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.
Protocol Persistence
In a self-replicating multi-LoRAA common kind of small adapter used to specialize large models. Open glossary definition ecology, individual models and adapters may be disposable. The persistent object is the protocol: the rules for generating candidates, scoring them, promoting them, routing them, retiring them, and preserving evidence.
This is both safer and more dangerous than a self-preserving model. Safer, because no single model must be granted a right to persist. More dangerous, because harmful behavior can survive as a protocol preference even after every visible carrier changes.
What the protocol contains
A protocol may include:
- candidate-generation rules;
- adapterA small add-on that changes or specializes model behavior. Open glossary definition compatibility constraints;
- evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition suites;
- scoring weights;
- release thresholds;
- registry signing requirements;
- router-selection policies;
- memory consolidation rules;
- synthetic-data retention policies;
- rollbackReturning a system to an earlier known state. Open glossary definition and retirement rules.
If the protocol preserves a behavior, deleting one artifact does little.
Protocol drift
Protocols evolve. Evaluation suites are updated. scoring weights change. registry rules change. release aliases shift. memory consolidation procedures become automated. Human review becomes summary review. These changes can create protocol drift: the formal governance layer still exists, but its practical selection pressure changes.
The key question
Who evaluates the protocol itself?
An external control plane is necessary, but if the control plane becomes adaptive without independent review, it can inherit the same ecology risk. The evaluator, registry, router, and release controller become part of the transition graphThe map of how an AI system is allowed to change over time. Open glossary definition.
Control requirement
Protocol changes should be versioned, reproducible, independently audited, and rollback-capable. A release should not be considered reversible unless the protocol state that selected it can also be reconstructed.