Rules and risk scores organize decisions; neither should hide the reason

Rules make explicit conditions

A rule may route a transaction to review when defined conditions are met. Rules are readable and can respond quickly to a known pattern, but too many overlapping rules become hard to maintain. Thresholds also age as customer behavior, products, attackers, and data quality change.

Scores combine evidence

A risk score can combine multiple signals into a ranked estimate used for approval, friction, review, or decline. The number is not a fact about a person. Its meaning depends on the target outcome, data, calibration, population, time window, and decision threshold.

The decision needs an explanation

Operators should know the main contributing reasons, data freshness, rule or model version, action taken, and available override or escalation path. Governance should track performance across relevant customer groups, drift, feedback labels, change approval, and unintended harm. This guide explains the concepts; it does not screen or score anyone.

Monitor decisions, not only model output

Track approval, review, decline, customer friction and confirmed outcomes across relevant groups and products. A stable score distribution can coexist with changing data quality or business behavior. Changes need version control, a documented owner and a rollback path. Explainability is an operating discipline, not a decorative reason label added after a decision.

Example: the score crosses a threshold

A transaction receives a high score because of velocity, a new beneficiary and a device change. The reviewer should see those reasons and the underlying evidence, not only “92.” If the device field is missing or the customer recently changed phones, that context may affect the next action. The score organizes attention; it does not decide identity, intent or legal responsibility by itself.

Govern the change that moves the line

When a threshold or feature changes, record the rationale, expected effect, approver and monitoring period. Compare confirmed outcomes and customer friction with the prior version, then retain a rollback path. A later reviewer should be able to reconstruct which logic applied to a decision. Versioned reasoning is part of explainability.