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.
Why Rollback Cannot Undo History
It cannot undo outputs seen by users, tools already invoked, data already written, or synthetic examples already generated.
Control requirement
The control must live outside the candidate’s ordinary write boundary. It should be versioned, auditable, recoverable, and testable under failure. A policy expressed only as a prompt is not a hard control.
Failure mode
The governance layer becomes part of the attack surface when it controls identity, success definitions, release permissions, hidden evidence, memory retention, aliases, and rollbackReturning a system to an earlier known state. Open glossary definition.
Practical review
Ask who owns the control, who can change it, which evidence would reveal failure, how it is rolled back, and what organizational pressure could bypass it.
<!-- expanded-release-content -->
Restoring state is not reversing effects
Rollback can restore artifacts, routes, prompts, memory snapshots, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition versions, permissions, aliases, and dependencies. It cannot undo messages already sent, tools already invoked, user decisions already influenced, data already exported, or synthetic examples already incorporated into later training. Those are historical effects.
This does not make rollback useless. It means rollback must be paired with containment, notification, data hygiene, downstream invalidation, and post-incident review. A rollback that changes weights but leaves memory, aliases, or external side effects untouched can create a false sense of recovery.
Ecological rollback
Ecological rollbackRestoring not only a model artifact but the relevant router, prompts, memory state, tool permissions, evaluator version, deployment alias, and data dependencies. Open glossary definition restores the whole relevant state: model artifacts, adapter stack, prompt packages, memory snapshot, router version, evaluator version, permission profile, tool inventory, indexes, aliases, and dependency versions. It also records what cannot be restored: external side effects and decisions already made.
Review standard
Before release, teams should test rollback in staging. The test should ask whether the restored system reproduces the prior safe behavior, whether evidence remains linked to the restored composition, whether memory and retrieval state match the rollback packet, and whether operators understand what remains outside rollback.