Loan delivery data validation, explained.

Selling a closed mortgage, told through one Friday deadline

The road from closing table to investor, in five steps

1. The loan closesThe borrowers sign the note, the mortgage and the Closing Disclosure. A warehouse line funds the loan.
2. Post-closing assembles the fileTrailing documents arrive: the recorded mortgage, the final title policy, the mortgage insurance (MI) certificate. The loan system exports the delivery file.
3. Delivered data is checked against the signed fileEvery field against its page, the derived figures recomputed, mismatches corrected or sent to quality control. In the example, 11 mismatches on 8 of 48 loans.
4. The loan is delivered and the note certifiedThe agency's edits run on the file, the custodian certifies the note, and the investor buys the loan against the commitment.
5. The investor owns the loan; the lender keeps the promiseQuality control samples, post-purchase adjustments and repurchase requests can reach back to the delivered data for the life of the loan.

Step 3 is the subject of this guide. The lender, borrowers and figures in the example are fictional.

Four surprises in selling a loan as ULDD data

Four delivery terms: ULDD, fatal edit, note certification and pair-off

Thirteen terms are defined in the one-page summary.

Inside a ULDD file: the fields an investor buys, and the signed page behind each one

In shortA ULDD file is an XML export with one record per loan and a few hundred required or conditionally required fields. Almost every field has a signed document it must agree with: the note, the security instrument, the Closing Disclosure, the appraisal, the final underwriting findings, the credit report or the insurance certificate. Some fields are calculated, and the investor recalculates them.

Fannie Mae's Implementation Guide for Loan Delivery Data lists the fields in its Appendix D, which the Selling Guide describes as "a listing of the ULDD required and conditionally required data fields" with their definitions and implementation notes.2 "Conditionally required" carries much of the work: mortgage insurance coverage is required only when the loan-to-value ratio calls for insurance, and project type and dues only when the home sits in a planned unit development or a condominium.

The file is a picture of the loan system at the moment of export. That system was right when each value was keyed, and some values were keyed weeks before closing. A good check therefore starts from the document side: for each field, which page is the final word, and what does it say today?

What one batch of 48 loans brings to the check

The fields are the easy part to count. The example lender checks 286 delivered fields on each loan, 13,728 in the batch, against those 662 documents. The number of fields an investor requires varies by loan type; the conditional ones decide how many apply.

The agency delivery rules in force in October 2026, and the dates still moving

In shortThe rules sit in the agencies' guides more than in federal regulation. The core ones are old and stable: deliver true data, clear every fatal edit, certify the note, match the appraised value. What moves is the dataset. ULDD Phase 5 became mandatory in July 2025, VantageScore 4.0 opened to every seller in September 2026, and the redesigned appraisal dataset, UAD 3.6, is mandatory from November 2, 2026, with an exception that runs into June 2027.

Three changes a delivery team has to plan for

Fannie Mae, Freddie Mac and Ginnie Mae: how each one takes delivery of a loan

In shortFannie Mae and Freddie Mac buy conventional loans from lenders, through a cash window or a swap for securities, and both use the ULDD. Ginnie Mae buys nothing: it guarantees securities that lenders issue from pools of government-insured loans, and its custodians certify each pool against the issuer's schedule of loans. The data discipline is the same in all three.

The volume behind the deadlines

Freddie Mac funded 1.107 million single-family homes in 2025, 817,000 of them purchases and 290,000 refinances, for $389 billion of new business.26 Each of those loans arrived as a delivery file and a certified note, and each was delivered against a lender's commitment date.

Ginnie Mae works differently. The lender, called the issuer, keeps the loans on its own books inside a pool, sells the security, and stays responsible to investors. The issuer must get an initial certification from an approved document custodian before the security is issued, and a final certification "within 12 months of the issue date". An issuer with pools more than three years past due for final certification must post a letter of credit, and "Issuers must repurchase defective mortgages".18

Recomputing loan-to-value, insurance, payment and maturity the way the investor will

In shortSeveral delivered fields are calculations. The investor recalculates them from the other fields, so a wrong input produces a consistent but wrong answer. Recompute them from the documents: loan-to-value from the lower of price and appraisal, rounded up; insurance coverage from that ratio; payment from amount, rate and term; maturity from the first payment date.

