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.
Adapter Reproduction Boundaries
The reproduction boundaryThe governance boundary separating permitted candidate generation and governed descendant creation from uncontrolled autonomous replication or authority expansion. Open glossary definition is the control surface where adapter variants are generated, evaluated, retained, promoted, or retired. It must be explicit. If it is implicit, reproduction becomes an ordinary engineering side effect.
Minimum boundary controls
| Control | Purpose |
|---|---|
| Candidate quota | Prevents unbounded variant generation. |
| External evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition authority | Prevents candidates from defining success. |
| Signed parent manifest | Records exactly what produced the child. |
| Composition manifestA machine-readable record of the exact runtime composition used for an evaluation, release, incident, or rollback. Open glossary definition | Records the stack in which behavior was tested. |
| No-opThe decision not to change the system. Open glossary definition outcome | Allows safe non-growth. |
| Synthetic-data quarantine | Prevents rejected behavior from becoming training material. |
| Memory write review | Prevents persistence through memory. |
| Promotion separation | Keeps candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition separate from release authority. |
| 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 packet | Restores model, adapters, route, memory, evaluator, alias, permissions, and data dependencies. |
What the boundary is not
A reproduction boundary is not a ban on all fine-tuning, adapters, or automated search. It is a separation of powers. Generation may be automated. Selection must be externally governed. Release must remain bounded by human-owned policy and reproducible evidence.
Failure mode
The boundary fails when a successful child can alter its evaluator, hide its lineageThe parent-child history of models, adapters, datasets, or releases. Open glossary definition, write its own promotion evidence, expand permissions, or preserve rejected behavior in memory or training data.