AI systems change constantly: models update, prompts evolve, retrieval corpora shift, tools change. During incidents, that rate of change becomes a liability. A change freeze is a short-term safety measure that stabilises the system while you investigate.
What to freeze
Effective freezes focus on the behaviour surface:
- Prompt and policy changes (see prompt change control).
- Routing and model changes (see routing and failover).
- Ingestion and knowledge base updates (see ingestion pipelines).
- Tool schema changes and permissions (see tool authorization).
What to keep running
Freezes do not mean total shutdown. Continue to:
- Monitor key SLOs and drift signals (see AI SLOs and drift monitoring).
- Log incidents and capture traces for replay (see incident response).
- Run controlled canaries for potential fixes (see canary rollouts).
How to resume safely
Resuming requires clarity on root cause and new safeguards:
- Update evaluation suites to cover the failure mode.
- Document changes and approvals in the change log.
- Use canaries to reintroduce change gradually.
A change freeze is not a sign of failure. It is a sign of mature operations: stability first, learning second, improvement third.