METHODOLOGY v2.0 · REVIEWED SEP 16, 2026
A metric is evidence. Not a verdict.
Required inputs, methods, assumptions and stopping conditions remain attached to every result.
Raw and common-weight bureau-stratified rates are calculated from exported synthetic rows. All attempted runs are retained. No regression, BISG, causal inference or legal determination is implemented.
Data qualification
Establish population, dictionary, missingness, coverage, explicit exclusions, timestamps, provenance and system versions before analysis. Missing APR blocks pricing analysis.
Outcome screening
AIR is comparison-group approval rate divided by reference-group approval rate. Rate difference subtracts those rates. Show raw counts and denominators. A configured 0.80 screening threshold is not a legal verdict. Small samples, population composition and missing proxies limit interpretation.
Conditional analysis
Ask whether a difference persists after relevant risk factors. Regression and stratification require justified controls, overlapping populations and independent methodology review. The current method uses explicit bureau strata and pooled weights; it is not regression. Sparse strata are named and omitted. The full workpaper includes counts, weights and definitions; do not infer causation.
Comparative review
Document file selection and compare risk characteristics, policy exceptions and human overrides. The demo pair is purposively selected, not statistically matched. Similar observed fields do not prove equivalent risk.
Adverse-action reason integrity
Require the notice, authoritative scored-factor and principal-factor lineage, mapper version and overrides. A contradiction requires affirmative evidence. Missing lineage means Cannot verify, not Inconsistent.
Retesting
Preserve prior dataset, system, configuration and method. Identify changed variables and population differences. Both a persistent outcome screen and a corrected reason-map replay are demonstrated. Original results remain available. Closing a retest-required finding requires a supported linked retest; an accepted-risk exception is a separate disposition.
Protected attributes and proxies
Use appropriately supplied institution data with provenance and uncertainty. No proxy estimation is implemented. First-line sharing hides cohort details in the interface; the canonical synthetic demonstration has no server-side authorization. The separate managed workspace enforces analyst/reviewer actions but does not restrict cohort visibility between those roles.
Human interpretation
Neither disparity nor a null result establishes compliance. Screening evidence supports qualified investigation. No automated legal conclusion or continuous monitoring is provided.
Regulatory source register
Federal Reserve, OCC, FDIC · issued 2026-04-17 · reviewed 2026-09-16
Supersedes SR 11-7; letter most relevant to Federal Reserve-regulated organizations over $30 billion; applicability requires institutional assessmentCFPB · issued 2026-04-22 · reviewed 2026-09-16
Amended provisions; qualified interpretation required