What the GSEs already know: a credit score is not enough.

For conforming mortgages, the practical gate is the GSE-controlled automated underwriting system: Fannie Mae’s Desktop Underwriter or Freddie Mac’s Loan Product Advisor. The credit score is part of the mortgage ecosystem, but it is not the whole decision.

Fannie Mae now says Desktop Underwriter no longer requires a minimum third-party credit score and that DU does not use third-party credit scores to assess credit risk. DU uses Fannie Mae’s proprietary credit risk assessment instead. The score can still travel with the loan file for delivery, reporting, pricing, investor, or lender-policy purposes; it just is not the full automated underwriting judgment. Fannie Mae.

The GSE channel already knows the single number is not enough. The mortgage decision is a governed system, not a score lookup.

This distinction matters. FICO, VantageScore, and the credit bureaus are not blind to history. They are built from credit-history variables: payment history, balances, utilization, depth of credit, new credit, available credit, and related tradeline behavior.

So the claim is not that credit scores ignore history. The claim is narrower: credit-performance history is not the same as source-corroborated evidence that the claimed person, place, timeline, and obligation pattern cohere. A file can contain real tradeline history and still leave unanswered questions about identity coherence, post-shock recovery, source authenticity, or whether the evidence fits the loan in front of the lender.

Fannie Mae publishes the broad risk-factor taxonomy Desktop Underwriter evaluates. It includes credit-report factors such as credit history, delinquent accounts, installment accounts, revolving utilization, public records, collections, inquiries, and rent payment history. It also includes application factors such as LTV, reserves, loan purpose, term, amortization, occupancy, DTI, housing-expense ratio, property type, first-time-homebuyer status, and cash-flow assessment. Fannie Mae Selling Guide B3-2-03.

Fannie Mae also says DU weighs factors based on risk and importance to the recommendation, and that no one risk factor determines the recommendation. Several high-risk factors without sufficient offsets can drive a refer or ineligible recommendation; sufficient offsetting factors can support a stronger file. Fannie Mae DU V.12.0 whitepaper.

The public record does not reveal the GSE weights. It does reveal the architecture: factors, offsets, eligibility, and lender responsibility.

This is where the boundary has to stay clean. Kenshiki does not feed Desktop Underwriter or Loan Product Advisor, and it does not claim to improve a conforming-loan recommendation. The GSEs accept recognized input types through authorized verification channels; a proprietary signal outside that recognized set cannot simply upgrade a recommendation or make an otherwise-ineligible loan salable.

The asymmetry matters. Lenders still own material information they know, including derogatory information the automated system may not have seen. But that does not create a matching right to use a positive outside signal as if it were part of the GSE engine. Non-sanctioned data remains the lender’s responsibility for representation-and-warranty, legal-compliance, and fair-lending purposes. Fannie Mae Selling Guide B3-2-01.

Synthetic identity and fabricated-document risk are not solved by turning Kenshiki into a DU or LPA input. They are lender-side funding risks. The question is whether the file describes a real borrower with a coherent source pattern before principal leaves the lender.

That is where Kenshiki belongs in mortgage: not as a positive eligibility signal, but as pre-funding evidence that the claimed identity, source documents, device, residence, obligations, and timeline cohere. A clean credit file can still be engineered around a fabricated life; source-side corroboration gives fraud and underwriting teams evidence they can inspect before the loan funds.

Public reporting already frames the pressure: synthetic-identity fraud losses are industry-modeled in the tens of billions, and generative AI is making fabricated identities and supporting documents cheaper to produce. Those are lender exposure problems before they are model-input problems. Boston Fed.

Representation-and-warranty relief does not make every post-closing problem disappear. Data inaccuracies, misrepresentation, legal-compliance issues, and fraud can still matter after delivery. When a file is questioned later, the lender needs more than a point-in-time score or an unexplained internal flag.

Kenshiki’s value is the replayable evidence record: what was checked, which source patterns aligned or contradicted, what the reviewer saw, and why the lender treated the file as fraud risk, manual-review risk, or acceptable risk. That record does not change the GSE recommendation. It helps the lender defend the judgment it actually made.

That value is concrete because it lives where mortgage losses often surface: pre-funding stop decisions, post-closing review, repurchase exposure, and examiner questions. Fannie Mae’s own representation-and-warranty relief framework preserves enforcement paths for certain breaches, including misrepresentation and data-quality problems. Fannie Mae Selling Guide A2-3.2-02.

Kenshiki’s mortgage fit is lender-held evidence: a record the underwriter, fraud analyst, reviewer, auditor, or examiner can open later. It can support pre-funding fraud review, manual underwriting, overlay justification, adverse-action support, repurchase defense, and the lender’s own explanation of why the file was treated the way it was.

The product claim is not “another score.” It is not “Kenshiki replaces DU.” It is not “Kenshiki sees life trajectory while FICO only sees the present.” The sharper and more defensible claim is this: credit history is real evidence, automated underwriting is already multi-factor, and consequential mortgage decisions still need inspectable evidence about whether the claimed identity and supporting file cohere.

Adverse-action rules do not create a black-box exemption for complex algorithms. A creditor cannot justify noncompliance by saying the technology was too opaque to understand. CFPB Circular 2022-03.

GSE AI and model governance point the same way. FHFA requires the enterprises to manage AI and machine-learning risks, including third-party model risk, explainability, conceptual soundness, and fair-lending controls. FHFA AB 2022-02. SR 26-2 reinforces the broader model-risk posture for vendor and third-party models. SR 26-2 and model validation.