Skip to main content

Versioning & Lifecycle

AI Products must declare versions and lifecycle states to ensure clarity, reproducibility, and controlled evolution.
Without explicit versioning, consumers cannot trust outputs or reproduce results, and governance becomes impossible.

Why Versioning & Lifecycle Matter​

  • Reproducibility → Consumers must reproduce results tied to a specific version.
  • Traceability → Versioned history supports audits, compliance, and accountability.
  • Governance → Policies may apply differently to development vs production versions.
  • Evolution → Continuous learning and retraining require structured lifecycle transitions.

Versioning Requirements​

  • Semantic Versioning → Major.Minor.Patch format (e.g., 2.3.1).

    • Major → backward-incompatible changes (e.g., new input schema).
    • Minor → backward-compatible enhancements (e.g., new feature).
    • Patch → fixes or optimizations without interface changes.

  • Retraining Versions

    • Retrained instances must declare version metadata (date, dataset ID, retraining reason).
    • Semantic version increments should reflect retraining impact.

  • Metadata Linkage

    • Versions must link to training data versions, model artifacts, and governance approvals.

Lifecycle States​

Each AI Product must declare its lifecycle state:

  1. Experimental

    • Early-stage, for testing purposes only.
    • Not for production use.

  2. Development

    • Functional but incomplete.
    • May lack full monitoring, governance, or compliance.

  3. Production

    • Fully governed, monitored, and compliant.
    • Discoverable in enterprise catalogs or marketplaces.

  4. Deprecated

    • Actively discouraged; support phase-out underway.
    • Consumers notified of migration paths.

  5. Retired

    • No longer available or supported.
    • Archived for audit and lineage purposes.

Governance Integration​

  • Lifecycle state must be part of the catalog/marketplace entry (see Discoverability).
  • Retired versions must remain archived and auditable.
  • Deprecation requires clear communication to consumers.

Example​

LLM Product v2.4.0

  • Lifecycle State: Production.
  • Training Data Version: transactions-v2024.03.
  • Retraining Metadata: Triggered by drift in March 2024, approved by compliance.
  • Next State Transition: Planned deprecation in Q1 2026.

Summary​

  • AI Products must declare semantic versions and lifecycle states.
  • Retraining, governance, and metadata must link to specific versions.
  • Lifecycle transitions must be transparent, communicated, and auditable.

Principle: An AI Product without explicit versioning and lifecycle management cannot be trusted or governed.