Credit unions Model Risk Version Control | Moring
Version control gives credit unions a defensible answer to a deceptively simple model risk question: which model, prompt, and configuration produced this member outcome, and was it the version that was validated. It covers internal artifacts and, through contract terms, the vendor models that update on someone else's schedule.
Short answer
Version control gives credit unions a defensible answer to a deceptively simple model risk question: which model, prompt, and configuration produced this member outcome, and was it the version that was validated. It covers internal artifacts and, through contract terms, the vendor models that update on someone else's schedule.
How it works
Version control in credit union model risk is the practice of recording, for every AI system in member-facing or decision-relevant use, exactly which model version, prompt or template set, configuration, and data snapshot produced each period of behavior. Model governance guidance describes this as a core component of the end-to-end discipline, alongside documentation, back-testing, and monitoring, and the reason is reconstructability: when a member outcome is questioned months later, the institution must be able to say which version acted, and whether that version was the one that was validated and approved.
For internally built or configured systems, the mechanics are conventional: version identifiers in the inventory, a validation record bound to a specific version, and a rule that no version reaches production without its validation record. The subtlety is scope. Prompts, retrieval collections, tool permissions, and thresholds all change model behavior, so versioning must cover the full configuration, not just model weights. Two versions that differ only in a retrieval source can produce materially different lending support outputs, and a version history that misses the difference will mislead exactly when it is consulted.
Vendor-delivered AI makes version control contractual. Core vendors and fintech partners update models on their own schedules, sometimes silently, which means the version producing member outcomes can change without an institutional decision. The contract remedies are notice of material model changes, the ability to pin or defer versions, and change logs the institution can obtain. When a version change arrives, the credit union's model risk response is revalidation proportionate to the change: a material model swap triggers testing and approval before member exposure, a minor patch may need only monitoring confirmation. Revised interagency guidance issued as SR 26-2 in April 2026 is risk-based and scaled to the institution, and that proportionality is how a small institution applies it.
The supervisory logic is reconstruction. Technology-neutral expectations apply to AI as to any system affecting members, and fair lending exposure attaches to lending decisions whether a human or an algorithm made them, so the institution must be able to reconstruct which inputs and which model version produced a contested decision. FFIEC expectations on records and change management supply the familiar shape: controlled change, documented approval, and the ability to demonstrate what was in production at a given time. Institutions with vendors exposed to European markets should note the EU AI Act's technical documentation expectations point in the same direction.
Operationally, version control earns its keep in three moments: validation, where the record binds an approved version to its tested behavior; incident response, where the version history answers what changed before the anomaly; and examination, where the institution demonstrates that production matched approval. Metrics worth tracking include the share of production systems whose current version maps to a current validation record, and the median time between a vendor model change and the institution's completed revalidation.
Frequently asked questions
What does version control cover in credit union model risk?
Everything that changes model behavior: model versions, prompts and templates, retrieval sources, tool permissions, and thresholds, bound to validation records and production dates. For a credit union the purpose is reconstructability, being able to state which version produced a member outcome and whether it was the validated one. Vendor systems extend this through contract terms covering change notice, version pinning, and access to change logs.
How does version control work for vendor AI models?
Contractually. The credit union negotiates notice of material model changes, rights to pin or defer versions, and obtainable change logs, then responds proportionately: material swaps trigger revalidation before member exposure, minor updates may need monitoring confirmation only. Without these terms, the version producing member outcomes can change silently, leaving the institution unable to say what was in production when a contested decision occurred.
What triggers revalidation when a version changes?
Materiality of the change, judged against the use case's risk rating. A new underlying model in a lending support tool warrants testing and approval before member exposure; a patch to an internal drafting assistant may need only monitoring confirmation. The rule worth writing down is that no version reaches member-facing production without a validation record matched to it, with proportionate depth set by the credit union's risk-based framework.
Why is version control important for exams and disputes?
Because both turn on reconstruction. Reviewers and auditors ask which system produced a decision and whether its controls were current, and fair lending exposure attaches to algorithmic lending decisions exactly as to human ones. A version history that binds production dates to validated versions lets the credit union answer precisely, while gaps in the history turn a defensible decision into an unanswerable one.