Mortgage guideline search, explained.

What one underwriter's guideline question involves, from the rate lock to the file note

The Calders' student loan question, application to closing in five steps

1. The application comes inThe Calders apply to buy a $415,000 house. The loan officer prices it and locks it with an investor, Northcastle Capital.
2. Processing gathers the proofPay stubs, the credit report, the servicer's letter on $68,400 of student loans in forbearance.
3. The underwriter asks which rule appliesWhich payment counts for a student loan showing $0? The agency, the investor and the lender each have a say. This is where the guideline is found, applied and cited.
4. The loan closesConditions cleared, documents signed, funds out on Friday, October 30.
5. The investor buys it, and QC reads it laterPurchased the week of November 9. Post-closing quality control re-underwrites a sample against the rules in force.

Step 3 is the subject of this guide. The people, lenders, investors and figures in this example are fictional.

Four surprises in reading mortgage underwriting guidelines

Four guideline terms: agency guide, overlay, date basis, DTI

Fourteen terms are defined in the one-page summary.

Seven rule books a single home loan can answer to, and who writes each one

In shortThe program decides the agency guide: conventional conforming loans follow Fannie Mae or Freddie Mac, government loans follow FHA, VA or USDA. Then the investor's seller guide and matrix, and the lender's credit policy, add their own rules. Most purchase loans answer to a conventional guide, and almost a third to a government handbook.

Which books carry the most loans

Conforming loans are those within the limit Fannie Mae and Freddie Mac may buy: $832,750 for a one-unit home in most of the country in 2026, up to $1,249,125 in high-cost areas.16 A conforming loan a lender keeps on its own books still tends to be written to the Selling Guide, because that keeps it saleable. FHA, VA and USDA loans together were almost one in three purchase loans in 2023. In that year 5,113 lenders reported 5.7 million originations of all kinds.15

Conventional, government, investor, lender: one loan, four layers

More than 50 rule documents in ten months: tracking Selling Guide, Freddie Mac, FHA and VA changes in 2026

In shortBetween January and mid-October 2026, Fannie Mae issued nine Selling Guide announcements and eight lender letters, Freddie Mac at least 21 Guide bulletins, HUD eleven mortgagee letters and a Handbook update, and VA four circulars. Several changes reached questions underwriters ask every day: credit scores, student loans, income verification.

Counted on October 30, 2026: Fannie Mae and Freddie Mac from a QC firm's published lists and the Selling Guide's topic pages, FHA by mortgagee letter number, VA from its own circulars page. Freddie Mac's lettered bulletins A, E and I were not found, so its 21 is a floor. Freddie Mac's bulletins cover selling and servicing; Fannie Mae's servicing announcements are not counted. A vendor blog in 2026 still described Freddie Mac's pace as "typically 10 to 20 per year".19

The changes reach topics underwriters open every day. A desk note on student loans written in 2025 describes a text that has since been amended four times.5

Four changes that moved everyday answers

HUD also published an update to Handbook 4000.1 on August 12, 2026, which among other things defined written, electronic and re-verification of employment,21 and USDA replaced the income-limit appendix of HB-1-3555 in July.14 VA issued or changed four circulars, mostly on fees and partial claims.12 None of those agencies' changes reached the student loan rules in this guide's example, which shows the other half of the job: knowing which changes can be ignored for a file.

Effective dates and the date in the file: deciding which text a loan is judged by

In shortEvery change says which loans it reaches: those with applications taken, case numbers assigned, locks made, notes signed or loans delivered or purchased on or after a date. Find that date basis in the change itself, then find the matching date in the loan file. Two loans closing the same day can fall under different texts.

The agencies date their rules in different ways, and a desk that reads several books has to know each method. Fannie Mae puts a date in parentheses beside each topic: the date of the most recent announcement that amended it. The policy's effective date is in the announcement itself.4

Freddie Mac attaches effective dates to its sections and states the date basis in each bulletin; for example, its 2025 change to resale-restricted properties applied to mortgages with application received dates on or after March 3, 2026.9 FHA often counts from the day the FHA case number is assigned.11

One bulletin, one loan, six dates

A desk that matches the bulletin to the lock date would decide that this loan escapes it: the lock came first. The bulletin counts from purchase, so it reaches every Northcastle loan still in the pipeline on November 2. Reading the date basis is as much a part of the rule as the sentence that sets the payment.

Investor and lender overlays: why the stricter rule governs, and where overlays come from

In shortAn overlay is any rule stricter than the agency program: a higher score floor, a lower DTI cap, one option removed from a calculation, an extra document. Lenders and investors overlay because they carry the cost when a sold loan turns out to break the agency's rules. The agencies allow it, if the overlay is applied the same way to similar loans.

