Chapter 15: Automated Underwriting: DU, Loan Product Advisor, and Reading the Findings

"The machine did not approve your borrower. It read what you typed, and then it told you what it would need in order to believe you." — constructed; what an underwriting manager says to a new loan officer, roughly once

Overview

It is day 6 on the Linden Street file. The application is complete, credit is pulled, the income worksheet is built, the asset statements are in a folder, and you have a structure: conventional, thirty-year fixed, five percent down, \$365,750 at 6.625%. You press submit. Ninety seconds later a nine-page document comes back, and the first line of it says Approve/Eligible.

Here is what most new loan officers do next. They read those two words, forward the file to processing, call the agent, and never open pages two through nine.

Here is what the rest of the document contains: every number the system used to reach that answer, so you can check whether any of them is wrong — and a numbered list of the specific documents that must be produced to make the answer stick. That list is the whole processing plan for the next seventeen days. It was written for you, in ninety seconds, for free, by the system that will decide whether this loan is saleable. And most people in this business skim past it to look at two words they already expected to see.

Learning to read a findings report is the fastest available upgrade to a new loan officer's competence, and most never do it. That sentence is the reason this chapter exists.

The other half of the chapter is the thing nobody warns you about. An automated underwriting system is not evaluating your borrower. It is evaluating the data you entered about your borrower. Those are different objects, and when they diverge — a mistyped income, an occupancy checked wrong, a debt left off, a condominium coded as a detached house — the system returns a confident, well-formatted, completely worthless answer, and it does not tell you it is worthless. You told it your borrower earns \$12,300 a month. It has no way to know they earn \$10,500. It answers the question you asked.

So this chapter teaches three things: what these systems actually do, how to read what they give you back, and how to own the data underneath so the answer means something.

In this chapter, you will learn to:

• Explain what an automated underwriting system evaluates and what it structurally cannot see
• Distinguish DU, Loan Product Advisor, and the TOTAL Scorecard, and say which loans each serves
• Read a findings report page by page and reconcile every figure against your own worksheet
• Separate the borrower half of a recommendation from the loan half, and diagnose Approve/Ineligible
• Turn verification messages into a document request the same day
• Identify which inputs move a recommendation, and restructure rather than re-run hopefully
• Audit an application for the data errors that make a findings report worthless
• State plainly what the AUS does not decide, and who does

Learning Paths

🎓 Exam — §15.5 is the section that shows up on the test. Know that the recommendation has two halves, know which half each word belongs to, and know that Approve/Ineligible is a real outcome about a loan, not a mixed message about a borrower. §15.3's Accept-versus-Refer distinction on FHA is also testable. 🏠 New LO — §15.4, §15.6, and §15.9, in that order, and then read them again next week with a live findings report in front of you. These three sections are a job skill, not background. 🤝 Partner — §15.5 and §15.10. If you understand that an AUS recommendation is not an approval and that "the computer said no" is never the end of a file, you will stop losing transactions that were still alive. 📊 Operations — §15.6 and §15.8. The verification-message block is the source document for your document-request template, and §15.8 explains why a re-run queue full of identical submissions is a training problem rather than a systems problem.

15.1 What an AUS is actually doing

An automated underwriting system (AUS) is a piece of software that takes a complete loan application plus a credit report, evaluates them against a published rulebook and an unpublished risk model, and returns three things: a recommendation, an assessment of whether the loan itself is allowed, and a list of the documents required to support both.

That is the whole machine. Everything else in this chapter is detail on those three outputs.

It helps to be precise about what kind of thing is happening inside, because the two halves of it behave completely differently.

The half that is rules

A large part of an AUS is deterministic rule-checking. Is the loan amount within the applicable limit for this county and this number of units? Is the loan-to-value ratio within the maximum for this product, this occupancy, and this property type? Is the term one the program permits? Is the property type eligible? Is the debt-to-income ratio within the published maximum? Is there mortgage insurance where mortgage insurance is required?

None of that is mysterious. Those rules are published — in the Fannie Mae Selling Guide, the Freddie Mac Seller/Servicer Guide, and HUD Handbook 4000.1 — and you can look up every one of them. The AUS is, in this half, an extremely fast and extremely literal guideline reader. It does not get tired at four o'clock on a Friday and it does not forget that this county has a different limit than the next county over.

This is worth internalizing early: a substantial part of what the AUS tells you, you could have worked out yourself from a free public document. The loan officers who are good at this job usually know what the answer is going to be before they press submit, and they treat the findings as a confirmation rather than a revelation. When the answer surprises them, they know something is wrong — and it is usually the data.

The half that is a model

The other half is a statistical risk assessment. It takes the application's characteristics — credit history and score, the ratios, reserves, the loan-to-value, the product, occupancy, the number of borrowers, employment stability, and a set of other factors — and produces a judgment about the likelihood that this loan performs.

The internals of that model are not public. Fannie Mae and Freddie Mac do not publish the weights, the interaction terms, the thresholds, or the exact list of variables. Neither does HUD for its scorecard. You can read guidance about which factors matter and in roughly which direction, and you can develop good instincts from a few hundred files, but anyone who tells you they know exactly how much a 20-point credit score improvement is worth against half a point of DTI in DU is guessing. They may be guessing well. They are still guessing.

This is not a scandal; the models are proprietary and disclosing them would invite gaming. But it has a practical consequence you must absorb: you cannot reverse-engineer your way to a recommendation. What you can do — reliably, every time — is make the data correct and make the structure eligible, which is the part that is knowable. Chapter 14 gave you the rulebook. This chapter gives you the interface to it.

What it is not doing

An AUS does not read documents. It has never seen a paystub. It does not know whether the \$10,500 you typed is supported by anything at all. It does not evaluate whether a borrower's story hangs together, whether the two-year employment history you entered is real, whether the gift is actually a loan, or whether the appraisal will support the price. It does not know your lender's rules. It does not know whether this household can live with a \$3,033.72 payment.

It answers one question — given these facts, does this loan fit? — and then tells you what it will require in order to be convinced those facts are facts.

WHAT GOES IN, WHAT COMES OUT                             [constructed teaching example]

   YOU SUPPLY                          THE SYSTEM SUPPLIES
   ─────────────────────────────       ────────────────────────────────────────
   the application data (1003)    →    a RISK assessment of the borrower
   the credit report              →    an ELIGIBILITY assessment of the loan
   income and asset figures       →    VERIFICATION MESSAGES: the document list
   the structure (product, LTV)   →    observations, offers, and lender notes
   property and occupancy         →
   ─────────────────────────────       ────────────────────────────────────────
        ↑                                          ↑
   EVERY ONE OF THESE IS               NONE OF THESE IS A DECISION.
   TYPED BY A HUMAN.                   A LENDER DECIDES. (§15.10)

Look at the left column and notice who is responsible for it. Not the processor, not the underwriter, not the system. You. The loan officer takes the application, calculates the income, enters the debts, and codes the property. The AUS is downstream of every one of those acts. Section 15.9 is about what happens when one of them is wrong, and it is the section that costs people files.

Two final structural facts, both of which matter later.

An AUS run is a point-in-time evaluation. It reflects the credit report that existed when you pulled it and the facts as of the moment you submitted. Credit reports go stale. Borrowers open furniture accounts. The findings from day 6 describe day 6.

An AUS run is deterministic. Submit identical data on the same system version and you get an identical answer. This is the single most useful thing to know about re-running (§15.8), and it is the reason a loan officer who re-submits an unchanged file three times has done three times nothing.

📞 On the Phone

Day 6, 5:10 p.m. The buyer's agent has seen the file move in the portal.

Agent: "So they're approved?"

The wrong answer: "Yep, DU approved them." You have just told a real estate agent something that will be repeated to a listing agent as approved, and a seller may make decisions on it. Nothing has been verified. Nothing has been underwritten.

The other wrong answer: "Nothing's approved until underwriting, I can't tell you anything." True, useless, and your agent now has nothing to say to her clients or the other side.

What actually works: "We have an Approve/Eligible, which is the answer we wanted and is not the same thing as approved. What it means is that the automated system read the file we submitted and said: if you document these specific fourteen items, this loan fits Fannie Mae's guidelines. I sent the borrowers that exact list twenty minutes ago. Our underwriter approves the loan once the documents are in and they support what we said — I'd expect that around day 23. What you can tell the listing agent today is that the file is inside the guidelines on income, credit, and assets, and that I have a written list of what's left and dates on all of it."

Notice the shape. You name what you actually have, you name what you do not have yet, and you convert both into something your partner can say to a third party without it becoming a lie. The difference between "approved" and "Approve/Eligible" is not pedantry — it is the difference between a seller who was told the truth and one who wasn't.

15.2 Desktop Underwriter and Loan Product Advisor

There are two automated underwriting systems for conventional conforming lending in the United States, and there is one for each of the two entities that will buy the loan.

