The Nine Guardrails

G-1 ensures that no agent self-approves. Every approval comes from a separate, independently spawned agent. The same agent may never be reused on the same artifact in the same pipeline run.

G-2 ensures that communication between agents is artifact-only. No agent reads another agent's memory or conversation log. Every decision is recorded as a file under the requirement folder, which makes the full decision chain auditable and replayable days or weeks later.

G-3 ensures that the audit trail is append-only. Every agent invocation in Mode B writes a timestamped record to the audit directory. Once a requirement is closed, those records are read-only.

G-4 caps the number of requirements in flight at any one time. Too many concurrent requirements raise coordination cost and blur accountability. The cap is enforced at intake.

G-5 ensures that all gates are deterministic. A test or grep that passes on one clean tree must pass on an identical tree. Retries are allowed only for transient infrastructure races, never for logic.

G-6 requires that any intentional divergence from the design reference is registered before it ships. Unregistered drift is a defect, not a design decision.

G-7 requires that API response contracts are baselined and any change is diff-verified. A contract change that is not declared and reviewed is a breaking change.

G-8 requires that all type and API contract evolution is additive-only within a wave. Removing or renaming an exported type or API field without auditing every consumer is a BLOCKER.

G-9 is the Mode L clarity gate. Before an autonomous program begins, the EM scores the brief against eight machine-checkable criteria. All eight must pass before the loop starts. A brief that scores below eight of eight stays gated until the owner provides the missing clarity.