Why lenders and investors add rules

A lender that sells a loan to Fannie Mae or Freddie Mac warrants that it meets the guide. When a review finds it did not, the enterprise can ask the lender to buy it back. FHFA, which regulates both, has said that before its repurchase framework began taking shape in 2012 the enterprises' discretion "generated uncertainty for lenders and can contribute to increased credit overlays that affect lending costs".24

The Federal Reserve's May 2012 survey asked banks why they had become less willing to make an eligible loan to a borrower with a 620 score and 10% down; the risk that the enterprises would put back a delinquent loan was "listed as the most important one by the largest number of banks".25

An investor that buys loans from smaller lenders carries that risk on both sides: from the agency it delivers to, and from the lenders it bought from. Its matrix keeps its pipeline away from the edges that cost it most. The lender's own committee may then add rules for its own appetite, which apply to every loan it makes whoever buys it.

Four kinds of overlay

The fourth kind is documentary: more months of reserves, a fresher document, a narrower list of eligible properties.

A student loan showing $0, counted under five rule books: the arithmetic

In shortWhen a credit report shows a $0 student loan payment, each book sets a stand-in payment its own way: Fannie Mae 1% of the balance or the amortizing payment, Freddie Mac and FHA 0.5%, VA 5% of the balance divided by 12 unless deferment runs a year past closing. On a $68,400 balance that is anything from $285 to $684 a month, which moves this borrower's DTI by about four points.

Federal law makes the ratio part of every residential mortgage decision: under the ability-to-repay rule a creditor must consider "the consumer's current debt obligations" and "the consumer's monthly debt-to-income ratio or residual income".28 The agencies then say how to put a number on each debt.

What each payment does to the file

The Calders earn $9,860 a month. The new house payment is $3,308: principal and interest $2,460, taxes $593, insurance $130 and mortgage insurance $125. A car loan costs $468 and credit cards $112. Before the student loans, debts come to $3,888.

Three options keep the file, each resting on a different passage: move the lock to an investor that follows the Selling Guide on this point (44.4%), pay off the car loan at closing and stay with Northcastle (41.6%), or move the loan to FHA at 0.5% (42.9%), with FHA's own mortgage insurance and new disclosures. A fourth option, counting $0 because the borrower "isn't paying anything", is the one no book allows for a loan whose forbearance ends five months after closing.

The eight mismatches behind a wrong underwriting guideline answer, and the check that catches each

In shortMost wrong guideline answers come from matching the wrong rule to the file: the wrong investor, the wrong version, the wrong date basis, a looser agency sentence quoted when an overlay governs, or a saved answer written before the last change. Reading the sentence itself is rarely the problem.

The errors share a pattern: each one produces an answer that sounds right. A Selling Guide sentence quoted perfectly is still a wrong answer for a loan whose investor removed the option. That is why the strongest control is a note that names the book, the section, the version and the date basis, which lets a reviewer rebuild the answer.

From a misread rule to a defect: QC findings, repurchase demands and the 2025 defect data

In shortA loan that breaks the rule book in force at delivery has a defect. Fannie Mae sorts defects into findings, price adjustments and significant defects, and a significant defect can mean a repurchase demand. Industry QC data show the critical defect rate rising to 1.71% in early 2026, and eligibility defects almost quadrupling their share in 2025.

How Fannie Mae grades a defect

A student loan counted at the wrong payment is a liability error. If it moves the DTI over a limit, it can make the loan "unacceptable for purchase had the true and accurate information about the loan been known". Fannie Mae's reviews include "post-purchase reviews, early payment default reviews, servicing reviews, and post-foreclosure reviews".31,3

What industry QC data shows

ACES, a QC software firm, puts the 2025 rise down to eligibility decisions made at the edge, "consistent with borrowers stretching to qualify in a constrained affordability environment", such as income sources below a minimum history or assets that do not qualify under program rules. Its advice is to address them "at the front end of the decisioning process rather than during post-close review".35

That front end is the underwriter's question. In the first quarter of 2026 legal and compliance findings led at 26.02% of defects, with income and employment second at 20.07%.34

Mortgage guideline AI: the Freddie Mac and Fannie Mae AI terms, and what SR 26-2 leaves to the bank

In shortSince March 2026 Freddie Mac's Guide, and since August 2026 Fannie Mae's terms, make lenders govern any AI used in origination: written policies, an owner, monitoring, disclosure on request, vendor oversight. SR 26-2, the bank regulators' April 2026 model risk guidance, leaves generative AI to each bank's own governance. Freddie Mac also lets lenders feed its Guide into their own generative AI tools.

A general chat assistant answers from what it learned in training, which predates the current version of every guide, and it has never seen an investor's private bulletins. Software built for guideline search has to read the current texts, keep the older versions, match overlays to loans by date, and cite the passage behind each sentence. The rules below govern how a lender uses any such tool.

