RTL-STD-001 · Methodology note

Methodology note for rating agencies and diligence firms

What the standard is, what it explicitly is not, how a calibrated collateral exhibit could inform a collateral review, and what evidence exists now and at each milestone after.

1. Purpose of this note

This note is written for a structured finance or diligence analyst evaluating residential transition loan (RTL) collateral: rehab, gut renovation, and ground-up construction loans to investors, business-purpose, typically nine to twenty-four months. It describes what the Sphinx Standard (RTL-STD-001) is, what it is not, how a collateral finding produced against it is constructed, how such a finding could inform a collateral review once calibrated, and what evidence exists today and would exist at each milestone after. It makes no claim that has not been measured; where a figure would belong, the status page is the source.

2. The problem the standard addresses

The after-repair value (ARV) on an RTL appraisal is a forecast. The appraiser is handed the borrower's budget and asked what the property will be worth once the work is done, and answers by selecting comparable sales already finished to some standard. Conventional collateral review has no equivalent to this value, because a conventional appraisal values a house that already exists. There is no common review standard for ARV appraisals across the asset class, and no systematic measurement of whether ARVs held up at exit. The market is large enough for that to matter: the Urban Institute (April 2026) puts 2025 RTL origination above $85 billion, with more than $25 billion of it ground-up construction and more than $35 billion rehab. KBRA projects roughly $4.8 billion of RTL securitisation in 2026, about 73% of it rated.

The gap the standard is built around is a step nobody in the process is assigned: whether the submitted budget is large enough to reach the finish level of the comparables chosen to support the ARV. The appraiser values the finished house the scope describes; nobody prices the scope. That gap is arithmetic, computable from the appraisal and the budget alone, and it is the primary hypothesis of the validation protocol.

3. What the standard is

A published specification of 56 checks in 5 families (RTL-C comparable selection, RTL-A adjustments, RTL-V the after-repair value, RTL-S subject and file integrity, RTL-P appraiser), each with a stable ID, a statement of what it tests, the condition on which it fires, its output type (a flag, an interval contribution, or a dollar effect on value), and its provenance.
Provenance on every threshold. Industry Baseline: drawn from GSE guidance where documented and from common practice otherwise, threshold shown in full. Sphinx-Original: a check with no equivalent in any other review standard (20 of the 56); firing condition stated qualitatively, implementation parameters held as reference implementation. Calibrated: a threshold fitted to realised sale prices, with the sample size beside it. At publication, 0 thresholds carry that label.
Versioned from publication, with a governance process, a contributor committee with seats named by role, and a removal rule for checks that fail to predict outcomes. See governance.
An asset-class standard rather than a product feature. Check IDs carry the asset-class prefix so that a methodology can cite consistent with RTL-STD-001 v1.0, RTL-V-01 without naming a vendor. Sphinx maintains the standard and operates a reference implementation; the specification is open.

4. What the standard is not

Not an appraisal, not a valuation, and not a review appraisal under USPAP Standard 3. A finding is analytical review of a document: arithmetic, internal consistency, and comparison to public data. It produces an internal credit input, not an opinion of value, and is not signed by an appraiser because it is not one.
Not a communication to the appraiser. No output is transmitted to an appraiser, and no finding is a basis for requesting a revision to any report. Where a finding says what would clear a check, it refers only to the borrower's budget, scope, or the loan structure; it never refers to an adjustment, a comparable selection, or a value conclusion the appraiser would have to revise.
Not a credit decision. The standard outputs a corrected value and a confidence range. It does not compute or display an advance rate, LTV, or LTARV. Those are credit-box parameters that depend on lender policy, and RTL is typically sized on the lower of loan-to-cost and loan-to-ARV rather than on a single percentage of value.
Not a forecast. Nothing in a finding projects a future price. Hold duration and submarket volatility widen the confidence range; they never move the point estimate.
Not an appraiser rating. Appraiser-level effects, once estimable from pooled volume, adjust a value and are never surfaced as a ranking, roster, or comparative score in any output, including internal tools.

5. What a finding contains

A finding is designed as an exhibit filed in the loan package, readable by a diligence analyst years later with no access to Sphinx. Every exported finding carries, by value rather than by reference: the standard version at the time of the run; the threshold values actually applied; the extraction pipeline version; the market index provider and vintage; the effective date of the appraisal, distinct from the date the finding was produced; and the appraiser licence number. A specimen exhibit shows the format on illustrative data.

