Superalignment

Not a score, not a passing test suite, and not the model saying it is confident. Evidence is a record of something that actually happened, attached to the claim it supports, in a form somebody else can go and check. It is the load-bearing word on this whole site: when we say a system has earned the right to act, we mean there is evidence for the specific thing it is about to do.

Why it matters

It is the only thing that survives the question “how do you know?”