A physical eval is, by construction, a public-facing physical system that gives partial control of real hardware to whoever holds the current slot.
把网络安全的威胁模型直接套到物理世界：开放评测等于把真实硬件的控制权租给不特定的人。作者列的时间片审计、动作空间沙箱、影子模式，本质是云安全和 bug bounty 的搬运。隐含前提是护栏能在动作空间层面完备定义——而 Goodhart 那节自己已经承认，指标达成与真实伤害可以并存。
Programming agents might work, but there are a lot of ways they can go wrong. The amount of guardrails necessary to keep the model in check doesn't scale. The more steps of a process rely on generative output, the higher the potential for error and catastrophic failure. Any process that involves these models must strive to maximize determinism and minimize model variance.
stay close to deterministic. n:: Guardrails grow faster than keeping stuff in check, doesn't scale
There’s an idea in the science-fiction community called steam-engine time, which is what people call it when suddenly twenty or thirty different writers produce stories about the same idea. It’s called steam-engine time Âbecause nobody knows why the steam engine happened when it did. Ptolemy demonstrated the mechanics of the steam engine, and there was nothing technically stopping the Romans from building big steam engines. They had little toy steam engines, and they had enough metalworking skill to build big steam tractors. It just never occurred to them to do it.
People aren’t encouraged to think about little further changes they might want to make, or naturally make tweaks as they go; instead, they just adapt their behavior to whatever the app encourages