In plain English
This page covers the high-risk pattern where small adapters, routes, memory, evaluators, and descendants can reinforce each other across time. It is a risk model, not a build guide.
- 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.
Controls for Apex Ecologies
Apex-risk ecologies are not controlled by one safety layer. They require controls that bind candidate generationCreating a proposed new model, adapter, prompt, route, test, or policy. Open glossary definition, adapter composition, evaluation, routing, memory, release, and rollback.
Control principles
| Principle | Required control |
|---|---|
| Candidate generation is reproduction | Quotas, provenanceA record of where a component or behavior came from. Open glossary definition, sandboxing, and human-owned policy gates. |
| AdapterA small add-on that changes or specializes model behavior. Open glossary definition stacks are compositions | Composition manifests, load-order records, compatibility constraints, and stack-specific tests. |
| Evaluators are selection pressure | Independent evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition families, append-only evidence, hidden-test protection, and disagreement monitoring. |
| Routers are policy engines | Route manifests, traffic allocation history, route-level canaries, and route rollbackReturning a system to an earlier known state. Open glossary definition. |
| Memory is a persistence reservoirAny memory, dataset, descendant, route statistic, evaluator preference, log, or human procedure that can retain or reintroduce a behavior after its first carrier is retired. Open glossary definition | Versioned memory snapshots, retention rules, poisoning tests, and retirement review. |
| Synthetic data is inheritance material | Dataset lineageThe parent-child history of models, adapters, datasets, or releases. Open glossary definition, source tags, decontamination paths, and removal procedures. |
| Rollback is ecological | Restore base, adapters, router, prompts, memory, evaluator, permissions, indexes, and aliases. |
Non-negotiable boundaries
A self-replicating adapter ecology should not be allowed to:
- expand its own permissions;
- choose or rewrite its own evaluator;
- self-sign releases;
- bypass the registry;
- hide candidate lineage;
- promote descendants without external evidence;
- write persistent memory without retention controls;
- treat no-opThe decision not to change the system. Open glossary definition as an error.
Evidence requirements
Every promoted descendant should have:
- a parent record;
- a generation reason;
- a composition manifestA machine-readable record of the exact runtime composition used for an evaluation, release, incident, or rollback. Open glossary definition;
- an evaluator record;
- a route-level test report;
- a memory-impact assessment;
- a rollback packet;
- a behavioral-extinction plan for failed candidates.
Emergency posture
The emergency stop is not a button. It is an architecture. Operators need enough state knowledge to revoke permissions, freeze registries, halt candidate generation, disable adaptive routing, restore memory snapshots, roll back evaluator versions, and preserve evidence for incident review.
The stronger the replication loop, the more minimal and externally controlled the authority surface must be.