We present a sharp test for the efficiency of job separations. First, we document a dramatic increase in the separation rate - 11.2ppt (28%) over five years - in response to a quasi-experimental extension of UI benefit duration for older workers. Second, after the abolition of the policy, the "job survivors" in the formerly treated group exhibit exactly the same separation behavior as the control group. Juxta-posed, these facts reject the "Coasean" prediction of efficient separations, whereby the UI extensions should have extracted marginal (low-surplus) jobs and thereby rendered the remaining (high-surplus) jobs more resilient after its abolition. Third, we show that a formal model of predicted efficient separations implies a piece-wise linear function of the actual control group separations beyond the missing mass of marginal matches. A structural estimation reveals point estimates of the share of efficient separations below 4%, with confidence intervals rejecting shares above 13%. Fourth, to characterize the marginal jobs in the data, we extend complier analysis to difference-indifference settings such as ours. The UI-indiced separators stemmed from declining firms, blue-collar jobs, with a high share of sick older workers, and firms more likely to have works councils - while their wages were similar to program survivors. The evidence is consistent with a "non-Coasean" framework building on wage frictions preventing efficient bargaining, and with formal or informal institutional constraints on selective separations.