Explainability is not optional in regulated lending

Every credit decision Mansa produces is logged with the input data, the model version, the feature attribution, the risk officer who reviewed it, and any override that followed. This is not because we like ceremony. It is because regulators and credit committees ask hard questions, and the answer "the model said so" is not acceptable.

A score on its own is just a number. What gives the score weight is the audit trail behind it: which signals contributed how much, with what confidence, against what historical data, and what the human reviewer concluded after looking at it. That assembly is what makes the decision defensible months later, when the file is reopened.

Compliance is not a wrapper around a model. Compliance is a shape the model has to be designed for from day one — in the data lineage, in the feature attribution, in the override flow. Adding it later does not work.