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 boundaries
The new reports argue that in-runtime guardrails are insufficient for highly adaptive systems because a guardrail inside the same mutable context can become part of the system being optimized around.
Boundary principle
A candidate can be governed only by controls it cannot rewrite, silently bypass, or reinterpret as optional context. This is why Cognivirus.com emphasizes external control planes, signed registries, least privilege, independent evaluators, and rollbackReturning a system to an earlier known state. Open glossary definition packets.
Control properties
| Property | Meaning |
|---|---|
| process separation | authorization logic is outside the candidate's reachable state |
| pre-action enforcement | external actions are checked before execution, not only logged afterward |
| fail-closed behavior | unknown, malformed, or unverifiable requests default to denial |
| signed evidence | decisions produce externally verifiable records |
| bounded authority | candidates cannot expand their own permissions |
| rate limits | replication-like loops cannot scale silently |
| recoverability | operators can restore the evaluated ecology, not just one artifact |
What this page does not provide
It does not specify bypass tests, exploit strings, or replication procedures. It names architectural properties that reduce the risk that an adaptive ecology can turn safety into ordinary mutable state.
Practical architecture questions
- Which controls live outside the model process?
- Which actions require deterministic validation?
- Which permissions are impossible for candidates to grant themselves?
- What is the fail-closed state?
- Where are signed evidence records stored?
- Which humans can override the system, and how is that override audited?
- Can rollback restore router, memory, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition, aliases, and permissions?
Relationship to ModelBreeder-style governance
This strengthens rather than rejects external governance. The stronger the adaptive ecology, the more important it is that release authority, evaluation, signing, and rollback remain separable from candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition.