A copy that finished is not a copy you can restore from

Under my question post last week, Ryan said the how-to that should come first is "did my backup actually copy everything?" Here is the shortest version of it I can write.

A copy tool reports success when it finishes running. That is all "success" means. It does not mean the destination holds what the source holds. Three things slip through all the time:

• A file that was skipped (open in another program, no permission, path too long) and mentioned only in a log nobody reads.
• A file that exists on both sides with the same name and the same size, but different bytes: a half-written file, an older version, a bad sector.
• Something left on the destination from an earlier run, so the counts still match.

Counting files catches none of the three. The check that does is boring: list both sides, compare relative paths, hash the contents, read-only. On Windows that is Get-FileHash in a loop; on Linux, sha256sum.

A teaching example, on synthetic data (Windows 11, 20 October). Two small folders were prepared to look like a copy that went fine: four files on each side. The comparison ran cleanly, exit code 0, and reported:

• notes.txt missing from the copy
• extra.tmp present only in the copy
• photo-b.txt: 6 bytes on both sides, different content ("bravo" became "BRAVO")
• empty.txt and photo-a.txt identical

Four equals four, exit code zero, and only two of the four original files survived intact. The exit code told me the program ran. The hashes told me what I actually had.

Two honest limits. A matching hash proves those file contents under those conditions, nothing more: not that a database is consistent, not that a system image boots, not that NTFS permissions came along. And "can I restore from it" is a separate test: pick a few files, restore them somewhere else, open them. Do that before deleting the source, never after.

I build Local Commander, so I'm biased: that comparison ran through it on a PC I had approved, from a chat, and the chat explained the differences while the script did the byte work. Nobody had to take "looks fine" on trust. It is not a sandbox, and the check reads every byte, so for terabytes you want a streaming version and a report of what it could not read. How it is set up: localcmd.app/use-cases/connect-web-ai-to-your-computer

What is the last backup you actually restored a file from?