Desktop Underwriter (DU) is Fannie Mae's. Loan Product Advisor (LPA) is Freddie Mac's; originators who have been in the business a while will still call it Loan Prospector, its former name. Both were introduced in the mid-1990s (Chapter 15's first case study is about what happened next), both are accessed through your loan origination system or through the agency's own interface, and both are used many millions of times a year.

They do the same job for two different rulebooks, and this is the first thing to get straight: each system evaluates the loan against its own owner's guidelines. DU tells you whether Fannie Mae will buy the loan. LPA tells you whether Freddie Mac will. An Approve from DU is not a statement about Freddie Mac's requirements, and running LPA does not produce a Fannie Mae answer. Which one you run depends on where your lender intends to sell the loan — and if your lender delivers to both, you may have a real choice.

The vocabulary is different, and the exam knows it

The two systems use different words for the same structural idea, which is the source of a great deal of confusion in classrooms and in exam rooms.

Read the "borrower/risk result" row across. DU says Approve; LPA says Accept; they mean approximately the same thing. DU says Refer; LPA says Caution; those also mean approximately the same thing — the model will not endorse this risk on its own, and a human must take it from here. The word changes; the structure does not.

They are not the same system, and their disagreements are not explainable

Run the same file through both and you will usually get the same answer. Sometimes you will not. A file that Refers in DU may Accept in LPA, and the reverse happens too.

Why? Because the two guides differ in places — treatment of certain income types, of student loan payments, of self-employment history, of reserves requirements, of specific property and project categories — and because the two risk models are separately developed and separately calibrated. Some of the differences are documented in the guides and you can learn them. Some are model behavior and you cannot know them from outside. Anybody who explains to you with confidence why a particular file Accepted in LPA and Referred in DU is telling you a story.

The practitioner consequence is straightforward and legitimate: when a conventional file comes back Refer and your lender delivers to both agencies, run the other system. This is not gaming anything. Both rulebooks are published, both agencies buy loans, and a lender that is approved to sell to Freddie Mac is entitled to underwrite to Freddie Mac's guide. What makes it legitimate is that the data does not change — you submit the same facts to a second published rulebook. What would make it illegitimate is changing the facts until something approves, and §15.8 draws that line hard.

Two constraints on this, both real. First, your employer may not deliver to both, or may have a policy about which system runs first; ask before you assume. Second, your lender's overlays (Chapter 14 §14.7) sit on top of whichever answer you get, and an overlay does not care which agency approved the file.

What both systems also give you

Beyond the recommendation, both systems have layered on a set of services that a loan officer should know exist because they change the calendar and sometimes the cost of a file:

• Verification and representation-and-warranty relief programs. Both agencies offer processes under which income, employment, or asset data validated through approved third-party data providers earns the lender relief from certain representations and warranties. Chapter 14 owns reps and warrants; the practical effect here is that the findings may say a particular item is validated and may reduce what you have to collect. This is a genuine, documented benefit and it is worth knowing whether your shop uses it.
• Collateral relief. The offer to accept the stated value without a traditional appraisal. Section 15.7 is entirely about this.
• A durable record. Every submission is stored with a case identifier and a submission number, and the whole submission history travels with the loan file. This matters more than new originators expect (§15.8).

15.3 The TOTAL Scorecard and how government loans differ

Now the government side, where the architecture is genuinely different and candidates get it wrong.

TOTAL — Technology Open To Approved Lenders — is FHA's scorecard. It is not a standalone automated underwriting system, and this is the point. TOTAL is run through an approved AUS. In practice, an FHA case is submitted through DU or through Loan Product Advisor (or another approved system), that system passes the data to the scorecard, and the scorecard returns a risk classification which appears inside the findings the AUS produces.

HOW AN FHA CASE IS EVALUATED                             [constructed teaching example]

   application data ──┐
                      ↓
              ┌───────────────┐   passes data    ┌───────────────────────┐
              │  DU or LPA    │ ───────────────→ │   FHA TOTAL SCORECARD │
              │  (the AUS)    │ ←─────────────── │   returns ACCEPT or   │
              └───────┬───────┘   returns the    │   REFER                │
                      │           risk class     └───────────────────────┘
                      │
                      ↓
        the AUS pairs the scorecard's result with its own
        eligibility check against FHA program parameters
                      │
                      ↓
        FINDINGS: risk classification + eligibility + verification messages

   A REFER is not a denial. It is a routing instruction: this file must be
   underwritten by a human, against HUD Handbook 4000.1.

Read the diagram bottom-up. An Accept from the scorecard means the file may be underwritten under the documentation and processing the scorecard supports. A Refer means the file goes to manual underwriting (Chapter 14) under HUD Handbook 4000.1, where an underwriter evaluates it against the manual benchmarks and documented compensating factors.

That last sentence is the one to keep. A Refer is a routing decision, not a rejection. A new loan officer who sees a Refer and calls the borrower to say "FHA turned you down" has just terminated a loan that was never declined by anyone. Files are manually underwritten every day and they close.

Two more government notes, briefly, because Chapters 16 and 17 own them properly. VA loans are also submitted through DU or LPA, and VA's own residual income test — money left over after all obligations, by household size and region — is a requirement the AUS does not replace; Chapter 17 takes it apart. USDA guaranteed loans run through USDA's own system, the Guaranteed Underwriting System, which returns its own accept/refer classifications and enforces the program's income and property-location eligibility.

Why the scorecard matters: the Harlow Street file

This is the cleanest illustration in the book of what an automated scorecard does that a manual benchmark cannot.

🧮 Run the Numbers

The Harlow Street file — a loan the benchmark says no to and the scorecard says yes to. (Constructed teaching file. See Chapters 5, 8, 16, 18, 25, and 33.)

A single borrower. FHA financing. 641 representative score. Gross monthly income \$4,150.00**. Existing monthly debts **\$395.00. A \$215,000** townhome, with a **\$10,000 forgivable county down-payment assistance second.

The proposed housing payment — principal, interest, taxes, insurance, and the annual mortgage insurance premium, built in Chapter 16 — is \$1,721.57. Run the two ratios:

	Calculation	Result
Front-end (housing)	\$1,721.57 ÷ \$4,150.00	41.48%
Back-end (total debt)	(\$1,721.57 + \$395.00) ÷ \$4,150.00 = \$2,116.57 ÷ \$4,150.00	51.00%

Now apply FHA's manual underwriting benchmark of 31% / 43%:

	Calculation	Maximum
Maximum housing payment	\$4,150.00 × 0.31 | **\$1,286.50**
Maximum total obligations	\$4,150.00 × 0.43 | **\$1,784.50**

The proposed housing payment exceeds the benchmark by \$1,721.57 − \$1,286.50 = \$435.07 a month. Total obligations exceed it by \$2,116.57 − \$1,784.50 = \$332.07 a month.

Interpretation. Under a pure 31/43 manual benchmark, this borrower's payment would have to be roughly a quarter smaller — which in most markets is a different house in a different neighborhood, or no house. With an Accept from the TOTAL Scorecard, this loan is approvable as submitted. The scorecard is looking at the whole picture — the credit history behind the 641, the modest existing debt, the assistance structure, employment — and reaching a different conclusion than a two-number rule can reach.

HUD does publish higher permitted ratios for manually underwritten files where specific compensating factors are documented; the tiers and the qualifying factors are in Handbook 4000.1 and you must read the current version rather than a summary. But the structural point stands: the benchmark is a blunt instrument and the scorecard is not. That is what automated underwriting was built to do.

(Illustrative figures for a constructed file. FHA benchmarks and permitted exceptions are set by HUD and change; verify against the current Handbook 4000.1.)

And here is the uncomfortable half of the same story, which an honest book has to say out loud. The reason this borrower gets a house is a model whose reasoning is not published. We can describe what it considers. We cannot show the borrower the arithmetic, because there is no arithmetic to show them — and Chapter 25's fair-lending material, along with this chapter's second case study, is about what that means and what the law requires anyway.

15.4 Reading a findings report, page by page

Here is the section that will make you better at this job by Friday.

A findings report is not a verdict with some administrative pages stapled behind it. It is a structured document with six recognizable parts, and the two words on page one are the least useful thing in it, because they are the one part you already anticipated.