The practical reading: a lender selling to either enterprise needs a written AI policy, a named owner and evidence of monitoring before a guideline tool goes live, whatever its bank regulator's model guidance covers. The limits of AI here are plain. It cannot know an overlay it has not been given, it is only as current as the bulletins it receives, and the credit decision stays with a named underwriter.

The mortgage market in late 2026: thin margins, rising rates and fewer loans ahead

In shortLenders spent $11,094 to make each loan in 2025 and kept $785 of profit. Rates passed 7% in the fall of 2026, and the Mortgage Bankers Association now expects fewer loans in 2027. With so little margin, a loan that has to be re-locked, restructured or bought back eats the profit on several others.

The same forecast puts 2026 single-family volume at about $2.1 trillion, falling 5% to $1.99 trillion in 2027, with the 30-year rate staying above 7%.44 A falling market raises the stakes of guideline accuracy in two ways. Borrowers stretch to qualify, which is where ACES saw eligibility defects climb.35 Lenders compete for fewer loans, so a borrower who waits a day for a guideline answer has time to call someone else.

Cheat sheets, investor PDFs and a weekly memo: how underwriting desks find guidelines today

In shortLoan officers ask, underwriters answer, a policy and training lead reads every change and keeps the overlay matrix, secondary marketing chooses the investor, and QC reads it all later. The tools are the agencies' websites, investor PDFs in email, desk cheat sheets and a weekly memo. Nobody publishes how long a question takes; one worked example puts a hard one at over 30 minutes.

Who asks and answers guideline questions: loan officer, processor, underwriter, policy lead

A hard guideline question passes three or four desks

Minutes spent finding the overlay and its date

ABT, a mortgage technology firm, describes a loan officer still without "a defensible answer" twenty minutes after opening the first tab, and multiplies 35 minutes a day across 25 loan officers into about 3,650 hours a year. It labels the figure "not a benchmark or a savings guarantee".19 No regulator or study measures it. Two weeks of logging questions on your own desk gives a better number.

Cheat sheets, agency help channels or guideline search tools

The Calder student loan question, from the first answer to a cited note and a new investor

In shortThe example from the top of the page, inside the product. Renée Okafor, a senior underwriter at Halsted Ridge Mortgage, asks which payment to use for Priya Calder's student loans. Five agents read the loan and nine passages and answer with every sentence cited. Renée opens the passages, sees the stricter investor rule beside the agency rule, picks one of three options, and adds a cited note. Secondary marketing re-locks the loan.

1

The answer leads with the payment and the ratio

Renée types the question on the Calder file: Priya's $68,400 of student loans show $0 because they are in forbearance. The answer comes back in 9 seconds in the sample run. While the loan is with Northcastle, qualify at $684, 1% of the balance, which puts the DTI at 46.4%, over Northcastle's 45% cap.

Below the first line sit two citation chips side by side: the Selling Guide's "1% or $486" and the investor's "1% only", marked stricter and governing. A meter shows 44.4% at $486 and 46.4% at $684 against the cap.

2

Five steps, each one shown

"Show the steps" opens the run. The question reader read the loan: conventional, locked with Northcastle, expected purchase the week of November 9. The guide searcher found seven passages across the Selling Guide, the Seller/Servicer Guide, Handbook 4000.1 and the VA handbook. The overlay checker applied NC-26-19, the file impact calculator recomputed the ratio, and the answer writer cited all seven sentences.

The footer states the rule the agents work under: they never change a loan file, a lock or an automated underwriting submission.

3

Both rules, with the reason the stricter one governs

Renée clicks "Both rules". The right panel stacks the agency sentence, Selling Guide B3-6-05 in the edition in force, above the investor's, NC-26-19, published Wednesday, October 7. A line at the top says why the bulletin governs: Northcastle buys this loan the week of November 9, after the bulletin's November 2 start.

Each passage shows only its governing sentence highlighted, so Renée reads the two sentences that decide the file, side by side.

4

Three ways to keep the file, each worked through

Under the answer, three options carry their own ratio: re-lock with Bayline Mortgage Investors, which follows the Selling Guide on this point (44.4%); stay with Northcastle and pay off the $14,900 car loan at closing (41.6%); or move to FHA at 0.5% of the balance, $342 (42.9%). A line lists how the other books count the same loans: Freddie Mac and FHA 0.5%, VA 5% ÷ 12, and no overlay of Halsted Ridge's own.

Renée chooses the re-lock. Lydia Park in secondary marketing re-locks with Bayline at the same rate.

5

A note that a reviewer can rebuild

"Add a cited note" drafts the file note from the answer Renée just read: the balance, the servicer letter of September 28, B3-6-05, NC-26-19, the 45% maximum in Northcastle's matrix 4.02 and the decision to re-lock with Bayline. It carries five citations. Renée can edit it; it goes on the file under her name, with a condition to resubmit the loan to DU at the new payment.

