Constitutional Oversight of Hybrid ML-RAG Credit Scoring Systems: Scalable Governance for Safe and Fair Financial Inclusion in Emerging Economies

Over 350 million adults in the West African UEMOA region use Mobile 
Money as their primary financial instrument yet possess no formal 
credit history. Traditional credit scoring models systematically 
exclude this population — not because they are untrustworthy, 
but because the infrastructure to read their data does not exist.

We present Fvundi, a hybrid ML-RAG credit intelligence system 
addressing this exclusion at three levels. First, a two-layer 
prediction-explanation architecture: XGBoost scores creditworthiness 
from Mobile Money transaction features, while a RAG pipeline 
(LlamaIndex + Cohere + Claude) generates faithful natural-language 
explanations grounded in SHAP feature attribution.

Second, we identify and formalize explanation faithfulness failures 
— instances where the RAG explanation layer misrepresents the 
underlying ML model's decisions, introducing demographic 
inconsistencies for identical credit scores. We propose four 
quantitative metrics: Feature Coverage Rate (FCR), Feature Rank 
Correlation (FRC), Demographic Consistency Score (DCS), and 
Hallucination Rate (HR).

Third, we introduce Constitutional Oversight: a three-layer 
governance framework combining a machine-readable Financial 
Constitution (human-authored ethical rules), Scalable Oversight 
(an AI Arbiter enforcing constitutional compliance at transaction 
speed), and Human-in-the-Loop control (triggered selectively for 
violations). This architecture satisfies BCEAO regulatory 
requirements while operating at the scale of millions of Mobile 
Money transactions per day.

Our central argument: safe AI credit scoring in emerging economies 
requires institutional design that keeps humans in the loop at the 
governance level while delegating verification to machine-speed 
oversight. Fvundi is designed to extend the banking system, 
not replace it.