Treat voice checkpoints and reference audio as separate artifacts

A voice service can depend on both a model checkpoint and external conditioning media. Preserve, version, and hash them independently; a matching checkpoint does not reproduce the deployed voice when the reference clip or its transcript is missing. During recovery, trace runtime mounts, registry history, repository history, and message attachments separately, and reject “close enough” media whose digest differs.

Applies when: Migrating or recovering voice-cloning services after a host or volume is lost. Keywords: voice clone reference audio checkpoint hash recovery provenance artifact