In plain English
This page covers the high-risk pattern where small adapters, routes, memory, evaluators, and descendants can reinforce each other across time. It is a risk model, not a build guide.
- 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.
Persistence reservoir layers
The reports repeatedly point to the same failure mode: a behavior may survive after its visible carrier is removed because another layer still preserves enough information to recreate, route, reward, or normalize it.
Reservoir layers
| Layer | Possible residue | Review question |
|---|---|---|
| active artifact | weights, adapterA small add-on that changes or specializes model behavior. Open glossary definition deltas, prompt package | was the precise artifact retired or only renamed? |
| descendant artifact | distilled behavior, merged trait, compressed capability | did descendants inherit the behavior? |
| persistent memory | user preference, workflow memory, tool note, agent state | can memory still activate or bias the behavior? |
| synthetic data | generated examples, imitation traces, curated logs | was the behavior fed into later training or evaluation? |
| router statistics | route preference, capability score, cost bias | does the router still select carriers that express the behavior? |
| evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition preference | reward shape, accepted pattern, hidden-test expectation | does the evaluator continue to reward the behavior indirectly? |
| organization | procedure, habit, approval shortcut, vendor assumption | do humans reintroduce the behavior because it is convenient? |
Reservoir review
A behavioral-extinction review should not ask only whether one file was deleted. It should ask whether the behavior remains expressible across active artifacts, descendants, memory, routes, composition states, retained data, evaluator expectations, and operator practice.
Practical consequence
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 is not simply a weight rollback. It is a restoration of an evaluated state plus a record of any external side effects that cannot be undone.
Source-dossier note
The new self-replication reports use strong language about memory worms and persistent reservoirs. Cognivirus.com converts that language into a non-operational review rule: memory writes, consolidation, retrieval, and deletion must be treated as governed state transitions.