FYI, the "Evaluating Alignment Evaluations" project of the current AI Safety Camp is working on studying and characterizing alignment(propensity) evaluations. We hope to contribute to the science of evals, and we will contact you next month. (Somewhat deprecated project proposal)