Danger ModelReasoned from system designv1.15.0

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, Returning a system to an earlier known state. Open glossary definition, and retirement.

Good varianting records more than weights

Evidence levelReasoned from system designTechnical label: Architectural inference

A useful lineage record names the base model, adapter stack, prompt package, tokenizer, data snapshot, training recipe, inference configuration, quantization, router version, The exact version of the evaluator used for a test or release. Open glossary definition, tool permissions, environment, release alias, approval ticket, and UTC timestamps.

The 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:

What to watch for

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.