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.
Retirement Is Not Extinction
A retired model can leave behavior behind in memory, synthetic data, descendants, adapters, routes, evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition preferences, and operating procedures.
Mechanism
Variation, evaluation, selection, inheritance, and succession can exist as properties of the broader development process. The model does not need to rewrite itself at runtime. The ecology changes because operators, pipelines, routers, and release controllers alter the population.
Assurance implication
A descendant needs fresh evidence for safety-relevant behavior. A content hash can identify an artifact, but it cannot prove that a related descendant preserved all relevant guardrails.
Review question
What behavior is being tracked, where could it be encoded, which descendants or reservoirs may carry it, and what evidence would count as absence across active compositions?
<!-- expanded-release-content -->
Artifact retirement versus behavior retirement
Artifact retirement is a registry event. Behavior retirement is an ecological claim. The first can be proven with a signed release record. The second requires evidence that the behavior is no longer expressible through active artifacts, descendants, memory, routes, retained datasets, adapters, evaluator expectations, or operational practices.
A model can be removed from production while its outputs remain in a training set. A specialist can be replaced while an adapterA small add-on that changes or specializes model behavior. Open glossary definition trained from it remains active. A route can stop invoking the old model while continuing to prefer descendants that imitate its style. A memory system can preserve decisions, summaries, or preferences that were originally produced by the retired artifact. A human workflow can also carry the behavior when operators keep copying outputs that once scored well.
Behavioral residue
Behavioral residueInformation or tendencies left in memory, synthetic data, traces, evaluator preferences, or subsequent training material after a component is retired. Open glossary definition is not necessarily malicious. It may be a useful phrase pattern, a domain heuristic, a compliance convention, or a formatting style. The risk appears when the residue is treated as absent merely because the first carrier is gone. Residue becomes safety-relevant when it affects refusals, tool use, user classification, prioritization, recommendation, or evaluator judgment.
Evidence for extinction
A behavioral-extinction review should name the behavior, list all plausible carriers, list retained data and memory reservoirs, identify descendants and merge products, test active routes, review evaluator preferences, and examine whether future training pipelines can reintroduce the behavior. The review should be dated, versioned, and scoped. “No longer observed in a sample” is weaker than “not expressible across defined active compositions under defined probes.”
Control implication
Retirement should produce a packet: artifact hashes, descendants, memory snapshots, synthetic-data references, route rules, aliases, evaluator versions, tool profiles, external side effects, and known residual reservoirs. Without that packet, rollbackReturning a system to an earlier known state. Open glossary definition and extinction claims are partial.