Skip to content
replicapong

I'm building a system

How do I make sure nothing gets lost when the AI revises a file?

Facts as of September 16, 2026

With a verification chain: the AI creates the new version, reads it back and compares it by machine with what was supposed to arrive. Only when both are identical does the old version go to the trash – never before, never permanently. The comparison rests on a checksum, not on the eye.

In detail

The chain has six steps:

  1. Fetch the file unchanged, byte for byte.
  2. Build the new version in a working copy.
  3. Create it anew under the same name.
  4. Read back what was actually stored.
  5. Compare by machine.
  6. Only once they are confirmed identical, move the old version to the trash.

The last step is the only one that destroys anything, and that's why it's the only one with a condition in front of it. Two things easily get cut: read back – what counts is not the text that was sent but the text that was stored – and to the trash, so that even a passed comparison can still be undone.

Why not by eye? Because the dangerous errors are invisible. A single swapped character turns “Bühne” into “Dühne,” with the same length and valid encoding. A swallowed blank line at the end costs one byte, and nothing about it looks wrong. The byte count is an early warning that costs nothing; the proof is the checksum. That two files are the same length or look the same doesn't count.

Before you have a text of your own revised, store the original version separately. Without it, you can later neither check what was changed nor undo anything, see How do I keep track of which part of a text is mine?

What the chain doesn't do: it checks sameness, not correctness. A wrong text arrives wrong, byte for byte. And its weakest point is the check itself, when long strings of characters are copied out by hand along the way – that's exactly where the errors you're looking for arise. The machine should do the comparing, not the retyping.

To copy

Discard the old version only after you have read back the new one and
compared it by machine with the source, and both are identical. Then move
it to the trash, not permanently. If they are not identical, leave both
in place and report it.

In practice

On August 17, 2026, a rules file was written back on Thomas J. Weiss's system, and Umkehrbar arbeiten (“work reversibly”) had become UmkehrBAR arbeiten. Nobody had typed that; it had happened along the way. Had the old version no longer existed by then, the damage could no longer have been detected.

Since then the chain runs on every replacement. On two days it found more than fifteen errors, every single one introduced in the copying, none found by eye. Once it even found old damage: an invalid byte in place of a line break, left over from an earlier session, which could have sat there for years.

Sources

  • Note „Die Prüfkette“, replicapong.de, 16.09.2026
  • Prompt „Erst nach bestätigter Gleichheit verwerfen“, replicapong.de, 16.09.2026
  • Prompt „Die erwartete Byte-Zahl mitgeben“, replicapong.de, 16.09.2026
  • Review of the rules of a theater project, 02.09.2026

Cite: Thomas J. Weiss: “How do I make sure nothing gets lost when the AI revises a file?” replicapong.de, as of September 16, 2026. https://replicapong.de/en/question/safe-file-replacement/ – CC BY 4.0

Related