I think we probably agree, but we're talking about slightly different things. In the proposed experiment, the hypothesis is, "Green jelly beans cause acne.", a

I think we probably agree, but we're talking about slightly different things. In the proposed experiment, the hypothesis is, "Green jelly beans cause acne.", and the variable under test is "whether or not green jelly beans cause acne". The hypothesis is basically a statement of what variable is under test AND a prediction of what the results of the test will be. What I'm saying is that the prediction part of the hypothesis is basically a statement of bias. It's useful in noting what an interesting result would be and what the biases of the researcher might be, but it arguably is counterproductive in that it takes the focus off of observing the variable with an open mind. If we simply preregister the variable under test, that gives us everything we actually need from the hypothesis, and avoids the problem you describe: instead of testing 20 different variables and only publishing on the one that yields a low-P result, we test the single variable and that's it.