In plain English
This page explains where an AI behavior can live. It may be in a model, but it may also be in a prompt, memory record, adapter, dataset, tool setting, evaluator rule, or human workflow.
- 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.
The Replication Surface
In a self-replicating multi-LoRA ecology, the replication surface is larger than the adapterA small add-on that changes or specializes model behavior. Open glossary definition file. It includes every place where a behavior can be encoded, rewarded, reconstructed, or reactivated.
The obvious surface is the adapter itself. A LoRA deltaThe behavior-changing weight difference stored by a LoRA adapter. Open glossary definition can carry task specialization, refusal changes, stylistic behavior, domain knowledge, or a fragile shortcut. But the more important surface is the set of relationships around that delta: the base model it modifies, the load order, merge coefficients, router path, prompt policy, memory state, evaluator expectations, and release alias.
Primary carriers
| Carrier | Persistence question |
|---|---|
| Base model | Does the behavior require this base, or can it transfer? |
| LoRA adapterA common kind of small adapter used to specialize large models. Open glossary definition | Does the delta carry the behavior alone or only in a stack? |
| Adapter stackA set of adapters loaded together, usually in a defined order. Open glossary definition | Does load order change expression? |
| Router policy | Does the behavior appear only when routed through a specific path? |
| Prompt-policy package | Does policy wording activate or suppress the behavior? |
| Memory store | Did the behavior write cues that outlive the adapter? |
| Synthetic dataset | Did generated outputs become future training material? |
| EvaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition | Did the judge reward the pattern or encode a blind spot? |
| Registry alias | Did the deployment name keep pointing users toward descendants? |
Why this differs from ordinary replication
A biological analogy can mislead if it implies a single organism copying itself. The Cognivirus concern is functional replicationThe reappearance or preservation of a behavior through descendants, memory, synthetic data, evaluators, routes, or adapters without copying a whole model. Open glossary definition. A behavior can be deleted from one file and still reappear through distillation, memory consolidation, synthetic training examples, evaluator preferences, or a descendant adapter.
This is why the phrase “self-replicating multi-LoRA ecosystemA proposed Cognivirus term for an adaptive model ecology where LoRA adapters or adapter-derived behavior can be generated, selected, copied, recomposed, promoted, or preserved across bases, routes, memory, synthetic data, and descendants. It is a risk model, not an implementation instruction. Open glossary definition” should be read as a transition-graph risk. The system may not have a self. The adapter may not know anything. The pipeline may be governed. The persistence can still occur if the process repeatedly generates variants and preserves whatever scores well.
What must be inspected
A complete anatomy review should inspect the adapter package, base compatibility identity, merge recipe, load order, route preconditions, memory write permissions, evaluator versionThe exact version of the evaluator used for a test or release. Open glossary definition, synthetic-data retention policy, canary results, and rollback dependencies. Anything less can certify a visible component while missing the surface through which the behavior actually persists.