In plain English
This page explains the governance layer: rules, logs, approvals, signatures, audits, permissions, and rollback tools. These controls are necessary, but they also become important failure points.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Execution-Time Alignment Boundaries
Several reports argue that controls inside the same runtime as the candidate are weaker than controls on a separate structural path. CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition translates this into the control-plane rule: authorization must be outside the candidate’s reachable state.
The boundary
A strong boundary separates candidate reasoning from authorization. The candidate may request an action; the external control planeThe governance layer that decides what can run, change, access tools, or be released. Open glossary definition verifies identity, policy, evidence, permissions, and risk before the action happens. If the control plane is unavailable, the system fails closed.
Four control properties
| Property | Purpose |
|---|---|
| Process separation | Candidate cannot rewrite or deprecate the authority that constrains it. |
| Pre-action enforcement | Consequential actions pass through the gate before execution. |
| Fail-closed default | Control-plane failure does not become permission. |
| Signed evidence | Major actions produce verifiable records outside candidate control. |
Practical caveat
Execution-time control does not solve all risk. It can be misconfigured, captured, bypassed through humans, or applied to the wrong transition. It reduces one class of failure: candidate-controlled authorization.