A central requirement in asymmetric quantum nonlocality protocols, such as quantum steering, is the precise reconstruction of state assemblages—statistical ensembles of quantum states correlated with remote classical signals. Existing steering works often rely on simplifying assumptions about detection efficiency and photon loss. Here we introduce a generalized loss model for assemblage tomography that uses conical optimization techniques combined with maximum-likelihood estimation. This approach allows us to accurately estimate assemblages without assuming uniform detection efficiency on the untrusted party's side. Using an evidence-based framework grounded in the Akaike information criterion, we demonstrate faithful reconstructions while balancing model complexity. We validate our results through numerical simulations and an experimental setup, showing robust performance in assemblage estimation when applied to experimentally relevant data.