Raj runs five experiments against a single Cloud Volume and refuses to treat a successful API call as proof of anything. Every artifact is verified by byte length and SHA-256 read back through the volume, not from inside the sandbox that wrote it. Sandbox A writes stage one and is terminated, Sandbox B mounts the same volume and carries on, a permanent snapshot preserves the known-good tree while the live filesystem moves ahead, and a snapshot-pinned read-only mount rejects an attempted write with EROFS. The line to keep is his own: replace the compute, preserve the workflow state. The guardrails section earns its place too, in particular that volume changes publish asynchronously and that concurrent writers still need a path each rather than a lock.
We didn’t write this one — it’s Raj Kumar’s piece, published on Medium. The note above is ours; the full article is theirs.