Loan-to-value, Fannie Mae's way

Fannie Mae's rule: "The property value is the lower of the sales price or the current appraised value." The result "must be truncated (shortened) to two decimal places" and "then rounded up to the nearest whole percent", so "94.01% will be delivered as 95%" while "80.001% will be delivered as 80%".4 In the example, $412,000 ÷ $509,000 = 80.943%, truncated to 80.94% and delivered as 81%.

That one point changes more than the loan-to-value field. Above 80%, a conventional loan needs mortgage insurance, and for a fixed-rate loan over 20 years at 80.01% to 85% the standard coverage is 12%.27 The loan also moves into a different price bucket. The same appraised value has to match the value the appraisal portal holds, or the loan breaks another rule.15

Payment and maturity

A relocked loan in the example batch was delivered at the original rate. At 6.750%, a $356,250 loan over 360 months pays $2,310.63 a month; at the 6.875% on the signed note it pays $2,340.31. The edit that tests payment against rate passes, because the delivered payment and the delivered rate agree with each other. Only the note shows both are stale.

Maturity follows the first payment: a 360-month loan with a first payment on November 1, 2026 matures October 1, 2056, the date of the 360th payment. Move the first payment to December 1 and the maturity moves with it, which is why a rescheduled closing can leave both dates wrong and still consistent.

Debt-to-income and the credit score: the final run counts

Fannie Mae requires that "The data submitted to DU must reflect the loan as it was closed". After closing changes, the loan must be resubmitted when the recalculated debt-to-income ratio exceeds 45%, or rises "by 3 percentage points or more (if the recalculated DTI ratio is 50% or less)".16 The delivered ratio has to come from that final submission.

In the example, one loan was delivered at 38.2% from submission 5; submission 7, the final one, reads 44.6% and is still approved. The delivered figure is wrong even though the loan is eligible.

Credit scores work the same way. A score from the first credit report is the wrong score when the report was pulled again: in the example, 724 on the first report and 716 on the re-pull of September 18, which moves the loan from the 720–739 price bucket to 700–719. Since September 2026 the delivered score model, Classic FICO or VantageScore 4.0, must also match the report.9

Fatal edits, the custodian and the value match: what each check before purchase can see

In shortThree checks run before an agency buys a loan. The edits read the delivery file and the agency's own data. The custodian reads the original note. The appraisal portal matches one value. None of them reads the final application, the underwriting findings, the credit report or the insurance certificate against the file, so those fields reach the investor checked only by the lender.

The edits are rules on the data: is a required field present, is a value in range, do two fields agree, is the loan eligible. Fannie Mae runs them throughout "the submission, loan documentation certification and acquisition purchase process", and a loan with a fatal edit is not bought.2 Freddie Mac's Loan Selling Advisor shows "purchase edit messages" in the same way.17 Fannie Mae's optional EarlyCheck service lets a lender check loans against the Loan Delivery rules before it delivers them, so many problems surface days ahead.23

The custodian holds the original note. Under Fannie Mae's process it reviews the documents, and where the note and the data disagree it proposes data changes that the lender must accept or decline; a clean package reaching the custodian by first-morning delivery is certified the next business day.14 Ginnie Mae's custodians certify that documents "are consistent with the list of loans" on the issuer's schedule.19

The example batch, mismatch by mismatch

The two the edits catch are cheap to find and to fix. The three the custodian catches are costly in time, because the loan waits unpurchased until the data and the note agree. The six nobody catches before purchase are the ones that matter after the sale: a value that changes the insurance requirement, a debt-to-income ratio from the wrong run, an occupancy that contradicts the signed application.

Four ways delivered data drifts from the signed file, traced through eleven mismatches

In shortMost wrong delivery values are well formed. They are a true value from the wrong moment, a value from the wrong document, a value typed instead of read, or a value that the documents themselves contradict. Each has a different cause in the loan's life. The first three are corrected from the signed page; the last needs a person to decide what the loan is.

