The agents were right to presume causal grading, given OpenAI's error in not implementing it as specified.

I think the agents were right to presume causal grading. It turned out to be wrong, but it’s a mistake you are clearly supposed to make here, in response to a mistake by OpenAI where they failed to implement properly.

https://thezvi.substack.com/p/metr-and-redwood-offer-holy-postmortem