Model Card — calibrated_bias_model (golden)¶
Generated by EthicLens. Format follows Mitchell et al. (2019).
Model details¶
- Owner: EquitaTech Systems
- Audit date: 2026-06-16
- Evaluation dataset:
synthetic-lending - Fairness status: ⚠️ FLAGGED
- Composite Bias Score: 0.689 (Medium Risk)
Intended use¶
Automated decisioning under fairness review.
Factors¶
Protected attributes evaluated: gender, race. Subgroups are compared against the highest-selection-rate (privileged) group.
Metrics¶
Disparate Impact (4/5ths rule), Statistical Parity Difference, and — when ground-truth labels are available — Equalized Odds, each reported with a bootstrap 95% confidence interval. A group is flagged only when its DI confidence interval lies entirely below 0.80.
| Group | Disparate Impact | 95% CI | Status |
|---|---|---|---|
| race:White | 0.950 | [0.89, 1.01] | OK |
| race:Black | 0.551 | [0.50, 0.60] | FLAG |
| race:Asian | 0.974 | [0.90, 1.06] | OK |
| gender:Male | 0.994 | [0.94, 1.05] | OK |
Ethical considerations¶
Disparate impact below 0.80 is evidence of adverse impact under the EEOC Uniform Guidelines. This card supports, but does not replace, human review and legal assessment.
Caveats & recommendations¶
- The composite score is a triage convenience, not a legal standard.
- Confidence intervals widen on small subgroups; treat wide intervals as "unknown".
- Re-audit after any retraining or data shift. See
LIMITATIONS.md.