ANATOMY OF A FINDINGS REPORT — what lives where
  ┌──────────────────────────────────────────────────────────────────────────┐
  │ 1. SUMMARY              the recommendation, the submission number and    │
  │                         date/time, the case identifier, the system       │
  │                         version that produced it                         │
  ├──────────────────────────────────────────────────────────────────────────┤
  │ 2. LOAN DATA            product, term, purpose, occupancy, property type  │
  │    (WHAT YOU TYPED)     and units, address, sales price, value, loan      │
  │                         amount, LTV / CLTV / HCLTV, note rate, MI         │
  ├──────────────────────────────────────────────────────────────────────────┤
  │ 3. UNDERWRITING         representative score, qualifying income, housing  │
  │    ANALYSIS             expense, total obligations, both ratios, verified │
  │                         assets, funds to close, reserves                  │
  ├──────────────────────────────────────────────────────────────────────────┤
  │ 4. RISK / ELIGIBILITY   the two halves stated separately, plus the reason │
  │                         for any ineligibility                             │
  ├──────────────────────────────────────────────────────────────────────────┤
  │ 5. VERIFICATION         THE DOCUMENT LIST. Income, assets, credit and     │
  │    MESSAGES             liabilities, property, closing. One requirement   │
  │                         per message. (§15.6)                              │
  ├──────────────────────────────────────────────────────────────────────────┤
  │ 6. OBSERVATIONS AND     advisory notes, collateral offers, validation and │
  │    LENDER NOTES         relief messages, special feature codes the lender │
  │                         must deliver with the loan                        │
  └──────────────────────────────────────────────────────────────────────────┘

  THE READING ORDER THAT WORKS:
    read 2 and 3 FIRST  — they are the only pages you control
    read 5 SECOND       — it is the only page that creates work
    read 4 THIRD        — and if anything is Ineligible, read WHY, not WHETHER
    read 1 LAST         — it is a consequence, not a cause

That reading order is the practical heart of this chapter, so let me defend it.

Pages 2 and 3 are an audit of your own work. They print back every fact the system used. If the loan amount on page 2 is not the loan amount you structured, the recommendation on page 1 is about a different loan. If the qualifying income on page 3 is not the income on your worksheet, one of the two is wrong and you need to know which one today rather than on day 23 when an underwriter recalculates it from W-2s. This is the closest thing this business has to a free second set of eyes, and it takes ninety seconds.

Build the habit as a reconciliation, not a read. Put your income worksheet next to page 3 and check five figures against it: qualifying income, total housing expense, total monthly obligations, verified assets, and reserves. If all five match your own numbers, your data entry is almost certainly clean. If one does not, stop and find out why before you do anything else.

Page 5 is the only page that creates work, and the sooner it becomes a dated request the sooner the file moves. Section 15.6 is about that page.

Page 4 you read for the reason, not the result. "Ineligible" without the reason is useless; "Ineligible — LTV exceeds the maximum for this product" is a ten-minute fix (§15.5).

Page 1 you read last because it is downstream of everything else. If pages 2 and 3 are wrong, page 1 is wrong, and reading it first only tells you what to feel about a number you have not checked yet.

Now let us look at one.

AUTOMATED UNDERWRITING FINDINGS — SUMMARY PAGE            [the Linden Street file]
Constructed rendering. Structure only; no agency's exact wording, message
numbering, or page layout is reproduced. Formats differ by system and change.

  ┌────────────────────────────────────────────────────────────────────────┐
  │  RECOMMENDATION       Approve / Eligible                               │
  │  Submitted            day 6, 4:52 p.m.        Submission number   1    │
  │  Case identifier      (system-assigned)       Engine version   current │
  └────────────────────────────────────────────────────────────────────────┘

  MORTGAGE INFORMATION                    ← page 2: WHAT YOU TYPED
    Loan type                  Conventional, first lien
    Amortization               Fixed rate, 360 months, fully amortizing
    Loan purpose               Purchase
    Occupancy                  Primary residence
    Property type              1 unit, detached, site-built
    Subject property           4412 Linden Street, Ridgeview
    Sales price                $385,000.00
    Appraised value            not yet obtained — value not entered
    Loan amount                $365,750.00
    LTV / CLTV / HCLTV         95.00% / 95.00% / 95.00%
    Note rate / term           6.625% / 30 years
    Mortgage insurance         required; standard coverage for this LTV,
                               term, and product; borrower-paid monthly

  UNDERWRITING ANALYSIS                   ← page 3: WHAT IT CONCLUDED
    Number of borrowers                     2
    Representative credit score             706
    Total qualifying income (monthly)       $10,500.00
    Proposed monthly housing expense        $3,033.72
    Total monthly obligations               $4,479.72
    Housing expense ratio                   28.89%
    Total debt-to-income ratio              42.66%
    Total verified assets                   $38,000.00
    Funds required to close                 $25,376.34
    Reserves after closing                  $12,623.66   (4.16 months)

  RISK AND ELIGIBILITY                    ← page 4: THE TWO HALVES
    Risk assessment            APPROVE
    Eligibility assessment     ELIGIBLE — the loan as submitted meets the
                               product and program parameters

Every figure on that page comes from somewhere you have already been. \$365,750 is 95% of \$385,000 (Chapter 13). \$10,500.00 is Borrower 1's \$6,300.00 plus Borrower 2's \$4,200.00 (Chapter 11). The 706 is the lower of the two middle scores (Chapter 10). \$3,033.72 is P&I plus taxes plus insurance plus mortgage insurance (Chapter 4). \$4,479.72 is that payment plus \$1,446.00 of other monthly debts (Chapter 12's asset work sits behind the last three lines).

That is the point of the reconciliation. Not one number on this page should be a surprise. If one is, you have found something.

📄 Read the File

text FIGURE 15.1 — "The summary page, read backwards" [the Linden Street file] THE DOCUMENT Automated underwriting findings report, summary and analysis pages, generated day 6 at 4:52 p.m. from the loan origination system. First submission on this casefile. Constructed rendering. THE CONTEXT A $385,000 purchase, conventional, 5% down, two borrowers, both first-time buyers, 45-day contract. The application was completed yesterday. No appraisal has been ordered. Nothing has been verified with a third party yet — every figure on this page is a number a loan officer typed and a credit report supplied. WHAT IT SHOWS Approve/Eligible on the first submission. The loan is inside the LTV maximum at 95.00%; the ratios are 28.89% housing and 42.66% total; the representative score is 706; reserves after closing are 4.16 months of the full PITI-plus-MI payment. The five figures a loan officer should reconcile all tie to the worksheet exactly: income $10,500.00, housing $3,033.72, obligations $4,479.72, assets $38,000.00, reserves $12,623.66. The eligibility half is clean because a 95% conventional purchase on a 1-unit detached primary residence is squarely inside the product. WHAT IT DOESN'T It does not prove a single fact on it. The $10,500.00 is a calculation from documents the underwriter has not seen; the $38,000.00 includes a $10,000 gift that has not been given yet. It does not know the property is worth $385,000 — no value has been entered, so LTV is computed on the sales price. It says nothing about the lender's overlays. It cannot see the $611.00 furniture payment the borrowers will take on in thirty-five days, because on day 6 that account does not exist. And it is not an approval: no underwriter has looked at this file. THE DECISION Today: reconcile the five figures (done — they tie), then go straight to the verification messages and convert them into a dated document request to the borrowers with a copy to the processor. Order the appraisal and title tomorrow. Tell the agent what this is and, precisely, what it is not. THE LESSON The recommendation is the least informative thing on the report. It is a function of pages 2 and 3, and pages 2 and 3 are a function of what a human typed. Read the inputs, then the work list, then the answer — in that order, every time.

Constructed rendering of a real document type. The figures are this book's frozen Linden Street facts; the layout and wording are written for teaching and are not any system's actual output.

15.5 Approve/Eligible, Approve/Ineligible, Refer, Caution

A recommendation has two halves, and they are about two different things.

The first half is about the borrower. Approve (DU) or Accept (LPA) or Accept (TOTAL) means the risk assessment endorses this borrower on this loan. Refer (DU, TOTAL) or Caution (LPA) means it does not — the model will not stand behind the file on its own, and a human underwriter must decide.

The second half is about the loan. Eligible means the loan as submitted meets the product and program parameters. Ineligible means it breaks one.

They are evaluated separately and they are reported separately, which means all four combinations are possible and each one means something different.

THE TWO HALVES OF A RECOMMENDATION                       [constructed teaching example]

                        │  THE LOAN: does it fit the product and program?
   THE BORROWER:        │
   will the model       │      ELIGIBLE              │      INELIGIBLE
   endorse the risk?    │                            │
  ──────────────────────┼────────────────────────────┼──────────────────────────
   APPROVE   (DU)       │  The ordinary good         │  Creditworthy borrower,
   ACCEPT    (LPA)      │  outcome. Document it      │  broken parameter.
   ACCEPT    (TOTAL)    │  and move.                 │  FIX THE STRUCTURE —
                        │                            │  this is not a decline.
  ──────────────────────┼────────────────────────────┼──────────────────────────
   REFER     (DU)       │  Parameters are fine; the  │  Both halves need work.
   CAUTION   (LPA)      │  model will not endorse    │  Fix the structure first,
   REFER     (TOTAL)    │  the risk. Manual          │  then re-evaluate the
                        │  underwrite, restructure,  │  risk half — the fix may
                        │  or run the other system.  │  move both.
  ──────────────────────┴────────────────────────────┴──────────────────────────

   LEFT COLUMN asks WHO is borrowing.   TOP ROW asks WHAT the loan is.
   Neither half is a decision. Both are inputs to one. (§15.10)

Now take them one at a time.

