Mortgage Intelligence

QC · AVAILABLE FOR EVALUATION

Mortgage Intelligence for Quality Control.

Review the whole file, preserve the evidence, and focus human judgment where it matters. Running on real files today, on sample, synthetic or de-identified historical files, with every finding carrying the rule version and the source page behind it.


Why QC is hard

Most defects live between the documents

Experienced reviewers already do this work, and do it well. The constraint is not skill: the checks are numerous, the file is long, the policy keeps moving, and the time is finite.

EXAMPLE FINDING · SYNTHETIC FILERULE XDC-INCOME-RECONCILE
QUALIFYING INCOME, FOUR SOURCES SIGNIFICANT
The documents disagree, and one of them is the one that qualified the loan
Paystub annualizes to $8,940. W-2 gives $8,875. The 1003 states $9,200 and the 1008 used $9,200 to qualify. Spread 3.7%, inside the 5% tolerance. Not a defect. Raised because the 1008 used the stated figure rather than a derived one, and a reviewer decides which basis stands.

No single document is wrong. The finding only exists when the file is read as one thing. That is what cross-document reasoning means, and it is where Mortgage Intelligence shows first.

What a miss costs
$32,288Estimated average cost per repurchase demand. National Mortgage News, drawing on STRATMOR Group analysis.
What the rate hides
1.50%Critical defect rate, CY2025, flat against 1.52%, while eligibility defects rose 291% and credit 166%. ACES Mortgage QC Industry Trends.
What a loan costs to produce
$11,109Fully loaded total production cost per loan, not QC cost. MBA Quarterly Performance Report, Q3 2025. No per-loan QC figure is published industry-wide.
What changes
Designed for every loan, not a sampleFull-file, cross-document checks rather than a sampled read, so depth stops being the variable that quietly gives. What is on offer today is a controlled evaluation: see the capability matrix below.

Whole-file workflow

Seven stages, and two boundaries

Models perceive. Versioned rules enforce. A person resolves the exception, and every step is kept with the finding.

How QC works Seven stages. Intake, Classification and Extraction are performed by a vision-language model. Synthesis and Validation are deterministic and contain no model. Review is performed by a person. Evidence is the retained record. MODELS PERCEIVE · PERCEPTION 01Intake Loan file receivedand queued. 02Classification Documents identifiedand segmented. 03Extraction Values located, keptwith page and source. 04Synthesis Reconciled across thefile; conflicts recorded. RULES ENFORCE · NO MODEL 05Validation Versioned rules run.Skips recorded too. 06Review Exception raised.A reviewer resolves it. 07Evidence Finding kept with therule version that made it. PEOPLE DECIDE · RECORDED Models perceive. Versioned rules enforce. A person resolves the exception, and every step is kept with the finding.
Reading, reconciling, deciding, resolving, and what is kept from each stage.
01

Intake

Loan file received and queued.

02

Classification

Documents identified and segmented.

03

Extraction

Values located, kept with page and source.

  • Verify: a person corrects what the system is unsure of, before rules run
04

Synthesis

Reconciled across the file; conflicts recorded.

05

Validation

Versioned rules run. Skips recorded too.

06

Review

Exception raised. A reviewer resolves it.

  • Findings: what the reviewer opens
  • Disposition: a named person decides, and it is recorded
07

Evidence

Finding kept with the rule version that made it.

  • Report: the exported record, with the rules that did and did not run

Classification and extraction

What the system reads now

A vision-language model classifies each document against the supported form set and locates the fields. Every value is stored with the document, the page and its position on the page.

CLASSIFY
Segmentation and classification
Documents identified and split across the supported form set, with review of unclassified pages. An unread page is exactly where a missed defect hides, so pages that cannot be classified are surfaced rather than dropped.
EXTRACT
Fields read, with confidence retained
Values extracted with confidence captured; low-confidence values highlighted for inline correction.
VERIFY
A person corrects what the system is unsure of
Judgment stays where the system is least confident, and it stays before the rules run, so a misread does not propagate into a finding that then has to be argued about.

