Cost-effectiveness of student programs for AI safety research

Similar to my comment on the other post [LW(p) · GW(p)]: it seems like this critically relies on guesses about the 'treatment effects' of these programs, as detailed in the Pipeline probabilities and scientist-equivalence [LW · GW] section. How did you come up with these guesses?

A meta-thought: you published several posts on related topics today and I'm not sure what order I'm supposed to read them in, or how they relate to each other. (this doesn't come with an immediate suggestion for what to do differently since I'm not sure what your goal is, just flagging I felt a bit overwhelmed deciding whether/which to read)