In plain English
This page is part of the technical reference. It keeps the expert detail but starts with a plain-language summary for first-time readers.
- 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.
Controlled Varianting, Lineage, and Rollback
Direct answer
Creating model variants is not inherently unsafe. It becomes unsafe when variant generation outruns lineage, evaluation evidence, approval workflow, rollbackReturning a system to an earlier known state. Open glossary definition, and retirement.
Good varianting records more than weights
A useful lineage record names the base model, adapter stack, prompt package, tokenizer, data snapshot, training recipe, inference configuration, quantization, router version, evaluator versionThe exact version of the evaluator used for a test or release. Open glossary definition, tool permissions, environment, release alias, approval ticket, and UTC timestamps.
LineageThe parent-child history of models, adapters, datasets, or releases. Open glossary definition is not identity. It tells reviewers where an artifact came from. It does not prove which behavior was inherited.
Rollback is ecological
Rollback cannot be only “restore the old model.” It must restore or account for:
- model artifact;
- adapter stackA set of adapters loaded together, usually in a defined order. Open glossary definition and load order;
- prompt-policy package;
- memory snapshotA saved state of what the AI system remembers. Open glossary definition;
- retrieval index;
- router version and traffic weights;
- evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition version;
- permission profile;
- tool configuration;
- release alias;
- generated synthetic data;
- external side effects.
What to watch for
- model alias stays the same while implementation changes;
- missing training-data snapshot;
- adapterA small add-on that changes or specializes model behavior. Open glossary definition load order not recorded;
- rollback rehearsed for weights but not memory;
- no owner for evaluator updates;
- released variant lacks a retirement plan;
- experiments promoted by dashboard metric alone.
Expert review note
Varianting should behave like safety-critical CI/CD: immutable artifacts, signed records, explicit stages, canaries, shadow deployment, failure thresholds, rollback drills, and independent review.