CausalVisionRI: Derived data and evaluation record for 24-h rapid-intensification warning in the western North Pacific

This record provides the derived data and evaluation artifacts accompanying the manuscript “Causal-prior-constrained deep learning for rapid-intensification warning of western North Pacific tropical cyclones.” CausalVisionRI integrates satellite infrared imagery, atmospheric and oceanic environmental fields, and an observationally discovered causal graph for 24-h rapid-intensification warning. The evaluation uses a forward split: 2000–2017 for fitting, 2018–2019 for model selection, and 2020–2024 for the final holdout evaluation. The deposit comprises two ZIP packages. record1_core.zip contains the processed analysis panel with split markers and row-level keys, imputation information, causal-discovery and effect-estimate bundles, sensitivity analyses, satellite-frame indices, SHIPS-12 predictor tables, frozen test-period predictions, analysis outputs, figures, and protocol-lock materials. record2_ritcf_recon.zip contains the RITCF reconstruction sample frame, pair tables, normalization statistics, dataset manifest, run audit trails, operating-point sweeps, and summaries. Each package includes an inventory and a SHA-256 manifest. Extract both archives into the root directory of the accompanying CausalVisionRI code release. The released predictions and analysis products support recomputation of the reported metrics without retraining. The code repository provides installation, verification, data-acquisition, and reproduction instructions. Raw ERA5, GLORYS12V1, IBTrACS, and HURSAT-B1 data are not redistributed. The large satellite-frame store, trained model weights, and multi-GB RITCF caches are also excluded; the code repository documents how to acquire or rebuild them. The derived data are released under CC BY 4.0, and the accompanying code under the MIT License.