In the 1990s, the federal government commissioned a bunch of randomized controlled trials of state-level welfare to work initiatives, during the transition between AFDC and TANF. Not all of them, but a great number, showed that the welfare recipients who were given strong incentives to get jobs and get off of welfare did in fact do so, and even that some of the training and support programs made a difference (beyond just telling people you were cutting off their check if they didn’t get a job.)
Then, in the 2000s, the federal government commissioned additional studies of both the population that had gotten off of welfare but was still near the poverty line and the population that was still on TANF. These studies (ERA for example) were a textbook case of the Iron Law— almost as many programs hurt outcomes as helped them, despite the incentives for researchers to find success rather than failure wherever they look.
Aside from the fact that the economy was growing much faster in the mid to late 90s than it was in the early 2000s, facilitating people finding jobs who genuinely wanted them, what made the difference? There were really two kinds of diminishing returns that the government ran into. First, the people who were easier to get off of welfare were already off. Second, the obvious changes to welfare to make it hard to stay on it forever were already made.
One of the problems conceptually with “evidence-based policy” is it tends not to admit of diminishing returns. The assumption is never that we might be at a local optimum, or that improving one aspect or another of our collective life might be simply not worth the cost.