The proof loop

North Star uses four visible stages for useful work: Plan, Run, Review, and Remember.

1. Plan

Planning loads the project goal and recent evidence, then chooses one useful move. The plan defines:

what the worker is trying to achieve
which files, systems, or records are in scope
what must remain unchanged
which checks will prove the result
the time, token, or spend limit
where the worker must stop and ask

A plan is useful when another person can understand the intended result before execution begins.

2. Run

The task starts now or at a scheduled time. North Star launches the selected AI worker inside a bounded workspace and records the actions and results that matter.

Running does not mean unrestricted autonomy. The plan, permissions, protected actions, and available capacity still apply.

3. Review

Review asks whether the result is true on the real surface. Tests and builds are evidence, but they are not the entire proof. A website is checked in a browser. A deployment is checked from a fresh client. A risky decision returns to a person.

Review can accept the result, request another bounded change, or reject it and use the recovery path.

4. Remember

The outcome becomes durable project memory:

what changed
what was verified
the exact source and deployed version
what remains uncertain
how to recover
where the next session should begin

Remember is why a new conversation can resume from evidence instead of rediscovering the project.

Next: Projects and context