Fictional example batch: 48 loans, 11 mismatches on 8 loans.

What industry quality control finds

ACES's data covers loans that lenders chose for full-file quality control reviews. Its overall critical defect rate rose to 1.71% in the first quarter of 2026, from 1.38% in the fourth quarter of 2025 and 1.31% a year earlier;28 it had reached 1.79% in the third quarter of 2025.29

ACES counts a defect as critical when it would make the loan uninsurable or ineligible for sale.30 ACES measures underwriting and compliance defects, and its categories overlap with delivery data: an income defect is often a debt-to-income ratio delivered from an earlier run, and a collateral defect can start with the wrong value in the value field.

Where to point the first checks

Post-purchase adjustments, repurchase requests and a data warranty that never expires

In shortA field found wrong after purchase becomes a post-purchase adjustment, a quality-control defect, a fee or a repurchase request. Fannie Mae reprices loan-level price adjustments only for corrections sent within 18 months. Repurchase risk on new agency loans is low and falling, and a growing share of defects now ends in a fee. A repeated data error, though, keeps the lender exposed for the life of every loan it touches.

Fannie Mae expects lenders to report it themselves: lenders must notify Fannie Mae when "they determine there are discrepancies in the data attributes submitted to Fannie Mae at loan delivery" and "follow the post-purchase adjustment process". Corrections are considered for repricing only if sent "within 18 months of the loan's acquisition date", with a $100 minimum adjustment per loan.2 Freddie Mac takes changes after funding through its Post-Fund Data Correction tool, and a change that alters credit fees is adjusted under its Guide.24

How likely a buyback is today

Milliman's index puts the lifetime repurchase probability of loans originated in the first quarter of 2026 at 0.168% for Fannie Mae and 0.252% for Freddie Mac, and notes the figures are "proprietary model-based estimates and not actual mortgage repurchase results".31 Small as the rates are, the loans that are bought back cluster in a few defect types. Freddie Mac's defect report, in Milliman's reading, attributes "more than half of repurchases" to income and collateral defects.31

Fees in place of buybacks, and a fee for poor quality

Freddie Mac's non-acceptable quality (NAQ) rate on incoming loans "reached a high point in the third quarter of 2022", and its repurchase requests peaked in the first quarter of 2023.

In its repurchase-alternative pilot, which lenders choose to join and which began in the first quarter of 2025, "Lenders with a NAQ rate above 2% will be charged a fee in a step-up approach" on the unpaid balance they deliver each quarter, if they deliver enough loans to be measured; lenders outside the pilot can take a fee-only remedy on the defective loan in place of a repurchase.25,33 FHFA announced on October 28, 2024 that Freddie Mac would open the pilot to all approved lenders.33

Commitments, extensions and pair-offs: the clock every delivery batch runs against

In shortA mandatory commitment promises an amount by a date at a price. Loans that miss the date because their data does not clear can be extended for a fee or paired off, and a pair-off fee follows the market. Fannie Mae measures a pair-off from the original commitment amount, so each blocked loan counts in full.

Fannie Mae's mandatory whole-loan commitments "can be taken for 1 to 90 days" and must expire on a business day.34 A lender can extend before expiry, for "up to a maximum of 30 days" in total; a commitment that expires unfilled and never extended is extended automatically by five calendar days, and if nothing arrives Fannie Mae pairs off the balance and drafts the fee from the lender's account.35

The fee comes from comparing the price at commitment with the price at pair-off; when only the market moved in the lender's favor, Fannie Mae may pay cash back.36,35 And "All pair-off amounts (including cash back pair-offs) are calculated from the original commitment amount".36

In the example, four loans worth $1,608,650 would stop at purchase: one on a fatal edit and three on notes the custodian could not certify. Fixed on Monday, they deliver with the rest.

