Learning causal directed acyclic graphs (DAGs) from data is complicated by a lack of identifiability and the combinatorial space of solutions. Recent work has improved tractability of score-based structure learning of DAGs in observational data, but is sensitive to the structure of the exogenous error variances. On the other hand, learning exogenous variance structure from observational data requires prior knowledge of structure. Motivated by new biological technologies that link highly parallel gene interventions to a high-dimensional observation, we present $\texttt{dotears}$ [doo-tairs], a scalable structure learning framework which leverages observational and interventional data to infer a single causal structure through continuous optimization. $\texttt{dotears}$ exploits predictable structural consequences of interventions to directly estimate the exogenous error structure, bypassing the circular estimation problem. We extend previous work to show, both empirically and analytically, that the inferences of previous methods are driven by exogenous variance structure, but $\texttt{dotears}$ is robust to exogenous variance structure. Across varied simulations of large random DAGs, $\texttt{dotears}$ outperforms state-of-the-art methods in structure estimation. Finally, we show that $\texttt{dotears}$ is a provably consistent estimator of the true DAG under mild assumptions.