OrbitalProof · AI agents · for a technical reader
Crash recovery for AI assistants’ multi-step changes: the lab’s exact wording
The lab’s own sentences and figures for this result, word for word, in its working terms, with its limits in full. The plain account is on the result’s page; it says the same things in plain words.
The lab’s plain-English sentence
When an AI assistant carries out a multi-step change to files and one database table, the lab's transaction layer is meant to either complete all of it or undo all of it — but only for plans it admits in advance
- which may contain at most one action that cannot be undone (such as sending a message), placed last and carrying a replay key, other plans being refused or left for a human operator
- killed at 29 moments the authors chose, the program each time recovered a clean all-or-nothing state from its on-disk log in a fresh process, while a deliberately broken version did not
- but killed 2,893 more times at random moments it recovered 31 times to a state that was neither, some of them undoing work already committed
At the moments the lab chose, in its words
atomic recovery recorded at every one of those 29 points
The limit to read first, in the lab’s words
The 2,893 random kills are process kills, not power cuts, so the fsync ordering the journal relies on is still untested
Kills at random moments
2,893 further kills at seeded random moments found 31 recoveries that were not atomic
Recorded in the lab’s claim; the run record is not published.
At the moments the authors chose
atomic recovery recorded at every one of those 29 points and a negative control that is genuinely non-atomic
The lab’s prior-art search concludes
So this is INCREMENTAL, a real test of one more system, done to less than the published standard
Claim
Twenty-nine real processes killed with `os._exit(9)` at declared interruption points, each recovered in a fresh interpreter from the on-disk journal, with atomic recovery recorded at every one of those 29 points and a negative control that is genuinely non-atomic; 2,893 further kills at seeded random moments found 31 recoveries that were not atomic, some undoing committed work.
Limits
The 2,893 random kills are process kills, not power cuts, so the fsync ordering the journal relies on is still untested; they sample the interleaving space (85 distinct crash states) rather than exhaust it; and the three windows they found are recorded, not fixed.
More
- The plain account of this result: the problem, what was shown, what it means for a buyer, who we expect would buy, why now, why you can trust the check, and what it does not show yet.
- This result’s file, the source of every sentence above.