Models, systems, and the messy space between theory and deployment.
Premise
Some problems do not get easier the more you study them. You learn more, certainly. But by the time you understand one, it is not the one you started on.
Complex adaptive systems produce emergent behaviour: cascading failures and tipping points. You see it in weather, in markets, in disease spreading across a subcontinent. The components are usually simple. The trouble is always in the interactions.
Rittel and Webber called a problem wicked when people cannot agree what it is, and every attempt at a solution changes it. You cannot solve one the way you solve an equation. There is no practice run, and what you do is part of what you are studying. We write differential equations as though populations are homogeneous, and then the forecast breaks because people changed their behaviour mid-outbreak.
Thompson et al.
Deployment tests what a system does when the model fails. Taleb calls a system antifragile when disorder makes it better. A forecasting model is not antifragile; bad data only makes it worse. The limit was rarely the model itself, more often whether anyone had designed for the person who had to act on it. The ones I have worked on that are still running were not the most technically ambitious. They kept working even when the data arrived late, in the wrong format, or not at all.
So far the world has been interfering with the model. The reverse is harder to admit. We build models of systems we are inside. A published forecast changes how people behave, which changes the case reports the model was fitted to. Sometimes it comes true. Other times it averts what it predicted, which is a success that scores as an error.
Model elegance loses to field reality. Facts uncertain, values in dispute, stakes high, decisions urgent: Funtowicz and Ravetz called that post-normal science. It is the ordinary condition of this work.