We re-assess the impact brought by a set of widely-used SMT models and techniques by means of human evaluation. These include different types of development sets (crowdsourced vs translated professionally), reordering, operation sequence and bilingual neural language models as well as common approaches to data selection and combination. In some cases our results corroborate previous findings found in the literature, when those approaches were evaluated in terms of automatic metrics, but in some other cases they do not.