Selection against deleterious mitochondrial mutations is facilitated by germline processes, lowering the risk of genetic diseases. How selection works is disputed: experimental data are conflicting and previous modelling work has not clarified the issues. Here we develop computational and evolutionary models that compare the outcome of selection at the level of individuals, cells and mitochondria. Using realistic de novo mutation rates and germline development parameters from mouse and humans, the evolutionary model predicts the observed prevalence of mitochondrial mutations and diseases in human populations. We show the importance of organelle-level selection, seen in the selective pooling of mitochondria into the Balbiani body, in achieving high-quality mitochondria at extreme ploidy in mature oocytes. Alternative mechanisms debated in the literature, bottlenecks and follicular atresia, are unlikely to account for the clinical data, because neither process effectively eliminates mitochondrial mutations under realistic conditions. Our findings explain the major features of female germline architecture, notably the longstanding paradox of over-proliferation of primordial germ cells followed by massive loss. The near-universality of these processes across animal taxa makes sense in light of the need to maintain mitochondrial quality at extreme ploidy in mature oocytes, in the absence of sex and recombination.