Science Sandboxes for AI Agents

From the Broad Institute of MIT and Harvard (authors include Sangeeta Bhatia, Pardis Sabeti's and Eric Lander's orbit): a framework arguing that scientific capability is not finding solutions but learning the rules that explain them through repeated cycles of experimentation, feedback and hypothesis revision. Science sandboxes let an agent query the natural world at three levels of empirical verifiability — "wet" physical experiments, "damp" predictive models trained on empirical data, and "dry" invented rules — under one common experimental loop and evaluation protocol covering both quantitative metrics and qualitative scientific reasoning.