Data and artifacts for "Explaining El Niño Forecasts with the Average Gradient Outer Product"

Processed data, trained neural-network checkpoints, analysis artifacts, numerical source data, manuscript Figures 3–12 and Table 1, timing benchmarks, and the standardized-ocean-current animation supporting “Explaining El Niño Forecasts with the Average Gradient Outer Product.” Version 1.0.1 applies the common-reference correction: AGOP matrices use the same 360,000 input states from the 10,000-year training block at every forecast lead, independently of the lag-trimmed supervised training pairs. The trained models, normalization, and processed simulation are unchanged. The archive includes the corrected explanations, scores, figures, and physical nudging results, with provenance and checksums. The processed data comprise a 12,000-year Zebiak–Cane ENSO simulation sampled three times per month. The manuscript models use SST anomaly, thermocline depth, zonal and meridional ocean currents, plus two annual-phase coordinates; the chronological split is 10,000 years for training, 1,000 for validation, and 1,000 for testing. Use this record with the BSD-3-Clause source release at https://github.com/rjwebber/agop-xai/releases/tag/v1.0.1. The simulation archive retains its v1.0.0 filename because it is unchanged. Follow README-ZENODO.md to assemble any split archives and unpack the required packages at the repository root. The main artifact package includes the saved native starting states. A separate optional package provides the losslessly packed covariance needed to replay the physical nudging calculations. The private legacy input, downloaded upstream Zebiak–Cane source, and executable binaries are not included. Except where third-party provenance states otherwise, these data and artifacts are released under CC BY 4.0.