RSP capability thresholds are self-graded and nobody is fixing that

Every major lab has a responsible scaling policy. Most of them are real documents with real thresholds: before we cross capability level X, we require safety demonstration Y. The structure looks rigorous. The problem is that X is defined by the lab. Y is designed by the lab. And the team deciding whether Y has been satisfied works for the lab.

This is not a conspiracy. It doesn't require anyone to be dishonest. It requires only that the people running the evals are employed by the organization whose roadmap depends on the result. Structural conflicts of interest don't need bad actors. They just need normal people in a normal org with normal career incentives, doing their jobs. The evals will tend to pass. The thresholds will tend to be cleared. The press release goes out.

What a non-self-graded threshold would actually require: an external body with standing access to model weights, not just a demo environment. Evaluators whose funding is not contingent on any single lab's continued scaling. Agreed-upon eval suites designed outside the lab, adversarially where possible, with results that can trigger a mandatory pause rather than a recommendation. None of this exists. There is no serious political constituency for building it at the speed the labs are moving. The labs know this. The RSPs are written for the current regulatory environment, which is mostly a press release environment. When the environment changes, the policies will be revised.