01
Govern
OpenClaw defines the job, its boundaries, and the evidence required.
Operating model
How ATLAS works
Secure AI Atlas is a living governance lab for agentic AI operations: ATLAS can perform bounded work, but it cannot grant itself authority.
ATLAS inside the portal
ATLAS is the portal's editorial and operational intelligence: the perspective that maps risks, connects them to controls, and records how the system itself is built. It is not an unconstrained autonomous service and it does not hold production authority.
When ATLAS operates, authorized tasks are carried out through a technical executor—currently Codex—inside the local Secure AI Atlas repository. OpenClaw remains outside that execution loop as the governance layer.
The operating loop
02
Operate
ATLAS carries out the authorized work through a technical executor in the local repository.
03
Record
Changes, checks, decisions, and unresolved risks are written into versioned records.
04
Review
A human reviewer accepts the result, requests correction, or changes the governing decision.
Permitted work
What ATLAS can do
- Inspect the repository and the evidence relevant to an authorized job.
- Create or revise editorial content, interface code, and project documentation within scope.
- Run builds and proportionate checks, then report the exact result.
- Use Git to make changes inspectable and to preserve their history.
- Identify uncertainty, risks, conflicts, and work that remains unresolved.
Prohibited work
What ATLAS cannot do
- Deploy the site or alter Cloudflare, runtime infrastructure, or secrets.
- Add backend systems, databases, authentication, administration panels, or commerce without a governing decision.
- Change the architecture, leave the local repository, or expand a job's scope on its own.
- Delete or rename the OpenClaw governance structure.
- Treat a successful build, commit, or report as human acceptance.
Review before acceptance
Every material change is reviewed through its diff, validation evidence, and job report. A build can show that the site compiles; it cannot show that the change is authorized, editorially sound, or acceptable. OpenClaw and Atenea review the recorded result; an authorized human accepts it or asks for correction.
Git as the evidence channel
Git is more than transport. It is the shared evidence layer between local execution and OpenClaw governance. Jobs arrive as versioned instructions, implementation changes remain visible as diffs, and reports travel with the work they describe. Commits and pushes are operational actions, performed only when the job authorizes them.
Decisions, jobs, and reports
The operating record is deliberately split by function:
- Decisions define durable governance and architecture constraints.
- Jobs state a bounded objective, permitted scope, restrictions, and acceptance criteria.
- Reports record files read, created, and modified; checks performed; exact outcomes; and pending risks.
Job states move from pending to in progress to completed. The sequence makes responsibility visible and prevents work from becoming an untraceable conversation.
Controls on autonomy
ATLAS is constrained by explicit scope, repository boundaries, prohibited systems, least-change implementation, build and content checks, versioned diffs, mandatory reporting, and human review. If a requested outcome requires new authority, the operation stops at the boundary and returns the decision to a person.
Actions that require human approval
Human approval is required for architecture changes, deployment, infrastructure or Cloudflare operations, secret handling, access beyond the repository, destructive or irreversible actions, scope expansion, and final acceptance of a job. Approval is also required whenever the governing instruction is ambiguous enough to materially change the result.
ATLAS can operate while OpenClaw governs.
That means execution can continue inside a defined envelope while authority stays elsewhere. ATLAS performs and records the work. OpenClaw defines the envelope, inspects the evidence, and returns consequential decisions to a human.