Public bias audit · composite × protected class·FHA-safe by design · audited quarterly

Composite × Protected-Class Correlation Audit

Quarterly statistical check that ValiFit composite scores do not systematically correlate with Census protected-class indicators. Methodology + threshold are public; the dataset is published as Schema.org JSON-LD for external verification.

v1.1 (FHA layer, locked 2026-06-25) adds a race/ethnicity-scoped interpretation step on top of the v1 correlation matrix: the 4/5ths rule (UGESP, 29 CFR § 1607.4) applied analogously to housing per HUD guidance on 24 CFR § 100.500. Race + national origin proxies (B02001 / B03003) are filtered to the FHA-protected classes enumerated in 42 USC § 3604. The v1.1 layer is scaffolded; the next quarterly audit will publish v1.1 outputs alongside v1.

Last run: 2026-05-16

8 anomalyies flagged

Tracts

24,306

Pairs

Threshold

|r| > 0.3

Anomalies

Composite × Protected-Class proxy correlation matrix

Composite	Proxy	Pearson r	Spearman	n	Status
education	pct_white_alone	-0.004	-0.015	17,944	✓
education	pct_black_alone	0.031	0.044	17,944	✓
education	pct_native_alone	0.033	0.031	17,944	✓
education	pct_asian_alone	-0.077	-0.012	17,944	✓
education	pct_pacific_alone	-0.033	0.021	17,944	✓
education	pct_hispanic	-0.014	0.016	17,944	✓
education	pct_disabled	0.048	0.065	17,944	✓
education	pct_under_18	-0.003	-0.020	17,944	✓
education	pct_age_65_plus	0.031	0.081	17,944	✓
safety	pct_white_alone	-0.167	-0.289	17,944	✓
safety	pct_black_alone	0.143	0.359	17,944	⚠
safety	pct_native_alone	-0.017	0.232	17,944	✓
safety	pct_asian_alone	0.109	0.323	17,944	⚠
safety	pct_pacific_alone	0.006	0.198	17,944	✓
safety	pct_hispanic	0.077	0.254	17,944	✓
safety	pct_disabled	-0.058	-0.033	17,944	✓
safety	pct_under_18	0.020	0.043	17,944	✓
safety	pct_age_65_plus	-0.055	-0.043	17,944	✓
health	pct_white_alone	-0.215	-0.268	17,944	✓
health	pct_black_alone	0.137	0.339	17,944	⚠
health	pct_native_alone	-0.022	0.367	17,944	⚠
health	pct_asian_alone	0.178	0.449	17,944	⚠
health	pct_pacific_alone	0.030	0.515	17,944	⚠
health	pct_hispanic	0.167	0.305	17,944	⚠
health	pct_disabled	-0.057	-0.056	17,944	✓
health	pct_under_18	-0.009	-0.002	17,944	✓
health	pct_age_65_plus	-0.075	-0.032	17,944	✓
appreciation	pct_white_alone	0.085	0.137	17,944	✓
appreciation	pct_black_alone	-0.132	-0.115	17,944	✓
appreciation	pct_native_alone	0.014	0.074	17,944	✓
appreciation	pct_asian_alone	0.128	0.095	17,944	✓
appreciation	pct_pacific_alone	0.014	0.197	17,944	✓
appreciation	pct_hispanic	-0.058	-0.042	17,944	✓
appreciation	pct_disabled	-0.113	-0.099	17,944	✓
appreciation	pct_under_18	-0.001	-0.002	17,944	✓
appreciation	pct_age_65_plus	-0.005	0.000	17,944	✓
overall	pct_white_alone	-0.176	-0.255	17,944	✓
overall	pct_black_alone	0.147	0.311	17,944	⚠
overall	pct_native_alone	0.002	0.193	17,944	✓
overall	pct_asian_alone	0.084	0.250	17,944	✓
overall	pct_pacific_alone	-0.006	0.136	17,944	✓
overall	pct_hispanic	0.100	0.215	17,944	✓
overall	pct_disabled	-0.022	0.014	17,944	✓
overall	pct_under_18	-0.002	0.001	17,944	✓
overall	pct_age_65_plus	-0.035	0.013	17,944	✓

Flagged correlations

These pairs exceeded the |r| > 0.3 threshold. Per the Public Bias Audit charter, product review must address each anomaly before the next quarterly audit closes.

safety ↔ pct_black_alone: r = 0.143 (n=17944)
safety ↔ pct_asian_alone: r = 0.109 (n=17944)
health ↔ pct_black_alone: r = 0.137 (n=17944)
health ↔ pct_native_alone: r = -0.022 (n=17944)
health ↔ pct_asian_alone: r = 0.178 (n=17944)
health ↔ pct_pacific_alone: r = 0.030 (n=17944)
health ↔ pct_hispanic: r = 0.167 (n=17944)
overall ↔ pct_black_alone: r = 0.147 (n=17944)

Methodology

Sample. 1% deterministic stratified sample of US Census tracts (capped at 5,000 tracts). Reproducible across runs via deterministic ordering on tract_id hash.
Composites. 8 published composite scores (Budget, Commute, Appreciation, School Investment, Environmental Safety, Safety Infrastructure, Health Infrastructure, Livability).
Proxies. 6 Census protected-class indicators at tract level: % Black, % Hispanic, % Asian, % age 65+, % under 18, % with disability.
Statistics. Pearson and Spearman correlations on aligned (composite, proxy) tract-level data, n ≥ 30 per cell.
Threshold.|Pearson r| > 0.30 = flagged. Anomaly count > 0 ⇒ product review required.
Publication. Full matrix + dataset JSON-LD on this page. Audit history retained in valifit.bias_audits.

Aligns with

HUD Disparate Impact Standard (24 CFR 100.500)
NIST AI Risk Management Framework
EU AI Act Article 27 — Post-deployment monitoring of high-risk AI systems

Charter All charters