AI Production Release Lifecycle

AI Production Release Lifecycle

Key jargon

Term Plain-language meaning
Release candidate A version proposed for deployment after testing.
Canary A limited deployment used to observe behavior before broad release.
Rollback Restoring a previously known-good version.
Feature flag A runtime control that enables or disables behavior without rebuilding the application.

Key concepts

Concept map

flowchart LR
    A["Build versioned candidate"] --> B["Run offline evaluation gates"]
    B --> C["Canary and monitor"]
    C --> D["Promote or rollback"]

Release unit

Version model, prompt/instructions, tool definitions, retrieval corpus/index, code, policy, evaluation set, and deployment configuration. “Same model” does not mean same system.

Gates

  1. Offline representative and safety evals.
  2. Dependency/image provenance and security review.
  3. Load, failure, and recovery tests.
  4. Shadow or internal traffic.
  5. Limited canary with kill switch.
  6. Metric and human-review comparison.
  7. Progressive rollout or rollback.
  8. Post-release review and dated documentation.

Exercise

Write a release checklist for changing only the system prompt. Identify which regressions could still occur and why “no code change” is not “no risk.”

Ownership

Name product, model, data, security, privacy, operations, and incident owners. An AI system without an accountable owner is not production ready.