The tolerance decides the rest. Fannie Mae counts a whole-loan commitment as filled when the loans delivered fall short of it by no more than the greater of $10,000 or 2.5% of the original amount.37 On $18,000,000 that is $450,000, so the 47 loans left after the check, $403,600 short, fill it. With the four blocked loans missing Friday, the batch falls $1,672,250 short, outside the tolerance, and the pair-off is measured from the full $18,000,000.36

Meanwhile each unpurchased loan stays on the warehouse line, accruing interest, and counts against the line's limits on how long a loan may sit before sale. No public figure measures what a day of purchase suspense costs; your warehouse agreement and your own suspense report do.

When the signed documents disagree with each other: occupancy, income and the loan that leaves the batch

In shortMost mismatches have one right answer, the signed page. Some do not: the documents themselves contradict each other, and the loan may have been underwritten and priced on the wrong facts. Occupancy is the classic case. Those loans go to quality control with both pages, and out of the batch until a person decides.

Occupancy sets eligibility and price. A primary residence, a second home and an investment property carry different loan-to-value limits and different price adjustments, and Fannie Mae lists occupancy among the attributes the underwriting data "must reflect" as the loan closed.16 In the example, the delivery file says primary residence. Page 4 of the signed final application marks an investment property with $2,150 a month of expected rent, and the appraiser found a tenant with a lease through May 31, 2027. The loan was underwritten, locked and priced as a primary residence.

Changing the field to "investment" would make the file agree with one document and disagree with the others, and with the underwriting. The right move is a referral: both pages to the quality control manager, the loan out of this delivery, and a decision on whether it can be sold at all, and on what terms.

Delivery work in figures: loan volumes, cost per loan, defect rates and wages

In shortIndependent mortgage banks and bank mortgage subsidiaries spent about $11,100 to produce each loan in 2025 and kept $785 of production profit on it, so a single price adjustment or a week of suspense is a large share of the margin. Critical defect rates in lenders' full-file quality control reviews rose again in early 2026. Nobody publishes how often delivery files disagree with their documents.

A margin of $785 on a loan means the example's Abernathy correction, $1,110 on a $296,000 loan, is larger than the profit on that loan. Price adjustments found before sale and after sale cost the same in price; found after sale they also cost the work of a post-purchase adjustment, and they can arrive in a batch of corrections for the same field.

Government statistics class most of the people who do this work with loan interviewers and clerks: $50,020 a year at the median in 2025, across 169,100 jobs.40 With wages at 70.0% of private employers' compensation costs in June 2026, a loaded hour costs about $34.41 Treat that as a starting point and use your own pay scale.

Software that reads the closing file: what the agencies' AI rules and the 2026 model guidance expect

In shortThe agencies' edits check the data; software can now read the documents too, tie each field to its page, recompute what the investor recomputes, and draft corrections. Freddie Mac has required AI governance from sellers since March 2026, and Fannie Mae's framework moved into a security supplement in October 2026. For banks, the April 2026 model risk guidance leaves fixed rules and generative AI to the bank's own controls.

What automation changes, and what it leaves

• Every field can be checked on every loan, including the conditionally required ones, at the same care on the last day of the month as on the first.
• Each value can carry the page it came from, so a correction, a custodian question or a later repurchase request is answered in seconds.
• Recomputed fields use the investor's rule, including Fannie Mae's rounding of loan-to-value.4
• Contradictions between documents still need a person. So do decisions about price, commitments and release, and the warranty stays with the lender that signs.1

Testing a delivery data checker with planted mismatches

Ask for the same proof you would ask of a new analyst: take a recent batch, plant known mismatches in copies of the files (a rate from an old lock, a sales price in the value field, a score from a first report, a misspelled name), keep some files clean, and count what is caught and what is flagged wrongly.

The number planted matters as much as the number caught. With twelve plants of a kind and none missed, the true miss rate could still be as high as 22% at 95% confidence; it takes 59 plants with none missed to bring that bound near 5%. We found no published independent accuracy test of any delivery data validation product, as of October 2026.

Who touches a loan between closing and sale, and where the month-end hours go

In shortA post-closing team stacks the file and chases trailing documents, a delivery analyst exports and checks the data, secondary marketing manages the commitment and releases the loans, and quality control samples them afterwards. Most checking time goes to comparing the export with the PDFs by eye, under the pressure of the commitment date.