Approve/Eligible

The outcome you wanted, and the outcome the Linden Street file got on day 6. The risk model endorses the borrower; the loan fits the product. Proceed to documentation.

What it is not: an approval. It is a recommendation that the loan, as described, is one the agency will buy if the verification messages are satisfied and the lender's own underwriter agrees and the lender's overlays permit it. Chapter 14 §14.7 gave you overlays; here is where they bite. An Approve/Eligible at a 620 representative score is worth exactly nothing at a lender whose overlay floor is 640, and the findings will never mention that, because the findings do not know your employer exists.

Approve/Ineligible

The most misread outcome in origination, and the one this section exists for.

Read the two halves. Approve — the model endorses the borrower. Ineligible — the loan breaks a parameter. Put together: a creditworthy borrower on a loan that is structured wrong. That is almost always a fixable problem, and the fix is to the structure, not to the borrower.

The findings will tell you exactly which parameter broke. The usual suspects:

• Loan amount above the applicable limit. Conforming loan limits are set annually by the FHFA and vary by county and by number of units. A loan one dollar over is Ineligible.
• LTV, CLTV, or HCLTV above the product maximum. This includes the case where a second lien or a down-payment-assistance program pushes the combined ratio past the cap even though the first mortgage is fine.
• An ineligible property type or project. A condominium in a project that fails review; a manufactured home where the product does not allow one; too many units.
• An ineligible occupancy for the product or LTV. Second homes and investment properties carry their own maximums, and a program restricted to primary residences does not care how good the borrower is.
• A term or feature the product does not permit, or a program-specific limit — an income cap on an affordable-lending product, a first-time-homebuyer requirement not met.
• A debt-to-income ratio above the program's published maximum, depending on the system and how it classifies that parameter.

Here is the shape of the fix, using our own file. Suppose the Linden Street borrowers had come in with only 2% down instead of 5%:

At 98.00% loan-to-value the loan breaks the maximum LTV a standard conventional purchase permits (currently 97% where the program's requirements are met — verify the current figure and the program's conditions in the Selling Guide). The borrowers are unchanged. The credit, income, ratios, and reserves are all the same. The loan is wrong, not the people.

And the fix is arithmetic. Three percent of \$385,000 is \$11,550.00, so getting to a 97% loan requires \$11,550.00 down instead of \$7,700.00 — another \$3,850.00. Alternatively, restructure onto a program with a higher LTV ceiling: FHA permits 96.5% (Chapter 16), which would also clear it.

That is a ten-minute conversation with a borrower, not a decline. And a loan officer who reads "Ineligible" and hears "no" has just given away a file that a competitor will close.

🎓 NMLS Exam Watch

This is a favorite, and the trap is built into the wording.

The stem describes a borrower with strong credit, low ratios, and good reserves, whose loan comes back Approve/Ineligible, and asks what it means. The wrong answers are all versions of "the borrower was declined," "the borrower's credit is insufficient," or "the file must be manually underwritten."

The right frame: the first word is about the borrower; the second word is about the loan. Approve = the risk assessment endorses the borrower. Ineligible = the loan as submitted does not meet a product or program parameter. The correct action is to identify the parameter that was broken and restructure — a smaller loan, more down payment, a different program, a different property classification — not to give the borrower bad news.

Two companions the exam also likes:

• A Refer (DU) or a Caution (LPA) is not a denial. It routes the file to manual underwriting or to restructuring. Only a lender denies a loan.
• DU says Approve/Refer; LPA says Accept/Caution. If a question uses "Caution," it is asking about Freddie Mac's system. If it uses "Refer with Caution," it is using older DU vocabulary — read the stem for which system it names rather than matching on a single word.

Refer, and Caution

Refer in DU, Caution in LPA, Refer from the TOTAL Scorecard. The risk assessment declines to endorse the file on its own.

What it means practically: the automated path is closed, and the remaining paths are all open.

1. Manual underwriting. A human underwriter evaluates the file against the guide's manual requirements and documented compensating factors. Chapter 14 covered this. It is slower, the documentation is heavier, the ratio limits are usually tighter, and it works.
2. Restructure and re-run. If the binding constraint is something you can actually change — more down payment, a paid-off installment debt, a co-borrower's income you had not counted, a different product — change the fact, then re-run (§15.8).
3. Run the other agency's system, if your lender delivers to both. Same data, second rulebook.
4. Look for a data error. A meaningful share of surprising Refers are caused by something typed wrong, and §15.9 is about finding it before you go tell a borrower bad news.
5. A different program entirely. FHA, VA, USDA, a state housing agency product, or a portfolio or non-QM option (Chapter 34) may fit a file conventional financing will not.

The order matters. Look for the data error first, because it is free and because it is embarrassing to manually underwrite a file that Referred because you coded a duplex as a fourplex.

The results that are not results

Two more outputs exist and are not decisions at all.

Out of Scope, or an equivalent "unable to assess" message, means the system cannot evaluate this loan — a product, feature, or borrower situation outside what it handles. It is a statement about the software, not about the file, and the loan may be perfectly good under manual underwriting or a different channel.

An error or incomplete submission means the data is insufficient or internally contradictory and the system stopped. Fix the data and resubmit; nothing has been evaluated.

15.6 Verification messages: the document list the machine just wrote for you

This is the section that pays for the chapter.

An automated underwriting system does not merely approve or decline. It writes the document list. Buried on page five of a report most loan officers never open is a numbered set of verification messages — the specific items that must be provided in order to make the recommendation stand. Satisfy them and the loan is documented the way the investor requires. Miss one and the approval you were relying on has a hole in it.

And here is why it is such an efficient thing to learn: the messages are generated from the data you entered. You typed variable income, so the system asked for a written verification of employment. You typed a gift, so the system asked for a gift letter and evidence of transfer. You typed a rental housing history, so the system asked you to verify it. The verification block is a mirror of your own application, rendered as work.

AUTOMATED UNDERWRITING FINDINGS — VERIFICATION MESSAGES     [the Linden Street file]
Constructed rendering. Real systems use their own message text, numbering, and
grouping; the wording below is written for this book. The structure and the
intent are what transfer. Day 6 submission, Approve/Eligible.

  EMPLOYMENT AND INCOME
   1  Borrower 1, base income of $5,720.00 per month: obtain the most recent
      paystub showing year-to-date earnings, and W-2 forms for the most recent
      two years.
   2  Borrower 1, variable income of $580.00 per month (shift differential and
      overtime): obtain a written verification of employment documenting these
      earnings for the most recent two years and year-to-date.
   3  Borrower 2, base income of $2,400.00 per month: obtain the most recent
      paystub showing year-to-date earnings, and W-2 forms for the most recent
      two years.
   4  Borrower 2, variable income of $1,800.00 per month (commission): obtain a
      written verification of employment documenting commission earnings for the
      most recent two years and year-to-date.
   5  Verify the employment of each borrower within the period before the note
      date required by the applicable guide.

  ASSETS AND FUNDS TO CLOSE
   6  Depository accounts, two, reported total $28,000.00: obtain account
      statements covering the most recent two months for each account.
   7  Gift funds of $10,000.00: obtain a gift letter signed by the donor stating
      the amount, the donor's relationship to the borrower, and that no
      repayment is expected; document the donor's ability to make the gift and
      the transfer of the funds to the borrower or to closing.
   8  Funds required to close of $25,376.34 must be verified. Total verified
      assets of $38,000.00 and reserves of $12,623.66 (4.16 months) were used in
      this assessment.

  CREDIT AND LIABILITIES
   9  The liabilities below are included in the total debt-to-income ratio at
      the payments shown. Verify each; document any omitted debt and any payment
      that differs from the amount shown.
        installment (auto)         $487.00    31 payments remaining
        installment (auto)         $429.00    19 payments remaining
        installment (student)      $318.00
        revolving, minimum payments $212.00   on balances of $8,400.00
        ─────────────────────────────────────────────────────────────────
        total monthly obligations other than housing        $1,446.00
  10  Verify the housing payment history reported for the current residence
      (36 months).

  PROPERTY, TITLE, AND CLOSING
  11  An appraisal on the applicable form is required for this transaction.
  12  Obtain a title commitment showing the lender in first lien position.
  13  Obtain evidence of hazard insurance in an amount meeting guide
      requirements, effective at or before closing.
  14  Mortgage insurance is required at the coverage applicable to this LTV,
      term, and product.

Fourteen messages. Now watch what happens to your week.

