Risk LabStrong architectural inferencev1.22.1
In plain English
This page provides local browser worksheets. They help plan reviews; they are not formal safety certifications.
- 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.
ModelBreeder Risk Review Worksheet
Evidence levelStrong architectural inferenceTechnical label: Strong architectural inference
Use this worksheet when a system borrows ModelBreeder-style ideas: populations, genomes, fitness vectors, novelty archives, speciation, local dashboards, adapters, or multi-parent merging. The worksheet is qualitative. It does not certify safety.
Quick triage
| Question | Lower concern | Higher concern |
|---|---|---|
| How many candidates were generated? | bounded batch with owner | unbounded or unknown generation |
| What selects candidates? | multiple independent evaluators | one composite score or one model judge |
| Is novelty quarantined? | novel items archived first | novelty helps promotion directly |
| Are specialists scoped? | niche owner and route limits | specialists can be routed broadly |
| Are parent records complete? | hashes and merge operator recorded | descendant has friendly name only |
| Can rollbackReturning a system to an earlier known state. Open glossary definition restore environment state? | model, adapter, prompt, memory, router, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition, permissions | model weights only |
| Are edge/local states visible? | local manifest and retirement receipt | browser or edge cache unknown |
| Does the dashboard show limits? | skipped checks and what-is-not-proven visible | only green badges and charts |
Reviewer prompts
- List every candidate and descendant created since the last approved baseline.
- Identify the selection pressure: metric, judge, benchmark, human review, cost target, latency target, or novelty score.
- Mark which pieces can carry behavior: adapterA small add-on that changes or specializes model behavior. Open glossary definition, prompt, memory, router, evaluator, tool policy, synthetic data, browser cache, local file, deployment alias.
- Pick one promoted candidate and reconstruct its lineageThe parent-child history of models, adapters, datasets, or releases. Open glossary definition back to all parents.
- Pick one retired candidate and prove its behavior cannot still be reintroduced through memory, routing, synthetic data, or descendants.
- Identify which evidence claims are demonstrated and which are architectural inferenceA conclusion or output produced from data. Open glossary definition.
- Record the no-opThe decision not to change the system. Open glossary definition option before the release decision is framed as a yes/no deployment.
Required artifacts
- Genome or equivalent candidate manifest.
- FitnessVector or equivalent evaluation report.
- Parent list and merge operator record.
- Adapter stackA set of adapters loaded together, usually in a defined order. Open glossary definition and load-order record.
- Memory snapshotA saved state of what the AI system remembers. Open glossary definition and RAG source labels.
- Router policy and evaluator versionThe exact version of the evaluator used for a test or release. Open glossary definition.
- Tool permissions and action-boundary review.
- Rollback packet and retirement record.
Related pages
- ModelBreeder Risk Side
- Model Breeding Risk Boundaries
- Risk Lab
- Transition Graph Danger Review
- Retirement Completeness Review
Cognivirus.com risk-side note
Evidence levelStrong architectural inferenceTechnical label: Strong architectural inference
This page belongs to Cognivirus.com because it translates ModelBreeder-style possibility into risk review and governance controls.