For credit teams at rehab and ground-up lenders
Underwrite the appraisal.
You underwrite the borrower, the property, and the deal. The appraisal gets read — not tested. Sphinx runs 56 published checks against every after-repair-value appraisal and returns the value the file actually supports, with a range around it, in minutes.
appraised
56 checks published in full, each threshold with its provenance → · findings in minutes · not an appraisal, not a valuation
An ARV is a value for a house that doesn't exist yet.
It rests on three things inside the appraisal: which comparable sales the appraiser chose, how they were adjusted, and whether the renovation budget can produce what those comparables look like. Every one of them gets looked at. Underwriters and review appraisers read the comps, the adjustments and the budget on every file. What differs from lender to lender is what they test, and against what: there is no common standard for the review, and nobody goes back a year later to compare the ARV to what the house sold for.
So the number that sizes the loan is the one number in the file whose track record nobody keeps.
Origination volume from the Urban Institute, April 2026 — roughly $25B ground-up construction, $35B rehab, and $25B bridge and stabilization. Appraisal count is our estimate at typical loan size. Plenty of review happens today: underwriters read every file and some lenders keep staff review appraisers. What doesn't exist is a common definition of what good looks like, or any measurement of whether the values held up.
Three things about every appraisal, before the loan closes.
The value the file supports.
Not what the appraisal says. What its own comparables, adjustments, and budget actually hold up. On the sample file: $402,100 against a stated $455,000.
How firmly.
A range, not a point. Thin comparables, stale sales, heavy adjustments, and a long hold widen it. Deep, recent, nearby sales narrow it. A wide range means the file supports its number loosely — information the credit decision should carry.
What's wrong, and what it costs.
Every flag carries a dollar effect and cites a published check. Ranked by size, so the flags that matter surface first and the 50 that passed stay out of the way.
Every finding is built to be filed. Reproducible in eighteen months by someone who's never heard of Sphinx: the standard version at the time, the thresholds used, the index vintage, the extraction quality, and what the review did not examine.
- For the credit file — the full finding, cited to the standard, formatted as an exhibit. Preview →
- For the borrower — plain language, no check numbers, what the gap is and what would close it. Preview →
Three steps. Nothing about your process changes.
The appraisal is read into structured fields — every comparable, every adjustment, the condition and quality ratings, the licence number, the budget. Each check tests one thing about that data and either passes or fires. Fired checks with a dollar effect move the supported value; fired checks about evidence quality widen the range. Every field carries a confidence score, and a check that can't read what it needs says so instead of guessing.
Every flag cites a check you can read. So can your counterparties.
The 56 checks are published in full — what each tests, what fires it, and where its threshold came from. When a finding says the value isn't supported, it's citing a numbered check in a public document, not a vendor's opinion. That's what makes it usable in a credit memo, a repurchase conversation, or a diligence file.
Of the 56, 20 exist nowhere else, because no other review standard was written for a value that's a forecast. 5 of them test the ARV directly.
Five checks that only exist because the value is a forecast
Conventional review has no equivalent to these, because a conventional appraisal values a house that already exists. Every one of them is computable from the appraisal alone.
- RTL-V-01 Budget versus comparable finish level Does the scope fund the standard the supporting sales were finished to?
- RTL-V-02 Are the supporting sales finished? If the supporting sales are rated C3 or C4, the value rests on unrenovated stock.
- RTL-V-03 As-is plus budget reconciles to ARV A $180k lift on a $90k budget needs explaining. So does a $40k lift.
- RTL-V-04 ARV versus neighbourhood ceiling Compared to the highest recent sale and the 90th percentile $/sf nearby.
- RTL-V-05 Marketing time versus comparable days on market Whether the stated exposure period matches how the submarket actually clears.
Comparable selection
15 checks · RTL-C →Whether the sales chosen support the conclusion, judged against how dense the local market really is.
Adjustments
12 checks · RTL-A →Whether the math is consistent and market-supported, or reached for after the number was already chosen.
Subject & file integrity
16 checks · RTL-S →The routine integrity work that takes twenty minutes and catches something rarely — but expensively when it does.
Appraiser
8 checks · RTL-P →License and discipline are table stakes. Whether they actually work this county is the one nobody looks at.
Thresholds carry their provenance — industry baseline, or fitted to realised sale prices. When one moves, the change publishes with its working. Checks that don't earn their place come out.
Reading this for a rating agency or a diligence desk? Start with the methodology note →
How a threshold earns its number.
Every threshold in mortgage lending is inherited convention. A comp beyond a mile. Gross adjustments over 25%. A desk review that busts at 10%. Nobody can tell you where any of them came from.
Sphinx thresholds are built to be fitted to evidence: appraisals matched to what the properties actually sold for, with the market's movement over each hold removed. Onboarding runs that fit on your own archive first, so live review starts with thresholds set to your markets and your loan profile. The validation protocol was published before any data was examined, so the first result will be a result rather than a story.
The primary hypothesis, on record: appraisals whose submitted budget falls short of the finish level implied by their supporting comparables realise a lower market-adjusted sale price, relative to the ARV, than appraisals where the budget matches. One check, one test, decision rules fixed in advance.
flagged files overvalued more often than clean ones. →
The first study will report, for a cohort of exited loans, how often files were overvalued when a check fired against when it did not — by check, with sample sizes, nulls included. The figures on this panel are placeholders for the format, not results.
Run this on your own book →* Placeholder figures, illustrative of the format only. No calibration study has been run; the status page is the canonical record.
Every lender's exits sharpen thresholds for every other lender. Your files stay yours; only derived statistics cross the line. Read the validation protocol →
Your files stay yours. Only the evidence is pooled.
Never leaves your workspace Your identity, loan numbers, addresses, borrower names, the PDFs. Deletable on request, in full.
Pooled, de-identified Comparable grids, adjustments, ratings, values, outcomes, appraiser licence numbers, geography at census tract.
Never stored Contact details, borrower financials, credit information.
Why the address is held at all Calibration means comparing an appraised value to the price that property actually sold for, and the match needs an address. Deeds are public record. The address does the matching and then stops — what enters the pool is the outcome and a census tract. Transition loans are business-purpose credit, usually to an entity, which sits outside the consumer privacy regimes governing conventional lending; we build to those standards anyway.
Why the licence number enters the pool Cross-lender volume is the only way to estimate anything about an individual appraiser, so the licence number enters the pooled calibration set. No appraiser is named in any published output. No score is surfaced as a ranking, list, or roster. The coefficient adjusts a value; it does not describe a person. We don't send findings to appraisers, and we don't produce a shared roster of who to avoid — that would be a different product, and a worse one.
Encrypted at rest and in transit. Identity and calibration data held in separate systems.
Setup is calibration. Calibration starts with your closed loans.
Point Sphinx at appraisals from loans you've already closed. The checks run in minutes per file. For the loans that have exited you get back how far the ARVs ran from exit prices, which checks fired on the misses, and in which markets; the rest is reported in aggregate. Those results fit the thresholds to your book before live review turns on.
Live review switches on when you say, with thresholds already fitted to your markets and your loan profile. From then on every new appraisal returns a finding, and every exit sharpens the fit.
Archive pilot at no charge, on closed loans. Annual agreements for high-volume lenders and capital markets.
Point Sphinx at the archive: a folder of PDFs, or an inbox forward. Nothing to integrate.
ARVs against exit prices, which checks fired on the misses, in which markets. Thresholds fitted to your book.
Live review switches on when you say. Nothing in your process changes until it does.
Every new appraisal returns a finding in minutes. Every exit sharpens the fit.