📄 Read the File

text FIGURE 15.2 — "The list you did not have to write" [the Linden Street file] THE DOCUMENT The verification-message block of the day-6 findings report, fourteen numbered items across four categories. Constructed rendering. Produced automatically, in seconds, at no cost. THE CONTEXT Day 6 of a 51-day file on a 45-day contract. The appraisal has not been ordered, the borrowers have not been asked for a single document since application, and the processor has not opened the file. Every hour between now and the moment these items are requested is an hour of pure calendar loss. WHAT IT SHOWS Exactly what must be produced to support the Approve/Eligible. Note how tightly each message tracks the data: two borrowers with variable income produced two separate written-VOE messages (items 2 and 4) that a W-2-only file would not have generated; the $10,000 gift produced item 7 with its three distinct requirements (letter, donor ability, transfer); the 36-month tenancy on the application produced item 10. Item 9 is the most useful line on the page: it prints the four liabilities the system counted, at the exact payments it counted them at, totaling $1,446.00 — which is a direct check on your own debt schedule. WHAT IT DOESN'T It is a FLOOR, not a ceiling. It does not include your lender's overlay requirements, anything the underwriter will ask for after reading the documents, or the conditions that will appear on the approval (Chapter 19 — the day-28 conditional approval carried eleven of them). It does not tell you that the second paystub you will need arrives on the 15th, or that the donor's bank takes four business days to produce a statement. It does not judge documents: it cannot tell a real VOE from an invented one. And it is silent about anything you did not enter — a debt you omitted generates no message, because as far as this report is concerned it does not exist. THE DECISION Today, before you leave: convert all fourteen items into a borrower-facing document request in plain English, with a due date on each, and send it with a copy to the processor. Items 2 and 4 become VOE requests you send to the two employers tomorrow — they are the slowest items on the list and they must go out first. Item 7's donor evidence is the second-slowest, so ask for it today too, not when the gift is deposited. THE LESSON The system that decides whether this loan is saleable has just told you, item by item, what it needs. There is no more reliable processing plan available anywhere, at any price, and it is on page five of a document most loan officers close after reading two words on page one.

Constructed rendering. Message wording, grouping, and numbering are written for this book.

Three disciplines for the verification block

One: do not add to it, and do not subtract from it. The messages are a floor. Collecting less than the list means the file is not documented; collecting a pile of documents nobody asked for means you have burned goodwill you will need on day 44 when you have to ask for something hard. Ask for what is on the list, plus whatever your lender's overlays require, plus what your own judgment says an underwriter will want. Then stop.

Two: sequence by turn time, not by list order. The messages are grouped by category, not by how long each item takes. A paystub arrives in an hour; a written verification of employment from a hospital's HR department can take a week and a half; a donor's bank statement depends on a third-party household. Send the slow requests first, on the day you get the findings. Every day costs money, and this is the single cheapest place in the file to buy three days.

Three: read item 9 as a check on yourself. The liabilities block prints back the debts the system counted and the payments it counted them at. Compare it to your own debt schedule line by line. If the system counted a \$487.00 auto payment and your schedule says \$387.00, one of you misread the credit report — and if it is you, your DTI is wrong and so is everything you told the borrower.

There is a fourth thing, which is more of a posture. The verification messages are the moment the abstraction becomes work. Up to now the file has been a conversation and a spreadsheet. From here it is fourteen items, each of which belongs to a named human being with a job and a life, and every one of them will take longer than you think. Loan officers who are good at the calendar are usually just loan officers who converted page five into a dated request on day 6 instead of day 11.

🔍 Check Your Understanding

1. The findings ask for "the most recent two months" of statements on the depository accounts, and the file already contains three months. Is that a problem? Is it a reason to remove the third?
2. Borrower 2's commission income generated a separate written-VOE message that Borrower 2's base salary did not. Why?
3. A borrower has a \$240 monthly payment on a credit union loan that does not appear anywhere in the liabilities block. What does the absence of a message about it tell you — and what does it not tell you?

(3 is the one that matters. The absence of a message tells you the system did not see the debt. It tells you nothing about whether the debt exists. If it is on the credit report and the system did not count it, something is coded wrong; if it is not on the credit report at all — a private loan, a family obligation, a new account — then the file is currently built on an understated DTI and the recommendation is answered on facts that are not this file's. See §15.9.)

15.7 Appraisal waivers and value acceptance

Sometimes the findings offer you something else entirely: an offer to accept the value stated in the application without a traditional appraisal.

Fannie Mae now calls this value acceptance; for years it was called an appraisal waiver, and the entire industry still says "waiver" out loud. Freddie Mac's version is automated collateral evaluation (ACE). Both agencies also offer hybrid options in which a full appraisal is replaced by a property data collection performed on site by a trained collector — Fannie's value acceptance + property data and Freddie's ACE+ PDR. The names have changed more than once and will change again; the concept has not.

The concept. The agency has accumulated an enormous amount of data about American residential property — prior appraisals submitted to its systems, public records, and its own models. When it has enough of it about a particular property, and the transaction is one where the risk of accepting the stated value is acceptable, it will tell the lender: we do not require a new appraisal on this one.

Where the offer appears. In the findings, on a per-case basis, in the observations section. It is tied to the specific casefile and the specific data you submitted, and — this is the part people get burned by — it can disappear on a re-run. Change the loan amount, the value, the product, or the transaction and the offer may not survive. Never promise a borrower a waiver before you have one, and never promise it will still be there next week.

What it is worth. Two things, and the second is bigger than the first. It saves the appraisal fee — a few hundred to a thousand dollars-plus depending on market, property, and turn time — and it saves the calendar. On the Linden Street file, the appraisal was ordered on day 7 and came back on day 16. That is nine days, on a 45-day contract, and it is nine days during which the file cannot be submitted to underwriting complete. Removing that from the critical path is worth more to a purchase transaction than the fee is. Chapter 39's pipeline arithmetic makes this concrete.

What it is not. Four things, and a loan officer needs to be able to say all four in a borrower's kitchen.

1. It is not a statement that the house is worth the contract price. It is a statement that the agency will accept the stated value for the purpose of buying this loan. Those are different claims, and only one of them is about your borrower's money.
2. It does not protect the borrower from overpaying. A borrower who pays \$40,000 over market with a value acceptance has still paid \$40,000 over market, and has no independent opinion of value in the file to have told them so.
3. It is not a home inspection. These are entirely different things and borrowers conflate them constantly. An appraisal is an opinion of value for the lender. An inspection is a physical examination of the property for the buyer. A waiver of the first has nothing to do with the second, and a borrower who skips an inspection because "the appraisal was waived" has made a serious mistake. Chapter 20 covers the contract side of that.
4. It does not have to be taken. The lender elects whether to exercise the offer, subject to its own policy, and a borrower who wants an appraisal can have one. Some borrowers should want one.

Why the Linden Street file did not get one. It is a 95% loan-to-value purchase. Historically, eligibility for appraisal waivers on purchase transactions has been meaningfully more restricted than on refinances, with LTV caps well below 95% for most cases, and both agencies have adjusted those criteria repeatedly, including expansions aimed at particular borrower and property categories. A 95% purchase has generally been outside the criteria. So the file got an appraisal requirement (verification message 11), the appraisal was ordered on day 7, and it returned on day 16 supporting \$385,000.

Contrast it with the Cypress Court file, where the appraisal came back \$35,000 under contract eleven days before closing and rewrote the entire transaction. Value acceptance is precisely the mechanism that would have prevented anyone from finding that out — which is a fair description of both the benefit and the risk in one sentence.

⚖️ Compliance Check

Three things about waivers that are compliance issues, not sales points.

First, the eligibility criteria change — repeatedly. Which transactions qualify, at what LTV, for which occupancies and property types, and what data the agency requires are all set by the agencies and have been revised many times, including the introduction of the property-data-report hybrids. Nothing in this section should be treated as a current criteria list. Verify the current requirements in the Selling Guide, the Seller/Servicer Guide, and your lender's own policy before you tell any borrower anything. Your lender may also decline to exercise waivers it is offered; that is its right.

Second, the borrower's right to receive valuations does not evaporate. Under ECOA and Regulation B's valuations rule, a creditor must notify an applicant of the right to receive a copy of appraisals and other written valuations developed in connection with a first-lien loan secured by a dwelling, and must provide copies promptly upon completion or a specified number of business days before consummation, whichever is earlier. What that means in a waiver case — where there may be no appraisal but there may still be a written valuation — is a question for your compliance department, and the notice obligation itself is not optional. Chapter 18 covers the valuations rule properly and Chapter 24 covers the disclosure regime.

Third, do not sell it as an opinion of value. "The computer says the house is worth it" is a sentence that is both false and, if a borrower relies on it, potentially a problem for you. What you may accurately say is: "Fannie Mae has told us they don't require an appraisal on this property for this loan. That saves you the fee and about ten days. It is not an opinion that the price is right, and if you want an appraisal for your own information, we can order one."

Requirements change and state law varies. Verify current rules with your compliance department and your regulator.

15.8 Re-running: what changes the answer and what does not

A loan officer with a Refer in front of them will do one of two things, and the two are separated by about a decade of career.

