The system uses a dedicated fact database where the LLM inserts records by key rather than recalling numbers from memory. An output gate verifies that every number in the text matches a database record; any number without a corresponding entry is blocked regardless of its plausibility.
The architecture incorporates three levels of attribution: official requirements (linked to government sources), practice (dated observations), and internal standards. The gate ensures that practice-level information is not presented as an official requirement, and upgrading a fact's level requires human intervention.
To prevent outdated information, the system implements a Time-to-Live (TTL) mechanism. Records older than 90 days are considered invalid and blocked by the gate, while entries older than 75 days are queued for re-verification.
The pipeline uses a separate verification pass—distinct from the generation context—to check for number matching, attribution levels, and a semantic stop-list. This stop-list uses an LLM to identify paraphrased prohibited phrases, such as guarantees of approval, which traditional regex patterns cannot detect.
Common industry myths are handled through an adversarial process where multiple independent agents attempt to debunk a statement based on primary sources. A consensus is reached only if independent passes agree, allowing the system to categorize certain widespread beliefs as myths.
Final publication is manually approved by a human via an approvedBy field. The system maintains an audit trail for every fact, documenting its origin, verification date, and the person who approved it to ensure accountability and the accuracy of the base data.
