The goal of this project is to study the modes of reasoning that text classification systems learn from training data, and the interaction between properties of the training data and properties of the trained classifier. The Recognizing Textual Entailment (RTE) task has been a staple in natural language processing for over twenty years. The RTE entailment relation is not formally defined, and its interpretation has been contested. In this project, we propose different formal readings of the RTE entailment relation, based on its description in the RTE task guidelines. We perform a comprehensive analysis of the meta-inferential properties of these relations and compare them to other well-defined entailment relations. We use this analysis as grounds for experimentally investigating to which degree the entailment relation encoded by the RTE dataset matches the proposed formal readings of the relation. Lastly, we perform a comparison with our previous study of the Stanford Natural Language Inference dataset, essentially striving for a better understanding of the different aspects of entailment these datasets seem to capture. We use datasets from the RTE 3–5 challenges, that are annotated according to a three-way classification scheme. To be able to analyse the meta-inferential relations and conduct a comparison with previously studied entailment relations, we need to generate new RTE-style sentence pairs that form implicational chains in order to be able to test transitivity and similar properties of the learned entailment relation. Such data is not available in the original RTE datasets. The results of this project will be presented as an invited contribution to the CLASP Concluding Conference (https://gu-clasp.github.io/concluding-conference/; October 5–6 2026) with an extended version published in a special issue of the NEALT Proceedings series (final version due March 15 2027).