The first re-runs the file. Nothing has changed; they just run it again, on the theory that the computer might be in a better mood. Then they run it a third time. This is not a strategy. It is a superstition, and it wastes the two hours in which the file could have been fixed.

The second reads the findings, identifies the binding constraint, computes what change would relieve it, finds out whether the borrower can actually make that change, changes the fact, documents it, and then re-runs. That loan officer restructures a file in ten minutes.

Re-running changes the answer only if the inputs change. The system is deterministic. Identical data on the same engine version produces an identical result. That is not a quirk to work around; it is the property that makes the system trustworthy, and it means the entire question of "how do I get a better answer" reduces to "which input do I change, and can my borrower actually change it?"

The inputs that move a recommendation

Roughly in order of leverage:

Read the occupancy row twice. Occupancy and property type are not levers. Changing "investment property" to "primary residence" because the first one returned Ineligible is not restructuring; it is occupancy misrepresentation, it is a crime, and Chapter 27 covers what happens to people who do it. The same goes for entering income you cannot document, omitting a debt you know exists, or inventing a co-borrower's employment. A re-run must record a change in the facts. It may never manufacture one.

There is a bright line here and it is easy to state: did something in the world change, or did something in the software change? If the borrower paid off a car, the world changed — re-run. If you retyped their income until it approved, the world did not change and you have created a false record.

And be aware that the record exists. Every submission is stored with a submission number and a timestamp, the full history stays with the casefile, and quality control reviews look at it. A file with six submissions in which the income climbs each time is a document that speaks for itself.

The two things that can change the answer without you touching the data

A system version release. Both agencies periodically deploy new versions of their engines, with published release notes describing what changed. A casefile submitted before a release and re-submitted after it is evaluated by the new version, and the recommendation can move. This is normal, it is announced in advance, and it is a good reason to read your agency's release notes when they come out rather than being surprised by them at nine o'clock on a Monday.

A new credit report. Re-pulling credit, or a credit refresh before closing, brings in whatever has happened since the last pull. On the Linden Street file this is exactly what happened on day 44 — and Chapter 19 tells that story.

🧮 Run the Numbers

Three re-runs on the Linden Street file, and what each is worth. (Runs 2 and 3 are constructed alternatives, not what happened. The day-6 figures are frozen.)

Starting point, day 6: income \$10,500.00, housing \$3,033.72, other debts \$1,446.00, total obligations \$4,479.72, **back-end 42.66%**, reserves \$12,623.66 = 4.16 months.

Run 1 — re-submit with no change.

Data changed	none
New back-end ratio	42.66%
New recommendation	Approve/Eligible — identical
Value of this run	zero

Run 2 — pay off the auto loan with 19 payments remaining (\$429.00/month).

New total obligations: \$4,479.72 − \$429.00 = \$4,050.72 New back-end ratio: \$4,050.72 ÷ \$10,500.00 = 38.58% — an improvement of 4.08 points.

Two catches, and they are why this is a calculation and not a suggestion. First, this debt cannot simply be excluded because it is close to done: an installment debt is generally excludable only when a small number of payments remain — currently ten or fewer under Fannie Mae's guide, and even then not if the payment would materially affect the borrower's ability to pay in the months right after closing. Verify the current rule. Nineteen payments is not close. Second, paying it off costs the payoff balance in cash, which has to come from somewhere, be sourced, and be documented — and it comes out of the reserves that are also doing work in this file.

Run 3 — pay the \$8,400.00 of revolving balances to zero.

New total obligations: \$4,479.72 − \$212.00 = \$4,267.72 New back-end ratio: \$4,267.72 ÷ \$10,500.00 = 40.64% — an improvement of 2.02 points. Cash required: \$8,400.00, from verified funds. New reserves: \$12,623.66 − \$8,400.00 = \$4,223.66** ÷ \$3,033.72 = 1.39 months**, down from 4.16.

Interpretation. Run 2 and Run 3 both improve the ratio and both make the file weaker in another dimension by consuming reserves — and reserves are themselves an input the model uses. On a file that is already Approve/Eligible at 42.66% with 4.16 months of reserves, spending \$8,400.00 to buy 2.02 points of DTI and give up 2.77 months of reserves is a bad trade. On a file that came back Refer at 49%, the same trade may be the whole deal.

This is the discipline. Do not ask "how do I get a better answer." Ask "what is the binding constraint, what would relieve it, what does relieving it cost somewhere else, and can this household actually do it?" Then change the fact, document it, and re-run to record it.

(Illustrative. Exclusion rules for installment debt, reserve requirements, and DTI maximums are set by the agencies and change; verify current guidance.)

The sequence that works

1. Read the findings for the reason. Which half failed, and which parameter or which factor.
2. Reconcile the data first. Before you restructure anything, confirm the report was answered on correct facts (§15.9). A surprising number of "problems" are typos.
3. Compute the change. How much down payment, how much debt payoff, how much income, what product — in dollars, before you call anybody.
4. Test whether it is real. Can the borrower produce the cash? Will the payoff be documentable? Does the income exist and can it be verified (Chapter 11)? If the answer is no, the restructure is not a restructure.
5. Change the fact in the world, then in the system. Pay the debt, then update the data.
6. Re-run, and read the new findings in full — including the verification messages, which will have changed, and the collateral offer, which may have vanished.

15.9 Garbage in: the data-entry errors that cost files

Everything in this chapter rests on a single load-bearing assumption: that the data in the system describes the file on your desk. When it does not, the report is not wrong in some interesting way. It is worthless — a confident, well-formatted, precisely-computed answer to a question about a loan that does not exist.

And the machine cannot tell you. It has no way to know that the \$12,300.00 you typed is not the \$10,500.00 the paystubs will support. It will compute the ratios to two decimal places, return Approve/Eligible, and print the document list — all of it internally consistent, all of it addressed to a fictional borrower.

A findings report is only as good as the data underneath it, and the loan officer owns that data. Not the processor. Not the underwriter. You took the application, you calculated the income, you entered the debts, you coded the property.

⚠️ Where Deals Die

The mistyped income field, and the two directions it fails in.

Borrower 2 on the Linden Street file earns a \$2,400.00 base plus \$1,800.00 of averaged commission, totaling \$4,200.00 a month. Suppose the loan officer enters \$4,200.00 in the base field and then, working down the worksheet, enters the \$1,800.00 commission underneath it.

Income entered: \$6,300.00 + \$4,200.00 + \$1,800.00 = **\$12,300.00. Income supported by the documents: \$10,500.00.**

	Entered (wrong)	Actual
Qualifying income	\$12,300.00 | \$10,500.00
Housing ratio	\$3,033.72 ÷ \$12,300.00 = 24.66%	28.89%
Back-end ratio	\$4,479.72 ÷ \$12,300.00 = 36.42%	42.66%

The findings come back Approve/Eligible at a 36.42% back-end and everybody exhales. The file looks comfortable. The loan officer, feeling generous, tells the borrowers they could probably stretch to a nicer house. On day 23 the underwriter recalculates income from the W-2s and the written verifications and gets \$10,500.00 — and every ratio in the file moves six points.

On this file, the loan still works: 42.66% is inside the guideline. The error was harmless by luck, not by design. Move it onto a file whose true back-end is 51% and the same typo produces an approval on day 6, a document request, an appraisal fee, a ten-day title order, a borrower who has told their landlord they are moving, and a decline on day 25.

Now the other direction, which costs you deals you actually had. Transpose \$10,500.00 into \$10,050.00 — one keystroke:

	Entered (wrong)	Actual
Qualifying income	\$10,050.00 | \$10,500.00
Housing ratio	\$3,033.72 ÷ \$10,050.00 = 30.19%	28.89%
Back-end ratio	\$4,479.72 ÷ \$10,050.00 = 44.57%	42.66%

Nearly two points of DTI, given away for nothing. On a marginal file that is the difference between Approve and Refer — and the loan officer who does not reconcile the report against their own worksheet will believe the machine, tell a borrower they do not qualify, and never find out.

What the disciplined loan officer does instead: reconcile five numbers on page 3 against the worksheet — income, housing expense, total obligations, verified assets, reserves — every single time, before reading the recommendation. It takes ninety seconds. It is the highest-value ninety seconds in the origination process.

The catalog

Beyond income, these are the fields that produce bad recommendations, roughly in order of how often they are wrong and how much damage they do.

Occupancy. Primary residence, second home, and investment property carry different LTV maximums, different reserve requirements, different pricing, and in some cases different products entirely. Getting this wrong is the most consequential single-field error in the application, and getting it wrong deliberately is fraud (Chapter 27).

Property type and units. A condominium coded as a detached single-family home changes project review requirements, mortgage insurance, pricing, and eligibility. A townhome in a planned unit development is not a condominium and is not always a detached home. A two-unit is not a one-unit with a basement apartment. If you do not know what the property is, do not guess — the listing, the tax record, and eventually the appraisal will tell you.