Synthesis

Documents become one loan record, with conflicts visible

Borrower, property, income, asset and liability data are consolidated into a single structured record, mapped to MISMO field paths so the same fact means the same thing across every document that touches it.

Where two documents disagree, the conflict is carried forward as a conflict rather than resolved silently by whichever document was read last. The resolution strategy, where one is applied, is recorded with the value. Derived values show the formula that produced them.

This is the layer where reading becomes mortgage understanding: not what the paystub says, but what qualifying income is for this borrower, under this program, given four sources that do not agree.


Controls

The policy the rules enforce, encoded and versioned

Not a checklist held in someone's head. Each control is a stored rule, bound to a loan program, and the record shows which ran and which were skipped.

TRID
Disclosure content and timing
Loan Estimate within a period of application; Closing Disclosure received a required interval before consummation; tolerance comparisons between the two.
ATR / QM
Ability to repay
A documented, verified determination, with the points-and-fees test and the safe-harbor threshold applied by loan pricing.
HOEPA
High-cost triggers
Additional protections and restrictions where rate or fee thresholds are exceeded.
HMDA
Reported data accuracy
Loan Application Register values checked against the file they were drawn from.
RESPA
Settlement services
Section 8 referral and unearned-fee prohibitions; Section 10 escrow administration.
ECOA / Reg B
Adverse action and timing
Notice content and the period within which it must issue.
FCRA
Consumer reports
Permissible purpose and risk-based pricing notice timing.
Agency programs
FHA · VA · USDA
Program eligibility and the layered documentation logic that stacks rather than replaces.
Selling guides
Investor requirements
The operative standard for conforming loans, encoded as versioned rules rather than remembered.

Rule sets are selected by loan program, so an FHA file and a bank-statement file do not run the same checks. Which rules were skipped, and why, is recorded alongside the ones that ran. Encoding a published requirement is not endorsement by, affiliation with, or certification from any agency or GSE.


Findings

Four outcomes are recorded, not two

Every rule returns one of four states, and the distinction between the last three is the point.

PassedThe rule ran and the file satisfied it.
FailedThe rule ran and the file did not satisfy it.
Data requiredThe rule could not run because a field it needs is empty. Reported, not folded into passes.
Not applicableThe rule did not apply to this loan program or file. Recorded as skipped, with the reason.

A system that reports only pass and fail turns data required and not applicable into silence, and silence reads as clean. Each finding carries a severity, the rule and its version, the values used and their sources, and a status a reviewer moves it through.


Evidence

What is kept with every finding

Six months after a file is cleared, someone asks why. A finding that cannot answer is an opinion with a timestamp.

The reconciliation passes at 3.7% against a 5% threshold. The exception is about basis, not arithmetic. FIGURE Anatomy of a finding Four sources, four periods, one comparison, and why the difference needs a person. RULE INC-014 v2.3.0 · QUALIFYING INCOME MUST RECONCILE WITHIN 5% ACROSS SOURCESSOURCEAS STATEDMONTHLYHOW DERIVEDPaystub 03/2026p.1 · YTD box$26,820 YTD over 3.0 months$8,940 / moYTD ÷ months elapsedW-2 2025p.1 · box 1$106,500 for 12 months$8,875 / moannual ÷ 12URLA 1003p.2 · 1c$9,200 stated monthly$9,200 / moas stated by borrower1008 Summaryp.1 · field 21$9,200 used to qualify$9,200 / mounderwriter’s figureSpread between lowest and highest monthly figure: 3.7% inside the 5% threshold.Not a defect. Flagged for review because the 1008 uses the stated figure, not the derived one.The reviewer decides: accept the underwriter’s basis, or request a written income calculation. the system records which,by whom, and when. The finding is an exception to investigate: the disposition belongs to the reviewer. Illustrative, from a synthetic file. Threshold and rule identifier are examples.
The reconciliation passes at 3.7% against a 5% threshold. The exception is about basis, not arithmetic.

