An underappreciated failure mode in iterative AI repair: an LLM's second attempt at fixing a workflow can genuinely be worse than its first, introducing a new error while resolving the original one. A naive healing loop that always keeps "the latest attempt" will happily ship that regression.
Our sandbox simulator tracks every attempt made during a healing run and compares the final result against the best one seen — not just the most recent one. If the last attempt is worse than an earlier one, we revert to the better version and report that honestly, rather than presenting a regression as progress.
This sounds like an obvious safeguard once stated, and it is — but it's the kind of thing that only becomes obvious after watching an AI confidently make something worse. Any iterative AI repair system that doesn't track this is quietly at risk of shipping regressions as if they were fixes.
