Model Versioning and Change Control

Objective

Ensure that every model promoted to production is uniquely versioned, traceable to its training data, evaluation results, and approval, and subject to controlled change-management with rollback capability.

Applicability

AI types

LLM, Agentic AI, Traditional ML, Computer Vision, Generative AI, Multi-modal, Recommender

Deployment models

Self-hosted, Hybrid, Edge, Embedded

Lifecycle stages

Training, Evaluation & Testing, Deployment, Operation & Monitoring

Risk domains

Model, Infrastructure

Regulatory regimes

EU AI Act, ISO 42001, NIST AI RMF, Banking, Healthcare

Company size

MidMarket, Enterprise

Rationale

Model changes can introduce regressions invisible to traditional change-control reviewers (a 2% accuracy drop on a protected subgroup, drift in calibration, new failure modes). Without versioning, organizations cannot reproduce a specific request's behavior, audit decision outcomes, or roll back when a regression is detected. EU AI Act Article 17 (quality management system) and ISO/IEC 42001 clause 8.3 require versioned, controlled change.

Control narrative

Every production model has a unique version identifier traceable to: training data snapshot (provenance records per AI-CTRL-002), training configuration, evaluation results (including fairness and adversarial robustness), approving authority, and deployment date. Model promotion to production follows a documented change-management flow with separation of duties between author and approver, mandatory pre-production evaluation gate (AI-CTRL-017), and rollback plan. Material change criteria are defined (e.g., > X% drift on key metric, new training data category, new tool/permission for agentic systems) and trigger re-evaluation. Customer-impacting model changes follow a notification cadence appropriate to the use case (e.g., enterprise tier of foundation models communicate model version with each response).

Test of Design

Procedures

1. Obtain the Model Change Control standard and confirm scope, versioning scheme, and required traceability fields.
2. Confirm separation of duties between model author and approver.
3. Confirm pre-production evaluation gate is required prior to promotion.
4. Confirm material-change criteria are defined and trigger re-evaluation.
5. Confirm rollback capability is required as part of the deployment plan.
6. Confirm customer notification policy for model changes where applicable.

Inquiries

• Who approves model promotion to production?
• How is rollback tested? How fast is it?
• How are silent model changes from third-party providers detected (e.g., a SaaS LLM's underlying model changes)?
• How are A/B or canary deployments governed?

Inspections

• Model Change Control standard.
• Model registry sample showing versioning, lineage, and approvals.
• Sample change records for sampled model promotions.
• Rollback procedure documentation and test records.

Test of Operating Effectiveness

Procedures

1. From the inventory, sample production models promoted in the audit period.
2. For each sampled promotion, inspect the model registry entry and confirm: unique version, training-data lineage, evaluation results attached, approver identity, deployment date.
3. Confirm separation of duties was enforced (author ≠ approver).
4. Confirm pre-production evaluation results meet the documented thresholds.
5. For any production rollback in the audit period, confirm the rollback procedure was followed and root cause was investigated.
6. Inspect material-change re-evaluation records for any qualifying change.
7. For SaaS-foundation-model dependents, confirm vendor-notified model changes were reviewed and re-evaluated.

Sample-size guidance

Reperformance

• For 1 sampled model, independently re-run a subset of the documented evaluation suite on the deployed version and compare to recorded results.

Evidence requirements

Required

• Model Change Control standard PDF/Word · At fieldwork
• Model registry export CSV/JSON · At fieldwork
• Change records for sampled promotions System export / PDF · Per sample
• Pre-production evaluation reports PDF / Notebook · Per sample

Supporting

• Rollback test records PDF / system export · Annual
• Vendor model change notifications and internal responses Email / ticket export · Per occurrence

Retention: Lifetime of the model + 7 years after retirement (regulated); + 3 years (otherwise)

Framework mappings

Related controls

• AI-CTRL-002

  Training Data Provenance and Lineage

• AI-CTRL-017

  Pre-Production AI Evaluation Gates

• AI-CTRL-018

  AI System Decommissioning

References

• EU AI Act Article 17 — Quality management system
• ISO/IEC 42001:2023, clause 8.3
• NIST AI RMF — MANAGE-2 Risk Management Actions