Who touches a loan before sale: post-closing, delivery, secondary, QC

Post-closing to investor purchase: the delivery hand-offs

About 30 analyst hours on a hand-checked 48-loan batch

Few delivery analysts get 30 quiet hours. The same week brings new commitments, trailing documents, custodian exceptions from the last batch and the month-end volume of closings. That is why teams sample: check every field on some loans, or a short list of fields on every loan, and trust the edits for the rest.

Stare-and-compare, a post-closing service or delivery edit tools

Batch DB-1012: 48 loans checked against their signed pages before a Friday release

In shortThis is the example from the top of the page, inside the product. Six agents read the delivery file and 662 closing documents and draw 13,728 lines between delivered fields and signed pages. Eleven lines break. Danielle Okafor, the loan delivery analyst, accepts ten drafted corrections, sends one loan to quality control and signs the data; Marcus Feld, the secondary marketing manager, releases 47 loans.

1

Forty-eight loans, four days, nothing matched yet

Monday, October 12. The match screen opens on the work: the 48 loans of batch DB-1012 as a row of borrower names, and the selected loan's deciding fields as lines, the delivery file on the left exactly as exported and the signed closing file on the right. Commitment CW-26-4471 is $18,000,000 and expires Friday, October 16.

Each line is labeled with the check the buyer will run on that field: the custodian's note certification, a delivery edit, the price grid, eligibility. Zero of 13,728 fields are matched so far.

2

662 documents read in under two minutes

Danielle presses Match all 48 loans. The delivery file reader maps 286 fields a loan to the page each must agree with. The closing document reader reads 48 notes, 48 security instruments, 62 riders and 48 Closing Disclosures; the appraisal and underwriting reader reads the appraisals, final findings, transmittals and credit reports. The field matcher recomputes LTV, CLTV, DTI, payment and maturity on every loan.

The run finishes in 1 minute 43 seconds with 11 mismatches on 8 loans, ranked by what they would cost.

3

Four lines snap red on one loan

On the Brandt loan the first payment and maturity lines break against Note page 1, labeled "fails · Custodian · note certification". Underneath, the reason, the consequence ("purchase stays in suspense") and a correction drafted from the note. The value line breaks against Appraisal page 3: the file carries the $515,000 sales price, the appraisal says $509,000.

The right rail totals the batch as it stands: 4 loans, $1,608,650, would stop at purchase and miss Friday; 3 would carry a repurchase trigger.

4

An occupancy the documents contradict

On the Pham loan, "Occupancy · Primary residence" meets "Investment property · Loan application p. 4". The run log adds that the appraiser found a tenant. The loan was underwritten and priced as a primary residence, so no correction is drafted.

Danielle presses Send to QC, take out of batch. The loan goes to Owen Whitaker in quality control with both pages, and the batch becomes 47 loans, $17,596,400.

5

Corrections accepted, lines healed

Back on Brandt, each correction cites its page: first payment 11/01/2026 and maturity 10/01/2056 from Note page 1; value $509,000 and LTV of 80.94% from Appraisal page 3, which the agency rule delivers as 81%; 12% coverage on certificate HM-5513082; planned unit development with $42 a month in dues. Danielle accepts all four and each line turns green with her name on it.

The exported file is never overwritten. Accepted corrections make a new version of it.

The solution's design goals are every delivered field tied to a signed document before sale, and zero fatal edits or note certification exceptions at delivery. They are targets, to be measured on your own batches. Edit numbers on these screens, such as W-118 and F-233, were invented for the product's sample; the agencies give their edits their own identifiers.

Eight more screens, from the note open beside its fields to the rules your managers set

In shortOne batch is the story. These screens show the rest of the work: the full field-by-field comparison with the document open, the mismatch list ranked by cost, the edits and the custodian preview re-run, the clean batch, the release, the queue of batches, the quality trend and the settings.

Select any screen to open it larger. The lender, investor, custodian, insurer, borrowers and figures are fictional; price effects come from the sample lender's own grid. One setting, rounding, shows LTV to two decimals; Fannie Mae delivers it rounded up to a whole percent (section 05). Edit numbers on the screens, such as F-417 and W-118, were invented for the product's sample; the agencies give their edits their own identifiers. Mortgage identification numbers (MINs) are masked.

