Skip to content

1. Model Governance

Purpose

Guidelines for managing AI models throughout their lifecycle: from development to production and retirement.

1. Purpose

This module defines how we manage AI models throughout their lifecycle: from development to production and eventual retirement. Good model governance ensures traceability, controllability and safe releases.

2. Core Principles

Every Model Has an Owner

  • Every AI solution has one designated Tech Lead responsible for technical quality.
  • The owner is the point of contact for incidents, updates and decommissioning.

Everything Is Version-Controlled

  • Model weights, configurations and System Prompts are in version control.
  • Changes are traceable: who changed what and when?

No Change Without Review

  • Changes to production models require review by at least one other team member.
  • For High Risk: Guardian review mandatory.

3. Model Registry

A central location where all models are registered with their metadata.

Minimum Metadata per Model

Field Description Example
Model ID Unique identification invoice-classifier-v2.1
Version Semantic version or hash 2.1.0 or abc123
Status Development / Staging / Production / Deprecated Production
Owner Responsible person/team Team Finance AI
Creation date When trained/deployed 2026-01-15
Data source version Which data used for training invoices-2025-q4
System Prompt Link to prompt/config version prompts/invoice-v2.1.yaml
Validation Report Link to accompanying evidence reports/invoice-v2.1.md
Risk level Classification according to EU AI Act Limited

Implementation Options

Option Suitable for Complexity
Spreadsheet/Wiki Starting teams, few models Low
Git repository with YAML Engineering teams Medium
Experiment tracking platform Mature MLOps environment, many models High

4. Approval Workflow

Standard Flow (Limited Risk)

[Development] → [Code Review] → [Staging Test] → [Gate Review] → [Production]
  • Code Review: At least one peer review
  • Staging Test: Golden Set test on staging environment
  • Gate Review: Validation Report meets Evidence Standards

Extended Flow (High Risk)

[Development] → [Code Review] → [Guardian Review] → [Staging Test] → [Fairness Check] → [Gate Review] → [Phased Rollout] → [Production]
  • Guardian Review: Independent assessment against Hard Boundaries
  • Fairness Check: Quantitative bias analysis
  • Phased Rollout: Start with limited user group, monitor, then full rollout

5. Model Lifecycle

Phase Characteristics Actions
Development Experiments, prototypes No production data, no external users
Staging Candidate for production Full Golden Set test, review
Production Live, actively used Monitoring, incident procedure active
Deprecated Being phased out No new users, migration plan active
Retired No longer available Archiving, documentation preserved

6. Change Management

Types of Changes

Type Example Required Approval
Configuration change Temperature from 0.7 to 0.5 Peer review
Prompt change Rewriting instruction Peer review + regression test
Model version update New base model (e.g. GPT-4 → GPT-5) Full Gate Review
Data source change Coupling new knowledge base Guardian review (High Risk)

Rollback Procedure

  • Every production release has a documented rollback plan.
  • Rollback must be executable within 30 minutes.
  • After rollback: incident analysis and documentation.

7. Model Governance Checklist

Model Governance Checklist

  • Model registry is set up and up to date
  • All production models have an owner
  • Approval workflow is documented and followed
  • Change management is set up with rollback procedure
  • Models are linked to Validation Reports