Feed an AI a Selling Guide version that was superseded last quarter and it will cite the old rule with total confidence. In mortgage, that is not a bad search result. It is a repurchase-worthy defect with a citation attached.
Garbage in, garbage out applies with teeth here, because the “garbage” is a rule that was true last year. An AI knowledge layer is only as trustworthy as the discipline behind what it ingests.
Knowledge lifecycle
- Ingestion gates. Nothing enters the index without an owner and an effective date. A guideline chunk with no date is a liability.
- Versioning. When Fannie updates a section, the prior version is marked deprecated and the system always cites the current one. The UAD 3.6 and redesigned URAR transition, mandatory for appraisals submitted to the GSEs from November 2, 2026, is the live example: an AI answering appraisal questions has to know which form regime applies to which loan.
- Review loops. Owners get pinged when a source has not been reviewed in 90 days: still current, or superseded?
Handling contradictions
The Selling Guide says one thing; an investor overlay says something stricter. A human knows the overlay wins for that investor. An AI needs that ranking made explicit through hierarchy metadata:
Agency Selling Guide > Agency bulletins and announcements > Investor overlays > Lender SOPs
When retrieval pulls conflicting chunks, the hierarchy resolves which one governs, and for which investor.
This is the difference between a chatbot that sounds right and a system you can cite in a QC defense. Currency, deprecation, and a clear authority order are not nice-to-haves in a regulated file. They are the product.