |                   |          |
|          +-------------------+-------------------+          |
|                              |                              |
|                              v                              |
|                 Fidelity Score Calculation                |
+-------------------------------------------------------------+

This verification process relies on two complementary analytical mechanisms:

1. The Veracity Engine's multi-phase evaluation: This subsystem executes grounded queries across target URLs, fetches the underlying source text via tools like url_context, and evaluates the output against source segments. It classifies content into distinct buckets:
2. Survived: Accurate, verifiable facts faithfully retained from the source.
3. Lost: Key source facts, statistics, or product details omitted during synthesis.

4. Distortions: Inaccuracies, conflated features, or contextual errors introduced by the model.

5. The Fact & Gap Checker's tripartite classification: By systematically comparing research findings against recalled assertions, this process categorizes claims into three diagnostic relations:

6. Confirms: The grounded evidence validates an existing parametric claim as factual.
7. Contradicts: Retrieved evidence directly disproves a parametric claim, flagging it as an active hallucination.
8. New (Gaps): Verified real-world brand facts that are entirely absent from the model's parametric memory.

By calculating per-URL and aggregate fidelity scores—measured as survived / (survived + lost + distortions)—the system quantifies informational decay. These insights feed directly into a persistent claims ledger system, providing a concrete baseline to guide structured content optimization, eliminate hallucinated claims, and force generative AI engines to reflect ground truth.