Consumer underwriting Q1–Q2 2026

{
  "version": "Stratified Screen v2.0",
  "formulas": {
    "rate": "approved / eligible",
    "AIR": "comparison rate / reference rate; undefined if reference rate = 0",
    "difference": "comparison rate - reference rate",
    "standardization": "sum pooled eligible stratum weight × cohort stratum rate; common weights sum to 1 over qualified strata",
    "uncertainty": "Wilson 95% intervals for raw rates with z=1.96; no adjusted interval, p-value or significance decision"
  },
  "transformations": [
    "Filter decision date, channel, model version",
    "Sort decision timestamp then ID",
    "Exclude non-final outcomes",
    "Retain first final application per applicant in selected scope",
    "Exclude missing proxy labels",
    "Restrict standardized comparison to bands meeting minimum per cohort"
  ],
  "assumptions": [
    "Single product and policy version; changes require new scope",
    "Proxy assignments are uncertain, not factual protected classes",
    "Within-scope applicant deduplication does not prove independence across all reviews",
    "Bureau bands are chosen for this synthetic example, not certified legitimate controls",
    "Multiple exploratory runs are retained; no multiplicity correction or confirmatory inference"
  ],
  "unsupported": [
    "Regression, marginal effects, SMD, equalized odds, causal attribution",
    "Pricing, continuous or ranked outcomes",
    "Joint applicants, multiple-offer data, unseen outcomes",
    "Proxy estimation or legal sufficiency of reasons"
  ],
  "fairness": "Descriptive outcome parity is one diagnostic, not a universal fairness definition. Equal outcomes can coexist with problematic treatment; unequal outcomes alone do not establish unlawful treatment. Scoring, policy, override and reason mapping are separate layers."
}