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.
Anatomy
This page explains where an AI behavior can “live.” It may be in a model, but it may also be in a prompt, a memory record, an adapterA small add-on that changes or specializes model behavior. Open glossary definition, a dataset, a tool setting, or an evaluation rule.
The system is the moving part.
Models, adapters, routers, memory, evaluators, and registries are shown as separate carriers. The moving signature is a behavior that can remain expressible even as its visible host changes.
The anatomy section defines the parts of an adaptive model ecologyA changing AI system made from many connected parts, not just one model. Open glossary definition: base models, specialists, adapters, prompt packages, persistent memory, synthetic data, routers, evaluators, registries, lineage graphs, release controllers, operators, and incentives.
- Pattern introduced
- Pattern expressed
- Output rewarded or retained
- Pattern encoded elsewhere
- Original carrier replaced
- Pattern reappears
- Lineage becomes less explanatory
- Assurance becomes stale
Not every pattern is harmful. The framework applies to useful, neutral, erroneous, biased, deceptive, brittle, and dangerous behaviors. The concern is not the moral character of a component; it is whether the system can still express a behavior after the apparent carrier has changed.
Start with hosts
A cognitive host is any component capable of carrying or expressing a pattern. In a modular ecology, the carrier may be a weight delta, a route, a retrieval rule, a memory record, a synthetic training exampleAI-generated or transformed data used for training or evaluation. Open glossary definition, or an evaluator preference.
Then inspect reservoirs
A reservoirA place where a behavior can remain after the first carrier is removed. Open glossary definition is any place a pattern can survive when an artifact is removed: memory stores, logs, datasets, descendants, adapter stacks, policy packages, release aliases, and human operating routines.