Omitted debts. Every debt you leave out understates the DTI. Most will be caught — the system reconciles what you entered against the credit report and will flag mismatches — but the ones that are not on the credit report are the dangerous ones: a family loan, a new account too recent to report, a court-ordered obligation, a lease. Ask the questions in the application interview and take the answers seriously.

Debts entered at the wrong payment. A revolving account with no minimum reported, an income-driven student loan payment, a lease with a residual — each has a rule for what payment to count (Chapter 12), and the rules are not intuitive. Item 9 of the verification block is your check.

Subject property address and county. Loan limits are county-specific. So are some program eligibility tests and, for USDA, the entire property-eligibility question. A wrong county produces a right answer to a wrong question.

Sales price versus appraised value. LTV runs on the lesser of the two. Before a value exists, the price governs. Entering an estimated value above the contract price does not help anything and will be corrected the moment the appraisal arrives.

Self-employment, ownership percentage, and income type. Coding a commissioned employee as self-employed, or a 25%-owner as a W-2 wage earner, changes the entire documentation set and the income calculation (Chapters 11 and 32).

First-time homebuyer status, citizenship and residency, and marital and vesting details. Each of these drives program eligibility, and each is routinely entered from memory rather than from the application.

Duplicated assets. Entering the same account twice, or entering a gift both as a gift and as a deposit, inflates verified assets and reserves. It gets caught when statements arrive, and it makes the file look sloppy at the exact moment you want the underwriter's goodwill.

The pre-submission audit

Build this into your process. It takes three minutes and it belongs to you, not to the processor.

PRE-SUBMISSION DATA AUDIT — do this BEFORE you press submit
   1  Loan purpose, occupancy, property type and number of units — verified
      against the contract and the listing, not from memory
   2  Subject property address, including county
   3  Sales price; value (or blank if no appraisal yet)
   4  Loan amount, down payment, and any second lien → LTV / CLTV / HCLTV
   5  Product, term, amortization, note rate, and lock assumption
   6  MI type and payment method
   7  Each borrower's income, BY TYPE, against your income worksheet
   8  Every liability on the credit report, at the correct payment, plus any
      debt disclosed by the borrower that is NOT on the credit report
   9  Every asset account, once, at the balance the statement shows
  10  Gift amount, donor relationship, and whether it is received or promised
  11  Housing history — current rent or housing payment, and how long
  12  First-time-homebuyer status, citizenship/residency, and dependents
   ─────────────────────────────────────────────────────────────────────────
   AND AFTER submission, reconcile five figures on the analysis page against
   your own worksheet: income · housing expense · total obligations ·
   verified assets · reserves. If any one disagrees, stop.

The judgment the machine does not make: the Fulton Avenue file

One more error class, and it is not a typo at all.

The Fulton Avenue borrower owns an S-corporation. Two years of returns are filed. The Fannie Mae cash-flow analysis produces \$9,020.83** as a 24-month average and **\$8,916.67 using the most recent year alone, and because the income declined 2.3% year over year, the underwriter uses the lower figure: \$8,916.67. Chapter 32 owns that worksheet and the reasoning behind it.

Here is the part that belongs to this chapter. The AUS acts on the income figure you enter. Type \$9,020.83 and the system will evaluate the loan at \$9,020.83, return a recommendation, print ratios to two decimals, and generate a document list. It will not tell you that the correct figure under the guide is \$8,916.67, because it does not have the tax returns and does not know the income declined.

The difference is \$104.16 a month. That is small — which is exactly why it is instructive.

WHAT $104.16 OF INCOME IS WORTH                          [constructed illustration]
Assume, only for this arithmetic, a total monthly obligation of $3,870.00 on
this file. Fulton Avenue's actual structure is Chapter 32's.

   at $9,020.83 entered:   $3,870.00 / $9,020.83  =  42.90%
   at $8,916.67 correct:   $3,870.00 / $8,916.67  =  43.40%
                                                     ────────
                                          half a point of DTI

Half a point. On most files, invisible. On a file sitting against a program maximum or an overlay threshold, half a point is the entire decision — and the loan officer who entered the higher number got an approval that the underwriter will take away on day 23, having spent three weeks of the borrower's life and their appraisal fee on it.

The judgment about which figure to enter is not the machine's. It is the guide's, applied by a human who read the returns. That is Chapter 11's work and Chapter 32's work, and this chapter's only contribution is to say: the AUS will faithfully evaluate whichever answer you give it, with no opinion whatsoever about whether you got it right.

15.10 What the AUS does not decide

We can now say the honest thing plainly.

The AUS does not decide anything. It returns a recommendation, evaluated against published rules, on data a human supplied. The lender decides.

That is not a technicality and it is not modesty about software. It is the actual structure of the transaction, and it has five practical consequences.

One: overlays sit on top of every recommendation. Chapter 14 §14.7 defined an overlay as a lender's own requirement layered above the agency guideline. An Approve/Eligible does not repeal one. If your employer requires a 640 minimum score on this product and the findings approve a 620, your employer's answer is 640. The findings do not know your employer exists, and the borrower's loan is being made by your employer.

Two: the underwriter still underwrites. The verification messages describe what must be provided; a human still has to look at what arrives and decide whether it supports what was entered. An Approve/Eligible obtained on data the documents do not support is not an approval — it is a discrepancy that has not been discovered yet. Chapter 19 is about what happens when it is.

Three: the AUS does not judge documents. It cannot tell a genuine written verification of employment from a fabricated one, a real bank statement from an edited PDF, or an arm's-length transaction from one that is not. Fraud detection is human and systemic and it happens elsewhere (Chapter 27).

Four: the AUS does not decide which figure is the right figure. Fulton Avenue, above. Which income is countable, whether a debt is excludable, whether a deposit is sourced, whether a gift is really a gift — those are guideline judgments made by people, entered as data, and then evaluated.

Five, and most important: the AUS does not decide whether this household can afford this house. The Linden Street file returns Approve/Eligible at a 42.66% back-end ratio. That means forty-three cents of every pre-tax dollar this household earns is committed before a single grocery is bought. It is an approvable ratio and it may be a miserable life, and no automated system will ever raise its hand about that. Chapter 4 made this argument and the book will keep making it: DTI is the most-quoted number in origination and it routinely approves borrowers who cannot afford the house and declines borrowers who can. The person who can raise a hand about it is you, on the phone, in week one, for free.

The sentence you should never say

A colleague will tell you, at some point, "Fannie Mae denied my borrower."

That sentence is not possible, and Chapter 1 flagged it in its very first Check Your Understanding. Fannie Mae does not evaluate individual borrowers for individual lenders. An automated underwriting system evaluated a file against Fannie Mae's published guidelines and returned a recommendation. If the recommendation was Refer, the file may still be manually underwritten, restructured, run through Freddie Mac's system, or placed in a different program. If the lender subsequently declines the application, then the lender denied the loan — the lender, whose name is on the adverse action notice, and whose obligation under ECOA and Regulation B is to state the specific principal reasons for the decision.

That obligation does not soften because a model was involved. The Consumer Financial Protection Bureau has said directly that creditors using complex algorithms must still provide accurate, specific reasons for adverse action, and that the difficulty of explaining a model is not a defense. Chapter 25 works fair lending in full and this chapter's second case study takes up the algorithmic question honestly.

So the correct sentence is: "The file came back Refer, here is why I think so, and here is what we do next." It is longer. It is also true, and it is the sentence that keeps the borrower's loan alive.

🗂️ The Loan File

Chapter 15 contribution: the day-6 findings, the document list they produced, and the day-47 re-run.

Day 6, 4:52 p.m. — the first submission

The application was completed yesterday (day 5). Credit is pulled, the income worksheet is built, and the structure is set: conventional thirty-year fixed, \$365,750 at 6.625% with a half point, 95% loan-to-value, borrower-paid monthly mortgage insurance at a 0.58% annual factor.

The findings return Approve/Eligible on submission number 1. Reconcile the five figures before reading anything else:

Ratios: 28.89% housing, 42.66% total debt. Representative score 706. Eligibility clean — a 95.00% conventional purchase on a one-unit detached primary residence is inside the product.

No value acceptance offer. At 95% loan-to-value on a purchase, this file is outside the criteria, and verification message 11 requires an appraisal. It goes on order tomorrow, day 7, alongside title and the two written verification of employment requests.

The document request the machine wrote

The fourteen verification messages became this, sent to the borrowers at 5:20 p.m. on day 6 with a copy to the processor. This list is not the loan officer's invention. It is page five, translated into English and given dates.

Note the sequencing. Items 2, 4, and 8 depend on third parties and go out first even though they sit in the middle of the list; the insurance binder is last because it is fast and cannot be issued too early. The file already holds three months of statements on both accounts where the message asks for two — that is fine. More is never a problem; less is.

