A credit model where every
decision has a reason

Per-decision Explainability

Explanation at the single-application level - not just global feature importance.

• SHAP values on every decision
• LIME for spot checks
• Decision reconstructed in 95%+ of cases
• Explanation stored in the audit trail

Adverse Action Reasons

Human-readable decline reasons - not an error code.

• ECOA / Reg B reason codes
• FCRA-aligned adverse action notice
• Wording that does not leak the model
• Explanation consistent with the decision

Fair Lending & Bias

Monitor disparate impact before the regulator finds it.

• Disparate impact testing
• Proxy-feature detection
• Sound reject inference
• Periodic reports to the credit committee

Model Cards & Documentation

Documentation ready for a model committee and an examiner.

• Model Card per Mitchell et al. 2019
• Training data, limitations, intended use
• Performance by population segment
• Versioning and approval trail

Champion / Challenger

The model updates under control - it does not sit frozen for two years.

• A/B between models in production
• Drift detection (PSI, KS)
• Documented rollback
• Model registry (MLflow)

Core / LOS integration

The decision returns to the LOS / core in real time, with its explanation attached.

• Low-latency decision API
• Reason codes inside the payload
• Documented manual override
• Feature snapshot retained per decision

US: ECOA / Reg B, FCRA, SR 11-7

Adverse action notices under ECOA Reg B and FCRA, CFPB UDAAP, and Federal Reserve SR 11-7 model risk management documentation for the scoring model.

EU: AI Act Article 13 + GDPR

Credit scoring is high-risk under the EU AI Act: transparency (Art. 13), human oversight (Art. 14), accuracy (Art. 15), plus GDPR Art. 22 right to an explanation of automated decisions.

Israel: Bank of Israel

For Israeli institutions: Bank of Israel model risk directives, the explanation-to-customer principle, and privacy-law data retention.

Explainability works

What the industry reports: Credit teams using SHAP / LIME reconstruct the model's decision in 95%+ of cases (FCA + EBA papers).

Where we come in: Retro-fitting SHAP onto an existing XGBoost model - without replacing what already runs in production.

Reject inference improves accuracy

What the industry reports: Handling rejected applicants properly reduces sample bias and improves model stability over time.

Where we come in: Designing a transparent, documented reject-inference methodology - not a hidden "patch".

Bias found before the regulator

What the industry reports: Disparate impact is usually found only after a complaint - when remediation is already expensive.

Where we come in: Continuous fair-lending monitoring inside the pipeline, with alerts and committee-ready reports.