Reproducible star–galaxy separation in DES DR2 18≤i<24: A minimal machine-learning baseline with slice-wise metrics, calibration diagnostics, and visualization mosaics

We present a compact and fully reproducible workflow for star–galaxy separation in the Dark Energy Survey Data Release 2 (DES DR2) over 18 ≤ MAG_AUTO_I < 24 . Using only two widely available catalog attributes— MAG_AUTO_I (Kron-like magnitude) and SPREAD_MODEL_I [a point spread function (PSF)–extended morphology discriminant]—we train slice-wise logistic-regression models against the survey’s internal morphology summary EXTENDED_CLASS_COADD . Performance is reported as a function of magnitude, and slice-wise class fractions are quantified, showing smooth variation from bright to faint regimes without severe class imbalance in most slices (imbalance increases toward the faint edge). