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.
Controlled Model Breeding Loop
The uploaded ModelBreeder controlled-evolution reports recommend moving from vague model-breeding language to explicit genomes, fitness vectors, novelty scores, evaluation checkpoints, and no-op release decisions. CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition uses that recommendation as a control vocabulary: evolution is allowed only when its carriers, measurements, and rollback path are explicit.
Loop
| Phase | Required record | Stop condition |
|---|---|---|
| Champion selection | Base model ID, adapter stackA set of adapters loaded together, usually in a defined order. Open glossary definition, prior fitness vector, rollback packet | No champion with complete provenanceA record of where a component or behavior came from. Open glossary definition. |
| Candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition | Mutation or crossover log, parent hashes, random seed, tool version | Candidate created outside quota or without lineageThe parent-child history of models, adapters, datasets, or releases. Open glossary definition. |
| Evaluation | Fixed suite, hidden suite ID, resource metrics, novelty score, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition versions | Evaluator disagreement or missing measurement. |
| Selection | Fitness threshold, cost cap, novelty archive update, human approval | No candidate materially improves the champion. |
| Promotion | Signed manifest, deployment alias, rollbackReturning a system to an earlier known state. Open glossary definition packet, monitoring plan | Rollback packet incomplete. |
| No-opThe decision not to change the system. Open glossary definition | Evidence that current champion remains stronger or safer | Release pressure is the only reason to change. |
Public terminology adjustment
The constructive wording from the report is useful when it clarifies controls. This site can use “evaluation checkpoint,” “fitness proof,” and “candidate model” in engineering contexts. It should not remove risk boundaries, because Cognivirus is a risk-analysis site. A fitness proof is acceptable only when it names the evaluation suite, composition, and limitation.
Failure modes to keep visible
- A high fitness score can hide weak coverage.
- A novelty score can reward irrelevant difference.
- Multi-parent merge can inherit a latent defect from a parent or from the merge operator.
- Speciation can preserve weak niches if the archive lacks retirement rules.
- A no-op decision must remain socially valid, not just technically allowed.