Objectif
Guarantee complete traceability for AI systems by enforcing unique identifiers, metadata capture, and lifecycle status tracking across all environments.
Champ d'application
Includes all AI models, prompts, datasets, feature pipelines, evaluation suites, monitoring configurations, and third-party components used by the organization.
- Production, staging, and development model versions
- Datasets and labeling assets feeding model training
- Evaluation artifacts (test suites, prompts, metrics)
- External models embedded into internal workflows
DĂ©finitions
- Lifecycle ID: Canonical identifier linking model, datasets, and documentation across environments.
- Inventory Record: Source of truth entry capturing metadata, ownership, risk tier, and control evidence.
- Status Tag: Current lifecycle state (ideation, development, validation, production, retired).
Politique
Every AI asset must be registered before testing or deployment. Inventory records must remain accurate, reflecting dataset versions, release approvals, monitoring owners, and regulatory mappings. Decommissioned assets are retained according to data retention rules.
Rôles et responsabilités
Model Registry Owner maintains the platform and audits completeness. Model Owners keep metadata current. Compliance relies on the inventory for regulatory filings. Security ensures access controls are enforced around sensitive entries.
Procédures
Inventory maintenance includes:
- Registration at project kickoff with assigned lifecycle ID and business owner.
- Metadata updates capturing training datasets, git commits, hyperparameters, and evaluation results.
- Linkage of approval tickets, QA reports, monitoring dashboards, and incidents.
- Access control enforcement and change logging.
- Archival checklist when retiring models, including final metrics, retention requirements, and successor models.
Exceptions
Temporary prototypes may defer registration for up to 14 days. Extended prototypes must enter the inventory even if they never ship to production.
Fréquence de révision
Monthly audits validate completeness, timeliness of updates, and linkage to control evidence. Findings route to Model Owners for remediation within 10 business days.
Références
- EU AI Act Article 11 (Technical documentation)
- ISO/IEC 42001:2023 Clause 7 (Support)
- Internal documents: Model Registry Handbook, Documentation & Traceability Policy, Data Retention Standard