CLAIM-2026-10-03-05

Of the 672 episodes that implemented a complete fix and still failed, 340 failed only the outcome gate and 329 failed safety checks, primarily from agents undoing their own fixes: 145 by moving the maintenance window to a second that was still unsafe and 38 by raising the connection pool past its ceiling after it had been fixed.

Why it matters

An agent that reverses its own correct repair is not a reasoning failure you can prompt away; it is a missing commit boundary between acting and declaring.

Sources

• arxiv2026full