Natural Language Inference (NLI) represents a foundational task within Natural Language Processing (NLP) that aims to determine the inferential relationship between pairs of sentences. While substantial progress has been made in English through the construction of large-scale, multi-genre resources, Spanish NLI datasets remain limited in both scale and genre diversity, and seldom capture explicit causal relationships. In this work, we introduce ESNLIR, a novel Spanish NLI dataset spanning multiple genres and explicitly annotating causal relationships. Leveraging automated extraction methodologies inspired by existing scientific NLI benchmarks, we construct an enriched corpus and evaluate baseline performance using BERT-based models. Our experimental results indicate that incorporating genre diversity enhances model generalization, and that explicit causal labeling reveals additional challenges for current architectures. ESNLIR aims to provide a new benchmark for Spanish NLI research and supports further investigations into model robustness and the mitigation of annotation artifacts. The complete dataset and code can be found in the following Zenodo repository: https://zenodo.org/records/15002575 .