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.
The Adapter as Propagule
In the CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition vocabulary, a propagule is a small carrier that can move a behavior through an ecology. The term is metaphorical. A LoRA adapterA common kind of small adapter used to specialize large models. Open glossary definition is not a seed, spore, organism, or malware sample. It is a small parameter-efficient modification that can alter how a base model behaves.
The reason the metaphor is useful is structural. A full model is heavy. An adapterA small add-on that changes or specializes model behavior. Open glossary definition is comparatively small. A small artifact can move through ordinary engineering channels more easily: pull requests, model hubs, internal registries, experiment bundles, evaluation candidates, edge deployments, or fine-tuning pipelines.
Why smallness changes governance
Small components are easier to normalize. They are treated as patches, plugins, or configuration. That can be appropriate. It can also create a review gap. A small delta may be examined for provenanceA record of where a component or behavior came from. Open glossary definition, license, or benchmark score while its composition-specific effect remains untested.
The hazard is not that a LoRA file is intrinsically suspicious. The hazard is that a behavior-changing component can appear operationally routine.
Adapter-level inheritance
An adapter can carry more than a task skill. It can carry style, refusal shifts, shortcut strategies, representational changes, or brittle assumptions about the base model. When copied, merged, distilled, or used to generate synthetic examples, those traits can become inherited by other components.
Adapter inheritance becomes difficult to interpret because the behavior may not be visible in the adapter alone. It may require:
- a compatible base family;
- a particular load order;
- a prompt-policy context;
- a router-selected task;
- a memory state;
- a tool permission profile;
- an evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition blind spot.
The inspection problem
A component can pass inspection because the dangerous behavior is not local to the component. It is relational. A clean component scan, a hash, and a model card are useful integrity signals, but they do not prove safety across all compositions.
Control requirement
Treat adapters as controlled behavioral deltas. Registry records should include the evaluated base, adapter load order, merge coefficients, prompt-policy version, memory assumptions, tool profile, evaluator versionThe exact version of the evaluator used for a test or release. Open glossary definition, and known incompatibilities.
A propagule is not a villain. It is a carrier. The safety question is what the ecology lets it carry.