I came here to say that one of the exceptions would be architecture and construction, or any undertaking where redoing later is way more expensive than foreseei

I came here to say that one of the exceptions would be architecture and construction, or any undertaking where redoing later is way more expensive than foreseeing now. Science is the art of collecting data and factoring out subjectivity as much as possible. So, ideally, "doing" is the only thing (non-theoretical) science should be doing. As for software construction, I don't think it's controversial to say that the "head-first" methodology is dangerous.