Six agents read, match and price. The analyst accepts and signs; the manager releases.

In shortEach agent has one job and limits it never crosses. None of them edits the exported file, changes a price or a commitment, releases a batch or sends anything to an investor. The loan delivery analyst accepts every correction and signs the data; the secondary marketing manager releases only with zero fatal edits and zero certification exceptions.

The delivery analyst signs the data; the secondary manager releases

Every agent run, correction, referral, signature, export and release is on the audit trail with who and when. Settings changes are versioned. Agent steps and human decisions are recorded apart.

Danielle, Marcus and Renee on delivery batch DB-1012

The ULDD export and closing packages it reads, and the signed batch it releases

In shortIt reads the files your team already produces and the documents your closings already create, and leaves behind a field-by-field match report, mismatches ranked by cost, corrections with their pages, a re-run of the edits and the custodian preview, and a signed release. Your loan system and the agencies' systems stay where they are.

Loan system, agency delivery, pricing engine and custodian links

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

Your loan system export and the agencies' edits stay in place

Every delivered value carries its page, and the export is never overwritten

What a wrong delivery field costs you: suspense days, price adjustments and buybacks

In shortStart from your own records. No agency publishes how often delivered data disagrees with its documents or what a day of purchase suspense costs, so your suspense report, your post-purchase adjustments and your quality control findings are the numbers that matter. A few published figures frame them: the margin on a loan, the clock on price corrections and the chance of a buyback.

Counting your own suspense items and post-purchase adjustments

Repeated mapping errors, warehouse room and your next trade price

Putting a number on checking every delivered field before sale

In shortPut in your own numbers. The calculator starts from a published wage and a published repurchase rate and asks you for the rest: loans delivered a year, the checking time they take now and after, how much repurchase loss you think a document check could prevent, and what the tool and your testing cost a year. Every default marked illustrative is an assumption to replace.

Analyst time and repurchase exposure found before sale

Your analyst hours and buyback risk across a year of deliveries

At the defaults, 4,800 loans a year release 2,400 analyst hours worth $81,600 and avoid about $130,560 of modelled lifetime repurchase loss, $152,160 a year after a $60,000 cost. Halve the hours saved and set the preventable share to 5%, and the result falls to $13,440: the case then rests on suspense and pair-offs, which you should count from your own records.

Score your secondary desk on eight questions before you automate delivery checks

In shortPick the answer closest to your team today. Eight questions score 0 to 2; the ninth 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.

Secondary desk scores: measure suspense, pilot on sold batches, or implement

Sixteen questions for any vendor that says it validates delivery data, with the answers to listen for

In shortAsk every vendor the same questions, us included. The answers show whether a tool checks the data against itself or against the documents, which documents it reads, how it treats contradictions, what record it leaves, and how you would measure it on your own loans.

We found no published independent test of any delivery data validation product as of October 2026. A planted-mismatch test on loans you already sold is the fairest comparison, and the same test measures your current process.

An eight-week pilot on batches you have already sold

In shortMeasure first, then test on delivered batches with mismatches you planted, then run alongside live deliveries, and decide with four numbers. The test set uses loans already sold, so it cannot touch a commitment; during the live weeks your usual process still runs, and your analysts sign every batch.

1. Pull your baseline: last quarter's suspense items, custodian exceptions, extensions, pair-offs and post-purchase adjustments, each traced to a field.
2. Choose three delivered batches, about 150 loans with their closing packages. Copy one batch and plant 40 known mismatches across the families in section 07.
3. Load your rules: investors, field maps, price grids by date, rounding, approvers.
4. Run the test set blind, then alongside two live month-end deliveries, with your analysts signing every batch.
5. Decide with four numbers: catches by type, false alarms on clean loans, analyst hours per batch, and mismatches found that the edits and custodian would have passed.

Measuring delivery validation: catches by family, false flags, hidden mismatches

ULDD, fatal edits, note certification, pair-offs and repurchase risk: delivery questions answered

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