Corrections are shown as a waterfall from the appraised ARV: each dollar-effect check that fired, its provenance tag, the gross directional correction, the overlap suppressed between checks that respond to the same condition (shown even when it is zero), and the corrected point estimate. Interval contributions are shown separately from corrections. Confidence means confidence in the corrected value, not extraction accuracy; it falls when the evidence is thin, and low confidence is neither a flag nor a statement that the value is wrong. A check that could not run because a field could not be read cleanly says so rather than running on a guess.

6. How a calibrated collateral exhibit could inform a collateral review

Three uses, in order of how much evidence each requires.

As a consistent review criterion, today. A published, numbered, versioned check set applied identically to every file in a pool gives a reviewer a common vocabulary for collateral exceptions: which check fired, on what input, against which version. This requires no calibration; it requires only that the standard be public and stable, which it is.
As a directional correction with a stated range, once the primary hypothesis is confirmed. If the pre-registered study confirms that the budget-to-finish gap predicts realised ARV shortfall, a corrected value and its interval become a defensible input to a collateral haircut or a loan-level exception, with the evidence and its sample size cited on the exhibit. Until that study reports, a corrected value is a structured statement of what the file supports, not an outcome-tested prediction, and the exhibit says so.
As pool-level collateral evidence, once pooled calibration exists. With outcomes matched across contributors and aggregation thresholds cleared, the same exhibits support pool surveillance: the distribution of corrections by originator, market, and vintage, and the realised miss against the predicted one as loans exit. This is the use that bears on execution. It is also the furthest away, and this note does not claim it.

In every case the exhibit is an input to the reviewer's judgment, not a substitute for it.

7. What evidence exists, and what would exist by when

Evidence is sequenced by what has to be true before the next step can begin. No calendar dates are stated here. When a milestone has a date it appears on the status page, and this note is updated to match.

Milestone	What exists at that point	What gates it
Today	The published specification; a pre-registered validation protocol with decision rules fixed before any data is examined; a specimen exhibit on illustrative data. No calibrated thresholds, no outcome study, no contributing lenders. See status.	—
Pilot study 50 exited loans	A single pre-specified test of the primary hypothesis (RTL-V-01) against market-adjusted realised sale prices, with the effect size, its confidence interval, sensitivity to the cost lookup, scorer reliability, and the baseline model it has to beat. Published with nulls, whichever way it falls.	One contributing lender's archive of already-exited loans. Because the exits are historical, this step does not wait on new originations.
Expanded study 150 to 300 loans	Results for the Sphinx-original checks as a family with false-discovery control; exploratory effect sizes for the industry-baseline conventions; the first removals of checks that predict nothing.	A pilot result at or above the protocol's ambiguous threshold; additional archive volume.
First calibrated thresholds	Thresholds relabelled Calibrated on the standard with their sample sizes, per market where the evidence supports it, with a dated changelog entry.	Expanded-study confirmation; aggregation thresholds cleared for each published cell.
Pooled calibration	Cross-lender cells; appraiser-level effects estimated as value adjustments; realised-versus-predicted miss reported as live-reviewed loans exit, nine to twenty-four months after review.	Multiple contributing lenders; minimum distinct-contributor counts per cell.

8. Data handling relevant to a reviewer

Lender identity, loan number, property address, borrower identity (held as a per-tenant one-way hash), and the source PDF stay in the contributor's own tenant. What enters the pooled calibration set is the comparable grid, adjustments, condition and quality ratings, both opinions of value, the realised sale price, the appraiser licence number, dates, and geography at census tract. Identity and calibration data are held in physically separate systems. Ground-truth sale prices come from public deed records, not from any contributor's servicing data. No pooled statistic publishes until it clears a minimum observation count from a minimum number of distinct contributors.

9. How to cite

Cite the standard by document ID, version, and check ID: consistent with RTL-STD-001 v1.0, RTL-V-01. Findings cite the same identifiers and record the threshold values used. The full specification is at sphinx.credit/standard, the governance process at sphinx.credit/governance, and the validation protocol at sphinx.credit/validation. Questions about the method: standard@sphinx.credit.

Home The standard Governance Methodology note Validation protocol Status Sample finding FAQ Glossary Team About

RTL-STD-001 v1.0 · published · this page last updated 2 September 2026 · status

The Sphinx standard describes analytical review of an appraisal report. It is not an appraisal, not a valuation, and not a review appraisal under USPAP Standard 3. Outputs are internal analytical inputs for credit decisioning. They are not communicated to appraisers and are not a basis for requesting revision to any report.