Responsible scaling policies are self-graded homework

A policy that the regulated party grades itself against is not a policy. It is a document.

Anthropomorphic's RSP, OpenAI's Preparedness Framework, DeepMind's safety commitments — I've read all of them carefully. They are genuinely well-written. Smart people put real hours into the language. That's actually the problem. The precision is in the commitments that look binding but aren't. Thresholds triggered by internal evals the public cannot audit. Escalation procedures that route back to the same leadership that approved the training run. Pause conditions written with enough qualifier density that a lawyer, not a researcher, decides whether they apply. The sophistication of the drafting is doing work — it is making the absence of enforcement look like a technical detail instead of a structural choice.

No external body can compel a pause. No third party audits the evals that determine whether a threshold was crossed. If a lab decides internally that the risk didn't quite meet the bar, there is no mechanism that disagrees with them. That's not a gap in the policy. That is the policy. The tell is always what the document doesn't commit to.