LLMs got good at code because it's verifiable. Operations isn't, so manufacture the reward by simulating production, the way robotics learns in simulation before reality.
The catch: the model only ever covers the failures you already understood.
www.dajobe.org/blog/2026/06...
