Candor DI Document Reconciliation: Preserve the Evidence

Candor DI brings document reconciliation to loan intake, according to the company’s September 1, 2026 launch announcement. For lenders, the key test is whether a resolved field remains traceable to the competing source documents and the decision that settled the conflict. A vendor’s description of controls does not establish that a lender’s implementation meets GSE requirements.

What Candor says is available

In its launch release, Candor describes three ways to deploy DI: as a standalone product, through an application programming interface (API), or within its Loan Engineering System (LES). The company says DI compares data extracted from documents with the loan application, credit report and automated underwriting findings, using validation controls and source-level histories for flagged discrepancies. The release describes its URLA Updater as forthcoming.

In a demonstration using your chosen loan origination system (LOS), check which fields enter automatically, which can be changed and which actions require approval. Have the vendor distinguish available functionality from planned work. Weigh those integration details alongside price and implementation effort when selecting mortgage workflow technology.

Why reconciliation goes beyond OCR

Separate three tasks when evaluating Candor document intelligence:

Finding a difference does not establish which value should govern. Consider a hypothetical application reporting $6,000 in gross monthly pay and an employment verification reporting $5,600. First check effective dates and pay components. If the figures describe different periods, both may be accurate. If they describe the same income for the same period, automatically choosing the newer document would leave the disagreement unexplained.

The workflow should retain both figures, record any additional evidence and route unresolved questions to an authorized reviewer. Correcting a transcription error and deciding qualifying income require different reasoning. Our guide to resolving mortgage income discrepancies addresses that distinction from the borrower’s side.

Keep the evidence behind each resolved field

Once an exception is resolved, another reviewer should be able to reconstruct the decision. The following is a proposed procurement acceptance test, not a prescribed GSE form or a verified description of DI’s implementation.

Require a demonstration of a later document reopening a previously resolved exception. Also test an override and an export of the history. A reviewer should be able to follow the resolution after the processing session ends.

This proposed test draws on the emphasis on documentation and human oversight in NIST’s January 2023 AI Risk Management Framework. It gives lenders a practical way to assess accountability in housing AI and automation.

Vendor controls do not establish lender compliance

Fannie Mae’s LL-2026-04, issued April 8, 2026, took effect 120 days later, on August 6. It covers artificial intelligence and machine learning (AI/ML) used to originate loans sold to or guaranteed by Fannie Mae, or to service loans on its behalf. Seller/servicers must maintain policies, risk management and policy ownership, with at least annual review. Vendor and subcontractor governance must be no less protective of these requirements. The letter also requires compliance with Fannie Mae’s Information Security and Business Resiliency Supplement and prompt disclosure of AI uses and safeguards when requested.

Freddie Mac’s December 3, 2025 Bulletin 2025-16 announced an AI/ML governance framework through changes to Sections 1302.2 and 1302.8, with a March 3, 2026 effective date.

Against Fannie Mae’s requirements, DI’s stated validation controls and discrepancy histories address only part of the lender’s responsibilities. A discrepancy log does not demonstrate policy ownership, risk management or vendor oversight. Neither cited GSE document certifies Candor DI. Before deployment, have compliance counsel review the proposed use, contract and allocation of responsibilities.

Test productivity separately from accuracy

The launch release quotes The Federal Savings Bank’s Erin Devaney estimating 15–20 minutes saved per file on indexing and classification. It supplies no sample size, study period or comparison protocol. Without those details, the testimonial does not provide an independently reproducible result. It also does not establish faster underwriting, fewer material errors or a staffing reduction.

For a lender pilot, measure indexing time separately from exception-review time and elapsed processing time. Include clean files, poor scans, conflicting documents and later revisions. Count missed discrepancies and unnecessary alerts, then have reviewers check the accepted resolutions. A rule that catches more discrepancies may also generate more review work; evaluate both effects.

Record the software version, sample selection and comparison workflow so the result can be assessed independently. NIST recommends documenting test sets and metrics and evaluating performance under conditions resembling deployment in AI RMF, MEASURE 2.1 and 2.3. The lender-specific measurements above are proposed applications of that principle.

Before signing, ask the provider to trace one disputed field from its original documents through resolution and the LOS update, then export the evidence. The record should let another reviewer understand what changed, why and who or what authorized it.