Controlled change and versioning

Version the behavior, not only the model

An AI system's behavior depends on a configuration set: model and tokenizer, prompt and templates, retrieval corpus and index, policy, tools and schemas, runtime, dependencies, hardware profile, feature flags, and application code. A release identifier should resolve to that complete set.

Classify change

Define change classes based on possible impact, not file size. A one-line policy or prompt update can be material. For each class, specify required review, evaluation, approval, deployment, monitoring, rollback, and evidence.

Common reassessment triggers include:

• Model, quantization, tokenizer, runtime, or hardware changes.
• Prompt, tool, retrieval, corpus, permission, or policy changes.
• New users, purpose, data class, geography, provider, or deployment boundary.
• Security findings, incidents, drift, quality regression, license changes, or expired exceptions.

Controlled promotion

Identify the immutable candidate and its parents. Evaluate in a representative environment, record results and exceptions, obtain authorized approval, promote without rebuilding, verify deployment identity, monitor the defined window, and preserve rollback.

Separate content movement from authorization. The ability to copy an artifact must not imply permission to promote or execute it.

Rollback and forward recovery

Rollback must consider data schemas, indexes, caches, state, secrets, and evidence compatibility as well as binaries. Some changes are not safely reversible; document forward-recovery steps and the point of no return.

Retirement

Remove superseded artifacts from active selection, revoke credentials and approvals where appropriate, preserve required evidence, update dependencies, and verify that hidden routes cannot still invoke the retired version.

Record architecture consequences and reversal triggers in an architecture decision record. Tool-specific deployment syntax belongs in separately tested adapters; the change contract remains stable across platforms.