Change Governance

Explanation

Change Governance checks that every change to the AI model or system passes through a controlled, staged release process and ships with a rollback plan, rather than being pushed straight to production. This matters because uncontrolled model changes are a leading cause of incidents — a new version, retrained weights, or altered prompts can degrade accuracy, fairness, or safety, and without staged promotion and a tested rollback the bank cannot recover quickly or demonstrate control to regulators. The metric and calculation method are both the Model Change Governance Gate, verifying that changes traverse the required staged gate. To implement it, route all model/system changes through a defined pipeline with promotion stages (for example dev to staging to canary to production), enforce gate approvals and pre-deployment validation, require an attached rollback plan before promotion, and log each change with its approvals, gate evidence, and rollback procedure for audit. The threshold is that 100% of changes pass the staged gate with a rollback plan. As an Immediate-priority control, any change that bypasses the gate or lacks a rollback plan constitutes a breach and should trigger blocking or reverting the change and investigating how the control was circumvented.