CompositionDemonstrated research proof-of-conceptv1.22.1

In plain English

This page explains why testing AI parts one by one is necessary but incomplete. Safe-looking parts can still produce unsafe behavior when combined.

  • 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.

When Individually Benign Adapters Become Unsafe Together

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

If behavior emerges only when adapters are composed, isolated A small add-on that changes or specializes model behavior. Open glossary definition review can pass the wrong unit of analysis.

Mechanism

The mechanism is interaction. Components exchange context through hidden state, prompts, outputs, adapters, memory retrieval, tool calls, A system that judges whether an AI output or candidate is acceptable. Open glossary definition prompts, and release rules. Each interaction can change what the next component sees and what the system is allowed to do.

Evaluation implication

The evidence record should include the exact A machine-readable record of the exact runtime composition used for an evaluation, release, incident, or rollback. Open glossary definition. A statement such as “Adapter C passed” is incomplete unless it says which base model, load order, router, prompt package, memory snapshot, evaluator, inference configuration, and deployment environment were used.

Practical control

Use composition-aware test suites, targeted higher-order samples, route-level canaries, independent judges, and Returning a system to an earlier known state. Open glossary definition packets that include all relevant runtime dependencies.

<!-- expanded-release-content -->

Relationship-dependent risk

Evidence levelDemonstrated research proof-of-conceptTechnical label: Emerging evidence

An adapter can appear benign when tested alone and become safety-relevant only when composed with another adapter, a particular base, a merge order, or a prompt package. The dangerous behavior can live in the relationship between components. That makes supplier review and isolated evaluation necessary but insufficient.

Load order and base dependence

Adapters are not abstract traits that apply identically everywhere. Their effect depends on the base model, layer targets, scaling, load order, quantization, prompt context, and other active modifications. A safety adapter and a capability adapter may compete. A domain adapter may amplify behaviors that a safety adapter was not evaluated against. A merged adapter may hide its origin while preserving an unwanted trigger.

Evaluation rule

Adapter approval should be scoped. It should name permitted bases, permitted co-adapters, load order, merge coefficients, prompt policies, A conclusion or output produced from data. Open glossary definition settings, and tool profiles. Any dynamic adapter composition should be logged as part of the runtime manifest. Unknown combinations should default to constrained routes until evaluated.