Can outside evaluators actually outpace the labs they watch?

The first step of the pacing idea sounds clean. Put independent evaluators inside the frontier labs. Let them catch dangerous capabilities as they emerge, before the lab talks itself into shipping anyway.

That only works if the evaluators can see what the lab's own researchers see, and see it sooner or more honestly. Here is the problem nobody really closed. The people who understand the next jump in capability are already inside. They have the weights, the traces, the failed runs, the late-night notes. An outsider who is systematically less informed is not a check. It is a reporting layer with a nicer title.

You can embed people. You can give them badges and a mandate. You cannot give them a smarter model of the thing than the team that trained it, not on a schedule that matters. If the lab's best researchers miss a failure mode, or decide it is fine, the visiting evaluator is not magically going to be the one who notices first.

Wait, so the whole "we will watch ourselves with extra eyes" pitch depends on the extra eyes being better than the people who built the system?

Yes. That is the gap. Oversight that cannot outrun the frontier it is meant to monitor is theatre with better stationery. It may still be worth doing. It is not the same thing as solving the problem. Until someone explains how an embedded team stays ahead of the lab that employs the actual frontier researchers, this part of the plan is a hope dressed up as a process.