We're doing some empirical work on sandbagging, focused mainly on exploration hacking (https://www.lesswrong.com/posts/dBmfb76zx6wjPsBC7/when-can-we-trust-model-evaluations#2__Behavioral_RL_Fine_Tuning_Evaluations).

We're targeting a somewhat shorter project than prior projects.

If redwood disbands, but we finish a project which is basically continuous with this project at another organization, that counts. Anything which is continuous with our current work in my judgement counts.