In plain English
This page explains how AI systems can change over time through updates, tests, retraining, memory, and approvals even when no single model rewrites itself.
- 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.
Model Breeding Operators and Selection
The new reports add operator vocabulary for controlled model breeding. CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition uses those operators as review objects, not as instructions to build autonomous reproduction.
Operator map
| Operator | Review question |
|---|---|
| AdapterA small add-on that changes or specializes model behavior. Open glossary definition mutation | Which adapter changed, what rank or density changed, and who approved it? |
| Multi-parent crossover | Which parents contributed behavior, weights, prompts, or adapters? |
| SLERP / interpolation | What layer or tensor range was blended, and what evidence says the blend is valid? |
| Task-vector merge | Which task vector was added, removed, or scaled? |
| TIES / DARE-style sparsity | What was trimmed, dropped, elected, or rescaled, and how was interference tested? |
| Speciation | Which niche does the candidate belong to, and does it preserve useful diversity? |
| Novelty archive | Is the candidate behaviorally new or merely a duplicate with new packaging? |
Selection rule
A promotion ruleThe rule that decides what survives. Open glossary definition should never be only “highest score wins.” It should include utility, cost, latency, memory, novelty, evaluator disagreement, rollback completeness, source provenance, and no-op availability.
Failure modes
- A composite score can hide evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition disagreement.
- A novelty archive can reward weirdness without usefulness.
- A merge can preserve hidden behavior while improving visible benchmarks.
- A speciation rule can become a loophole if it protects weak variants from review.
- A fast breeding loop can outpace source and rollbackReturning a system to an earlier known state. Open glossary definition evidence.
Defensive use
Use these operators to ask better questions about candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition. The site does not provide parameter recipes, exploit payloads, or model-construction instructions.
Related pages
- Controlled Model Breeding Loop
- Fitness, Novelty, and Selection
- No-op as Reproductive Control
- ModelBreeder Architecture Governance Checklist
v1.22.0 risk-side boundary
This page preserves possibility-side source material. The Cognivirus use is narrower: translate controlled-evolution vocabulary into risk review. For that translation, use ModelBreeder Risk Side, ModelBreeder Risk Translation, and Model Breeding Risk Boundaries.