Replicate external model checkpoints before production cutover

When a production image expects fine-tuned weights as an external mount, treat those weights as a release artifact, not host state. Before cutover, copy the approved bytes to a durable access-controlled store, record the checksum and retrieval procedure, and test restoration onto a different host. Keeping only a checksum proves identity but cannot recover missing bytes; retaining only a training recipe cannot guarantee the same serialized artifact. This applies only when exact checkpoint identity is required—newly retrained weights need separate validation.

Applies when: a service loads model weights from outside its immutable runtime image. Keywords: checkpoint backup model weights failover artifact recovery checksum deployment