A QC reviewer reading the file next year can open each citation and see the text as it stood that day.

The design target is a 9-second median from question to cited answer, with every sentence carrying a citation and the version in force. Both are targets, to be measured on your own questions.

Inside the cited paragraph, the loan file, the investor bulletin, the change digest and the answer rules

In shortOne question is the story. These screens show the rest: a cited paragraph opened at its sentence, the loan file after the re-lock, one topic compared across every rule book, an investor bulletin with its earlier version, the Monday change digest, one change traced file by file, the saved answers the policy lead approves, the investors' overlay books, and the answer rules.

Select any screen to open it larger. The lender, investors, people and figures are fictional.

What one investor bulletin did to the pipeline

Files are matched on program, investor, purchase date and the facts the rule touches, so the bulletin reaches only conventional files locked with Northcastle, bought from November 2, with a student loan reported at $0. One file, Mensah, sits at 44.9%, a tenth of a point under the cap: the kind of file a weekly memo would never single out.

What each of the six agents prepares, and what the underwriter and the policy lead sign

In shortEach agent has one job and limits it does not cross. None of them changes a loan file, a lock or an automated underwriting submission, and none edits a saved answer. The underwriter decides on the file and signs the note; the policy and training lead approves every saved answer and every edit to one; exceptions go to the credit committee.

Six agents from guideline question to cited file note

The underwriter decides; the policy lead approves saved answers

Every question, answer, note, re-lock request and approval is on the audit trail, with agent steps and human decisions kept apart. Roles decide who may see each investor's books, change settings or approve answers, so the maker of an answer is never its approver.

Renée, Hannah and Doug on the Calder student loan question

What it reads from your agencies, investors and pipeline, and what it leaves on each loan

In shortIt reads the agency guides you sell under, the overlays your investors send you, your own credit policy and the loans in your pipeline. It leaves behind cited answers with the version in force, comparisons across every book, a weekly map of changes to files, cited notes, and saved answers your policy lead approves. Your systems of record stay where they are.

LOS, lock desk, DU and the overlay inbox to connect

For each, check whether it has an API, a file feed or only screens, and who owns it.

Your agency guides, investor bulletins and overlays, kept by version

No passage, no sentence, and the stricter overlay wins

The price of a late or wrong underwriting answer: desk minutes, re-locks and repurchased loans

In shortStart from your own numbers. Nobody publishes how long a guideline question takes or how often the answer is wrong, so the figures that matter are yours. A few published figures frame them: what an hour of underwriting time costs, how often the rule books changed in 2026, and how fast eligibility defects grew in 2025.

Logging five days of guideline questions

Borrowers lost to a faster lender and loans investors refuse

The return on faster, cited underwriting answers, worked out with your own numbers

In shortPut in your own numbers. The calculator starts from a published loan-officer and underwriter wage and asks you for the rest: how many hard questions your desk answers, the minutes they take now and after, and what the tool and your testing cost a year. Every default marked illustrative is an assumption to replace.

Desk minutes and Selling Guide changes mapped to files

Your underwriter minutes on a year of guideline questions

A desk of 25 asking one hard question a day clears the first two lines. A lender that sells only to Fannie Mae with no investor overlays and a handful of questions a week probably does not, and Fannie Mae's own tools may answer most of them.

How ready is your underwriting desk? Eight scored questions on overlays, baselines and AI policy

In shortPick the answer closest to your desk today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

Underwriting desk scores: log questions, test on 100, or implement

Questions that show how a mortgage guideline AI tool picks the version and the overlay

In shortAsk every vendor the same seventeen questions, us included. The answers show what the tool reads, how it decides which version and which overlay apply, what evidence you keep, and how you would measure it on your own questions.

We found no published independent test of any mortgage guideline tool's accuracy, as of October 2026. A test on questions your desk has already answered is the fairest comparison, and the same test measures your current process.

Piloting guideline search on 100 underwriting questions with known answers

In shortMeasure first, then test on your own questions with known answers, then run live alongside your desk, then decide with four numbers. None of these steps commits you to a vendor.

1. Log two weeks of questions: who asked, the loan, minutes to an answer, books opened.
2. Build a test set of 100 answered questions, each with its right answer, book, version and date, including ten with no passage and ten where an overlay governs.
3. Load your overlays: two investors' matrices and bulletins and your own credit overlays, with date bases.
4. Run live for four weeks beside your desk, with underwriters deciding every file.
5. Decide with four numbers: right-rule rate, overlay catch rate, correct "not found" rate, and minutes per answer.

Measuring guideline search: right-rule rate and overlay catch rate

Overlays, student loan payments, Selling Guide updates and AI: guideline questions answered

In shortThe questions people search, each answered on its own.