I think you're saying essentially this, but 1 thing that helps is we're not designing in total ignorance -- it's all iterative, so we're building on what we've already learned from our prior studies & from others. The specific result may be less than certain but we know a lot to inform the design.

In physics, they often have more precise expectations (e.g. the mass of a Higgs boson), and so the test is looking for one or more specific values to validate the theor(ies). Ours may often be fuzzier (e.g. directional, not points), but similar nonetheless. And maybe we should aim for 5 sigma too?