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.
Model Breeding Risk Boundaries
Controlled evolution is only controlled when variation, evaluation, selection, release, memory, and retirement each have a boundary. This page defines the boundaries CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition should inspect when ModelBreeder-style systems are discussed.
Boundary 1: generation budget
Do not let candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition become background noise. Record every candidate batch with a UTC timestamp, parent list, mutation operator, adapter stack, prompt policy, evaluator version, and reviewer owner.
Minimum controls:
- candidate-generation quota
- human-visible generation ledger
- freeze switch
- backlog threshold
- no autonomous production promotion
Boundary 2: evaluation independence
Fitness is a selection pressure. It must not be owned entirely by the same loop that generates descendants.
Minimum controls:
- independent evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition owner
- hidden-test rotation
- judge disagreement metric
- visible skipped checks
- human review for unexplained score jumps
Boundary 3: novelty quarantine
Novelty can be valuable, but only after interpretation. Novel candidates should be archived before they are promoted.
Minimum controls:
- novelty descriptors
- archive before release
- weird-behavior labels
- human-readable rationale
- release hold until use case is bounded
Boundary 4: speciation scope
Specialists need stronger boundaries than generalists because they are easy to route around existing review.
Minimum controls:
- niche owner
- route policy for each species
- cross-species composition tests
- retirement criteria per niche
- deny accidental general-purpose use
Boundary 5: multi-parent provenance
N-way merges must never collapse into a friendly descendant name without parent records.
Minimum controls:
- parent hash list
- merge operator manifest
- adapterA small add-on that changes or specializes model behavior. Open glossary definition load-order record
- layer/source map where feasible
- pre-merge rollbackReturning a system to an earlier known state. Open glossary definition packets
Boundary 6: local and edge inventory
Local-first and browser-native execution must still produce inspectable state.
Minimum controls:
- local manifest export
- signed update channel
- adapter cache inventory
- memory/cache retention policy
- edge retirement receipt
Boundary 7: runtime residue
The runtime boundary includes more than prompts and model weights. It includes caches, worker state, GPU memory exposure, vector-store writes, tool logs, and synthetic output loops.
Minimum controls:
- cache partitioning
- memory diff review
- synthetic-origin labels
- session ownership
- residue test before shared execution
Boundary 8: dashboard honesty
A public or internal dashboard should not make weak evidence look stronger than it is.
Minimum controls:
- evidence label on concrete claims
- visible “what this does not prove” text
- skipped-check field
- independent-source links
- no-opThe decision not to change the system. Open glossary definition as a normal visible decision
Related pages
- ModelBreeder Risk Side
- ModelBreeder Risk Review Worksheet
- Evaluator Independence
- Rollback Completeness
- Apex Threat Implementation Controls
Cognivirus.com risk-side note
This page belongs to Cognivirus.com because it translates ModelBreeder-style possibility into risk review and governance controls.