In contrast, if you are in machine learning and you are extending an existing architecture you are very directly dependent on that original technique being useful. If it doesn't "replicate" the effectiveness of the original paper, you're going to find out quickly. Same for algorithms research. Some other comments here have mentioned life sciences being the same.
So I think there's a qualitative difference between sciences where we understand things in a mostly statistical way (sociology, psychology, medical studies) where the mechanism is unknown (because it's very very complicated), but we use the process of science mechanistically to convince ourselves of effectiveness. e.g. I don't know why this color makes people work faster/ this drug increases rat longevity / complex human interactions adhere to this simple equation, but the p value is right, so we think it's true. Versus sciences where we have a good grasp of the underlying model and that model is backed up by many papers with evidence behind it, and we can make very specific predictions from that model and be confident of correctness.