Even for that, though, if you define your criteria for success, it's mostly possible to measure the outcomes and learn whether your approach was successful or not. One key there is remembering that subjective metrics are valid - user happiness and developer happiness are two of the most useful ones I know of for measuring the system's overall health and effectiveness, and while they're not objective, they can still be measured reasonably quickly and regularly.
For fields like public policy, though, there are just so many confounding factors and the timeframes are so long that it becomes nightmarish, maybe even impossible, to genuinely learn anything. The subjective metrics of citizen happiness and functionary happiness that are analogous to my software ones above are worthless because of rollout times measured in years or decades and the infinite array of factors, confounding and otherwise, that go into people's satisfaction with society.
Even relatively straightforward things like farming have learning cycles measured in years for things like crop rotations.
The universe is just not a friendly learning environment.
Programming is an aberration in almost every respect, as a product of the human mind that can actually have useful impacts in the world (most friendly environments are human constructs that are not "useful" per se, like board games or musical performance or sports).
That's a sign something is wrong.
Constructing valid experiments in software development is incredibly easy. Add an assert, recompile, and see if it gets triggered.
We programmers consider it unacceptably painful if testing a hypothesis takes ten minutes.
Real life is never that clean or easy.