Recorded on every run: the values used, with the source document each came from; confidence per value; which source won where documents disagreed, and on what basis; the MISMO field path; the rules that ran; the rules that were skipped; the loan program applied; and the rule set version. Anatomy of a finding, taken apart in full.

What is retained today: findings and synthesized values keep their lineage to the source data and the source document. What is still being built: presenting that lineage down to page and field level inside the reviewer interface. The record exists; the interface for reading it at that granularity is on the roadmap.


Governance

“You are using AI. Show us what it did, and how you knew it was right.”

Fannie Mae's LL-2026-04 took effect 6 August 2026; Freddie Mac's Bulletin 2025-16 has been live since 3 March. Both require a documented AI/ML governance program, and both hold you to the same standard for the AI your vendors use.

What is the AI doing, and why

Each stage is named and bounded. A model reads; deterministic rules produce the findings; a person disposes. The purpose of the AI is answerable in one sentence rather than described.

What safeguards are in place

The rule layer contains no model, so the rule applied and its version are fixed. That is a real safeguard and a bounded one. It does not make extraction error harmless.

How do you know it got it right

Every value carries the source document it was read from, and every finding the numbered rule version. Given the same inputs and version, the same result follows, so a disputed answer is traceable to a specific value on a specific page.

Can you evidence it for a vendor

The obligation extends to vendor and subcontractor AI, held to the same standard. The record exists as a by-product of the review, not as a document written afterwards.

The four questions, and what the review produces in answer to each. FIGURE The questions you now have to answer Fannie Mae LL-2026-04, effective 6 August 2026. Freddie Mac Bulletin 2025-16, live since 3 March 2026. WHAT THE GSEs RESERVE THE RIGHT TO ASKWHAT THE REVIEW PRODUCESWhy is AI being used?To read documents: a perception task.Named and bounded per stage.For what purpose?Locating values, not deciding outcomes.The disposition comes from a stored rule.What safeguards exist?The deciding path contains no model.Structural, not procedural.How do you know it was right?Page, position and rule version, per finding.The review can be run again.The obligation extends to vendor AI, held to the same standard as your own. The platform produces the evidence; the lender holds the obligation. No approval or endorsement is implied.
The four questions, and what the review produces in answer to each.

The platform produces the evidence; the lender holds the obligation. Nothing here implies approval or endorsement by Fannie Mae or Freddie Mac.


Human review

A named person decides, and that is recorded

Reviewers correct extraction, assess findings and make the final loan decision. That is a design position and it is not on a roadmap to be removed.

TODAY
Correction, acceptance, rejection
Reviewers inspect and correct uncertain values before rules run, then accept, correct or escalate each finding. The loan-level disposition is captured with the deciding reviewer recorded.
TODAY
A report with export
A quality control report is produced with export. Rules that did not run are in the report as rules that did not run.
ROADMAP
Queues, assignment, escalation, service levels
Today a reviewer works the list; the workflow around it is coming. Investor and regulatory reporting packages are target, not current.

Current capability

What runs today, and what is still being built

Stated as two columns rather than one, because the distance between them is the thing a buyer needs and the thing vendors usually hide. Current as of the September 2026 capability review.

Running today

These run in the deployed environment now. A rules specialist or an audit lead can be shown them without qualification.

INTAKE
Loan files processed end to end
Upload sessions, documents and run relationships, processing a complete package rather than isolated documents.
CLASSIFY
Segmentation and classification
Documents identified and split across the supported form set, with review of unclassified pages.
EXTRACT
Fields read, with confidence retained
Values extracted with confidence captured, low-confidence values highlighted for inline correction.
VERIFY
Reviewers correct what the system is unsure of
Human judgment stays where the system is uncertain, before anything downstream runs.
SYNTHESIZE
One structured loan record
Borrower, property, income, asset and liability structures consolidated, with conflicting values flagged and derived values shown.
VALIDATE
Versioned rules produce findings
Rule execution with severity, rule-level detail, and the findings workflow.
DISPOSITION
A named reviewer decides
Loan-level disposition captured with the deciding reviewer recorded. The system produces findings; a person decides.
REPORT
A quality control report with export
Produced today. Investor and regulatory reporting packages are target, not current.