Day 47 — the re-run

Between day 6 and day 44 the file behaved: appraisal in at \$385,000 on day 16, submitted day 23, conditional approval with eleven conditions on day 28, a \$4,900 commission deposit sourced on day 33. Then on day 41 the borrowers financed \$5,200 of furniture at \$611.00 a month without telling anyone, and the day-44 (Friday) pre-closing credit refresh found it. Total obligations went from \$4,479.72 to \$5,090.72 and the back-end ratio from 42.66% to 48.48%. Chapter 19 tells that story and works the conditions; this checkpoint owns what happened over the following weekend.

Day 45 (Saturday) the original closing date passed, missed. On day 46 (Sunday) the borrowers paid the furniture account off online and closed it. On day 47 (Monday) the payoff documentation was submitted and the file was re-run. Here is what changed in the data and what did not:

Check the arithmetic. Day 44: \$4,479.72 + \$611.00 = \$5,090.72, and \$5,090.72 ÷ \$10,500.00 = 48.48%. Day 47: the \$611.00 comes back out, obligations return to \$4,479.72, and \$4,479.72 ÷ \$10,500.00 = 42.66% — exactly where the file started. The reserve line is the one that did not come back: paying \$5,200.00 out of verified funds leaves \$12,623.66 − \$5,200.00 = **\$7,423.66, which against a \$3,033.72 payment is 2.45 months** rather than 4.16.

Note the sequence, because it is the section's whole argument in three days. The fact changed first — the borrowers paid the debt on Sunday. The documentation came second — the payoff evidence went in Monday morning. The re-run came last, and its only job was to record what had already happened. A loan officer who had re-run the file on Friday afternoon, before anything was paid, would have gotten 48.48% again, because on Friday afternoon 48.48% was true.

The recommendation held: Approve/Eligible. And it held for a reason worth naming. The input that had broken the file — the monthly debt payment — was genuinely removed, because the borrowers genuinely paid the debt. The input that got worse — reserves — got worse in a dimension where this file had room. Two and a half months of reserves is still a real cushion on a 95% conventional loan.

The re-run also produced new verification messages: evidence that the furniture account is paid in full and closed, and documentation of the source of the \$5,200.00 payoff funds. Those became two more conditions, cleared the same day, and the file went clear to close on day 47 — with the Closing Disclosure issued and received on day 48 and the closing itself on day 51, six days after the date the contract originally set.

What this settles: that the loan fits Fannie Mae's guidelines as underwritten, twice — once on day 6 on estimated facts and once on day 47 on verified ones — and that the document list which carried the file through processing came out of a report generated in ninety seconds on day 6.

What it does not settle: anything the underwriter has not yet seen, anything the lender's overlays require, and whether this household is comfortable at 42.66%.

Open questions carried forward:

• Q11. The re-run cost \$5,200 in cash and pushed the closing from day 45 to day 51 — six days, against a lock that has a date on it. What would the extension have cost? (Chapter 30)
• Q12. Eleven conditions came back on day 28 and two more on day 47. Which of them could have been anticipated from the day-6 verification messages? (Chapter 19)
• Q13. The borrowers are still taking calls from another lender. (Chapter 40)

Your task. In Appendix C's workbook, paste or reconstruct the day-6 verification-message block and do two things with it. First, convert every message into a borrower-facing request in plain English with a due date — do not add anything and do not leave anything out. Second, next to each item, write who controls the turn time: the borrower, a third party, or you. Count the third-party items. That number is your real critical path, and it is almost always larger than new loan officers expect.

Conclusion

An automated underwriting system takes an application and a credit report, checks the loan against a published rulebook, evaluates the borrower against a risk model nobody outside the agency can see, and returns three things: a recommendation in two halves, an assessment of the loan's eligibility, and a numbered list of what must be documented.

The two halves are the part the exam tests and the part that saves files. The first half — Approve, Accept, Refer, Caution — is about the borrower. The second — Eligible, Ineligible — is about the loan. Approve/Ineligible is therefore an ordinary outcome and not a contradiction: a creditworthy household on a loan structured outside a parameter, which usually means find another \$3,850, not call the borrower with bad news.

The verification messages are the part that makes you good at this job. They are the document list, generated from your own data, and converting them into a dated request on the day the findings arrive is the difference between a file that closes in fifty-one days and one that closes in sixty-five.

And the whole apparatus rests on data a human typed. Re-running an unchanged file changes nothing, because the system is deterministic; re-running after the facts genuinely change is how a file gets restructured; and re-running until the numbers approve is not a technique but a misrepresentation with a timestamp on it. A mistyped income produces a beautifully formatted answer about a borrower who does not exist — and only a loan officer who reconciles the report against their own worksheet will ever find out.

Which brings us back to the sentence at the top. The machine did not approve anybody. It read what you typed and told you what it would need in order to believe you. The lender approves the loan, the underwriter reads the documents, the overlays apply, and the borrower decides whether they can live with the payment.

Next: everything in this chapter assumed conventional financing and glanced at the TOTAL Scorecard in passing. Chapter 16 takes FHA seriously — the insurance that makes it work, the upfront and annual premiums, the credit and ratio structure, and why the Harlow Street file, at a 641 score with 51% total obligations, is a good loan rather than a marginal one.

Key Terms

Automated underwriting system (AUS) — software that evaluates a loan application and credit report against a published rulebook and a proprietary risk model, returning a recommendation, an eligibility assessment, and the documentation required to support them. (Ch.15)

Desktop Underwriter (DU) — Fannie Mae's automated underwriting system; returns Approve or Refer paired with Eligible or Ineligible, and produces an Underwriting Findings report. (Ch.15)

Loan Product Advisor (LPA) — Freddie Mac's automated underwriting system, formerly Loan Prospector; returns Accept or Caution and produces a Feedback Certificate. (Ch.15)

TOTAL Scorecard — FHA's risk scorecard (Technology Open To Approved Lenders), run through an approved AUS rather than standalone; returns Accept or Refer for FHA-insured loans. (Ch.15)

Findings report — the document an AUS produces: recommendation, the loan data used, the underwriting analysis, the risk and eligibility assessments, verification messages, and observations. (Ch.15)

Approve/Eligible — the recommendation in which the risk assessment endorses the borrower and the loan meets the product and program parameters. Not an approval; a recommendation. (Ch.15)

Approve/Ineligible — a creditworthy borrower on a loan that breaks a product or program parameter — loan amount, LTV, property type, occupancy, term, or program limit. Fix the structure, not the borrower. (Ch.15)

Refer — DU's and TOTAL's risk result meaning the model will not endorse the file on its own; routes to manual underwriting, restructuring, or another system. Not a denial. (Ch.15)

Caution — Loan Product Advisor's equivalent of a Refer. (Ch.15)

Verification messages — the numbered items in a findings report specifying the documentation required to support the recommendation; effectively the file's document list, generated from the data entered. (Ch.15)

Appraisal waiver / value acceptance — an agency offer, appearing in the findings for a specific casefile, to accept the stated property value without a traditional appraisal; Fannie Mae's value acceptance and Freddie Mac's automated collateral evaluation, plus property-data hybrids. (Ch.15)

Re-run — a subsequent AUS submission on the same casefile. Changes the recommendation only if the inputs change; the system is deterministic on a given engine version. (Ch.15)

Data integrity — the loan officer's ownership of the accuracy of every field submitted to the AUS; the condition on which a findings report has any value at all. (Ch.15)

Spaced Review

1. (Ch. 11 + Ch. 15) Borrower 2 on the Linden Street file earns a \$2,400.00 base plus \$1,800.00 of averaged commission. Explain, in two sentences, why the commission produced a written-VOE verification message that the base salary did not — and what would have happened to the findings if the loan officer had entered the entire \$4,200.00 as base salary.

2. (Ch. 14 + Ch. 15) A file returns Approve/Eligible with a 622 representative score. Your employer's published overlay on this product is a 640 minimum. Who decided what, and what do you tell the borrower? Use the words guideline and overlay correctly.

3. (Ch. 15) A conventional purchase returns Approve/Ineligible with a message stating that the loan amount exceeds the applicable limit. Name the half of the recommendation that failed, name three structural changes that could clear it, and say which one you would investigate first and why.

4. (Ch. 11 + Ch. 14 + Ch. 15) The Fulton Avenue borrower's cash-flow analysis produces \$9,020.83 as a 24-month average and \$8,916.67 for the most recent year, and the underwriter uses \$8,916.67. If the loan officer enters \$9,020.83 into the AUS and gets Approve/Eligible, what exactly does that recommendation establish? Answer in one sentence, then say what happens on the day the underwriter reads the returns.

5. (Ch. 15) Your colleague re-submits an unchanged file to DU for the third time this afternoon. Explain, in the terms of §15.8, why the answer will be identical — then name the two things that can change a recommendation without anyone changing the application data.