Skip to main content
This is the vocabulary the product, the docs, and the website all use. Each concept is labeled by its claim class so there is never ambiguity about what is available to you today.
How we label claims. Live = running in production. Live-proven = running in production and exercised end to end by production acceptance and continuous security testing. Supported = available, used where the topology calls for it. Early Access = real, off by default, on request. Roadmap = planned, not shipping. Vision = direction, not a feature. We never describe an emerging capability as if it ships today.

​
The idea

A change can be authorized and still be dangerous. Your pipeline confirms the actor is allowed to make a change; it cannot tell you the service being changed is mid-incident, or that the change would exceed the namespace’s quota the moment it lands, or that the artifact was never evaluated. ChangeGuard AI makes the change the primary object. It evaluates a proposed change against the live state of the environment it will land in and returns a judgment before the change executes. When a change is authorized, it binds that authority to a short-lived Permit, lets an Executor carry out only the permitted action, and records what actually followed against the change that caused it.

​
Judgment, Permit, Executor, Observation

​
Advise and Auto

Authority is a per-environment decision with two modes. Advise (the default): ChangeGuard can diagnose, prepare and execute, but a human authorization must exist before anything executes. Auto (opt-in): it acts only inside an explicit policy — namespaces, fix types, a confidence floor, an hourly cap — and falls back to Advise outside it. Autonomous-capable by design, human-controlled by default. Details — Live-proven.

​
Fleet, Edge and the Executor

  • Fleet — the centralized evidence plane. ChangeGuard reads your Kubernetes API from outside, over read-only access you authorize through AWS Account Center. Nothing to install. Live.
  • Edge — an optional local-evidence extension: a small read-only collector for signals that need presence on the node, or clusters that can’t be reached centrally. Never required for judgment or execution. Supported.
  • Executor — the governed action plane, above. Inert without a Permit.
Fleet, Edge and the Executor explains which you need.

​
The rest of the product

​
What ChangeGuard AI does not do today

  • It does not learn from every change. The outcome record is instrumented — a verdict is recorded before a change and what followed is recorded after — so that prediction can be measured against reality. No model, threshold, policy or prompt is modified by an outcome. Recall of similar past incidents (Early Access) is retrieval, not retraining.
  • It does not act on its own by default. Every environment starts in Advise. Auto is opt-in and policy-bounded.
  • It does not cover infrastructure beyond Kubernetes. Terraform, IAM, cloud resources, networking and databases are vision, not shipping capability.
  • It does not hold a certification. SOC 2, ISO 27001 and FedRAMP are not claimed. See Testing and assurance for what is proven.

​
A note on honesty

If a capability isn’t labeled Live or Live-proven above, treat it as not-yet-generally-available and check with your ChangeGuard AI contact before depending on it. The website uses these exact same labels — so what you evaluate is what you can turn on.