Reasonable objections

No, and the distinction matters legally. An appraisal review under USPAP Standard 3 is a licensed appraiser forming an opinion about another appraiser's work, and it has to be performed and signed as such. Sphinx is analytical review of a document — arithmetic, consistency, and comparison against outcome data. It produces an internal credit input, not an opinion of value, and it isn't signed by an appraiser because it isn't one.

The appraiser shouldn't see the Sphinx report — it's your internal analysis. If a finding concerns you, you might adjust loan terms, request reconsideration through your AMC with fresh market data, or order a new appraisal. What you shouldn't do is send the report to the appraiser or use it to pressure a value change.

No. The appraiser has the scope of work — they're given it. What they're asked to do is value the finished house it describes, using comparable sales. What nobody asks anyone to do is price the scope: to check whether that budget can produce the finish level those comparables represent. That isn't in the appraiser's engagement, and it isn't in the underwriter's or the AMC's either. Sphinx performs that step. It doesn't second-guess the valuation.

You can, and some lenders will. But consider what you're actually building: a set of rules that fire on patterns in appraisals. The hard part isn't the AI — it's knowing which patterns matter. That requires outcome data you don't have yet, and years of exits to calibrate against.

Sphinx thresholds are built to be replaced by evidence. Every check is published with its provenance — whether the number is a stated industry baseline or fitted to realised sale prices — and when a threshold moves, the change publishes with it. No black box. Every exit across every contributor sharpens the calibration, which means the system improves faster than any single lender's data could support.

Build in-house if you want full control. Use Sphinx if you want a calibration set larger than your own book will ever be, continuously improving, and defensible to regulators and capital markets counterparties.

It runs in minutes on receipt, which is days before the file clears credit, so it sits entirely inside existing slack. If anything it pulls a problem forward: the alternative is discovering a value issue at the closing table, or discovering it eleven months later at payoff.

The same way you underwrite a new borrower without knowing if they'll default. Historical data is what teaches you which patterns correlate with problems — scope shortfalls, adjustment clusters, thin comp sets. Those patterns show up in new appraisals before the loan closes. We're not predicting the sale price; we're identifying the characteristics the outcome study is designed to test as predictors of a miss. Every exit adds another observation, and the thresholds are refitted as the sample grows.

We don't claim distant comps cause misses — we'll claim that appraisals with distant comps have historically missed more often, once that's been measured. The validation protocol sets the thresholds where correlation with realised outcomes becomes statistically meaningful. Until it's run, no threshold on the site is labelled calibrated: each carries its provenance — industry baseline or Sphinx-original — and says so. The protocol is published; where it stands is on the status page.

There's plenty to compare — new construction sells, and those sales carry quality ratings. Ground-up is arguably the cleaner test, because plans and specs are a precise document and there's no hidden condition behind drywall. What changes is the risk profile: longer timelines mean more market movement, and land comps are thinner than improved comps.

Fannie's CU scores appraisals for conforming mortgages — owner-occupied purchases with conventional financing. It doesn't touch transition loans, doesn't see rehab scopes or ARVs, and wasn't designed for post-renovation outcomes. Sphinx is purpose-built for RTL: fix-and-flip and ground-up construction where the question is "will this property be worth what the appraisal says after the work is done?" Different loan type, different risk, different tool.

Diligence verification confirms the file matches the tape — the appraisal exists, it says $455,000, the deed is right. It doesn't ask whether $455,000 was supportable. That's a different question, it needs a published standard and matched outcome data to answer, and it isn't what a per-loan verification workflow produces. Sphinx sits alongside diligence, not in place of it. The methodology note is written for that desk.

That's what the pilot solves. Your archive — ideally loans that have already exited — seeds the model for your markets before live review turns on. If you're in a market where pooled coverage is thin, your own historical data still calibrates the thresholds to your book. And every file you contribute improves coverage for that market going forward.

How firmly the file supports the supported value, shown out of 100 with a band (80 and above high, 60 to 79 moderate, below 60 low). Not a probability, and not our reading of the document. It falls when the underlying evidence is thin — few comparables, stale sales, heavy adjustments, a long project timeline, a volatile submarket — and when a field a check depended on could not be read cleanly.

Low confidence is not a flag. A clean, well-prepared appraisal in a thin market can score 60 with nothing wrong in it. Confidence and flags are independent.

Low confidence doesn't mean the value is wrong. It means the file supports its number less firmly — and that's worth knowing before you lend against it.

Because RTL was too small to matter and too fragmented to standardize. That's changing. Volume is scaling, the secondary market is emerging, and institutional capital wants consistent review criteria before buying pools. Unlike conventional lending — where FNMA guidelines have been embedded for decades — RTL has no standard. That's a window: establish it while the market is growing and you become infrastructure. Wait, and regulators mandate something worse. The time to define the standard is before it's needed everywhere, not after.

Because the alternative is that nobody ever tells you how your ARVs performed. Files stay in your own workspace, your identity never enters the pool, and everything is deletable on request. What crosses into the shared set is comp grids, adjustments, and outcomes — de-identified and aggregated.

Because every threshold in the industry is inherited convention — nobody can source where any of them came from. Sphinx thresholds are built to be fitted to your archive during onboarding: your markets, your project types, your loan profile. As contributors' loans exit, the pooled calibration is refitted for everyone, and every move publishes with its sample size.

It lands where appraisals already land — a per-tenant email address or the folder your AMC delivers into. The checks run on receipt, the finding attaches to the loan in your LOS or lands in the underwriter's queue, and the whole thing finishes days before the file clears credit. When that loan exits in 9–24 months, the sale price becomes a new calibration observation. Every file reviewed live today makes the thresholds sharper tomorrow.

$88 per appraisal reviewed, with no minimums and no platform fee. Founding partners pay $44 — half price for the first year, locked in.

High-volume lenders and capital markets: annual agreements with volume pricing. If you're an aggregator, warehouse lender, or due diligence team reviewing across originators, reach out — different model, same standard.

The 30-day pilot runs on your archive at no charge. You see the value before you pay anything.

Archive means appraisals from loans originated before your start date — the history you already have. The pilot runs on these at no charge to prove value on your own book. Exited loans are the most useful because we can compare the appraised ARV to what the property actually sold for. Live means new appraisals received after you go live — these are billed per file.

There's nothing to integrate in the traditional sense. Point us at the folder where your AMC delivers, or forward to a dedicated inbox. No API work, no LOS changes, no IT project. Most lenders are live within a week of signing.

Every production finding will carry a unique ID and QR code, so counterparties — loan buyers, auditors, warehouse lenders — can confirm at sphinx.credit/v/[finding-id] that the finding exists, see its key metadata (property, date, outcome), and detect tampering. If someone hands you a finding claiming a "Clear" outcome, you will be able to confirm it in seconds. Today the endpoint is a static demo behind the specimen finding; live verification switches on with the first production finding, and the status page says which it is.