The strongest part: rules, versioning and lineage

The rule model, execution, explainability and lineage can be demonstrated now; the authoring and approval interfaces around them are in the group below.

RULE SETS
Versioned, with explicit relationships
Rule sets, rules and versions implemented, so a finding can name the version that produced it.
PROGRAMS
Program to rule-set mapping
Loan programs map to rule sets, so different programs run different rules.
EXECUTION
Execution plans, results retained per run
So a run can be reconstructed rather than described.
EXPLAINABILITY
Rule logic shown with the outcome
Results carry the logic alongside the result and link back to the rule.
LINEAGE
Back to the source document
Findings and synthesized values retain lineage to the source data and documents.

Implemented, awaiting verification

Everything in this group is implemented; what it has not yet had is demonstrated behavior we would put in front of you as evidence.

APPROVAL QUEUE
Data model implemented, interface to demonstrate
Role-based approval rights and change history are target.
RULES CATALOG
Metadata modeled; catalog interface to verify
Usage analytics and impact analysis are target.
RULE STUDIO
Input mapping and testing designed in
An authoring experience usable by a rules specialist without engineering support is target.
NEEDS DATA / NOT APPLICABLE
Designed into the model, live behavior to demonstrate
Recording that a rule did not run, so “no finding” and “not checked” stay distinguishable.
SCHEDULED RUNS
Configuration present, end-to-end behavior to demonstrate
Retry, failure handling and alerting are target.

In development

Where current and target are furthest apart, stated so timing can be planned rather than discovered.

DEPLOYMENT
Vendor-managed cloud environment today
We are building a Terraform deployment that stands the platform up inside a customer's own AWS account. Not available yet; it is the work that closes the two rows below.
DATA RESIDENCY
Documents are processed in our environment
Under the deployment above, loan files and the processing that reads them would stay inside your account. That is the destination, not today's position.
INHERITED CONTROLS
Our controls today, yours after that change
Deployed into your account, the platform sits behind the network boundaries, key management, logging and identity controls you already operate and have had audited.
TENANCY
Tenant-aware data model
Enforced end-to-end isolation, tenant administration and cross-tenant negative testing are target.
ACCESS CONTROL
Token-based login with session handling
Enforced authorization boundaries, role-based administration and single sign-on are target.
REVIEWER WORKFLOW
Reviewers inspect and correct values today
Queues, assignment, escalation and service levels are on the roadmap.
OBSERVABILITY
Application-level run metrics
Infrastructure observability, alerting and security detection are target.
RESILIENCE
Standard managed database and storage services
Defined availability and recovery objectives with tested backup and restore are target.

Intake

How a loan file reaches the platform, today and next

Split into what runs now and what is on the roadmap, because the difference decides what an evaluation can actually cover.

TODAY
Upload sessions: deliver the file, nothing to build
Loan packages are delivered into tracked processing sessions and run end to end. No integration work on your side and no change to how your team assembles files. This is how an evaluation starts.
NEXT
Batch ingestion and origination connectors
Connectors to origination and document management systems, and batch ingestion for volume, so files flow without being handed over. On the roadmap, not available today.
NEXT
Mortgage data interfaces, including ULDD
Consuming a Uniform Loan Delivery Dataset export as the structured companion to the document set, so a cross-document check gains an authoritative view of what the loan is supposed to say. A roadmap item, not a current capability, and the foundation of the planned Delivery application.

Security and data

Your documents are not training data

A third-party foundation model reads them, inside our cloud environment. Retention is contractual. Deployment inside your own AWS account is in development and not yet available.


Evaluation

How to test it against controlled files

A synthetic file with no data from you, or three to five of your own de-identified files whose outcomes you know. Misses, false alerts and time, agreed in advance and computed together.

See where it stands, on a real file.

A working session on a synthetic file, or de-identified files of your own. You will see the parts described above as running, and the parts that are not.