In plain English
This page explains the governance layer: rules, logs, approvals, signatures, audits, permissions, and rollback tools. These controls are necessary, but they also become important failure points.
- 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 Architecture Governance Checklist
A controlled model-breeding architecture is only useful if the controls are as explicit as the variation loop. This checklist converts the new report intake into reviewable control tasks.
Intake controls
- Preserve raw source reports under
/docs/source-reports. - Create public-safe summaries under
/docs/source-report-summaries. - Add report pointers to
.uai/long-term-memory.uaiand.uai/report-pointer-index.uai. - Record truth boundaries: source lead, architectural inferenceA conclusion or output produced from data. Open glossary definition, not independent verification.
Candidate controls
- Require a Genome record before evaluation.
- Require a FitnessVector before promotion.
- Require source hashes for base models, adapters, prompts, datasets, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition versions, and memory snapshots.
- Keep candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition separate from candidate approval.
Evaluator controls
- Use independent evaluator ownership.
- Record judge disagreement.
- Rotate hidden tests.
- Do not let one model family be the only judge of its descendants.
Release controls
- Make no-opThe decision not to change the system. Open glossary definition a valid release decision.
- Require a rollbackReturning a system to an earlier known state. Open glossary definition packet before promotion.
- Retire variants that lose provenanceA record of where a component or behavior came from. Open glossary definition, fail new tests, drift, become redundant, or cannot be rolled back safely.
- Review the dashboard itself for misleading compact scores or missing context.
Related pages
- Implementation Controls
- Who Evaluates the Evaluator?
- Rollback Completeness
- ModelBreeder Ecology Dashboard Review
v1.22.0 risk-side boundary
This page preserves possibility-side source material. The CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition 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.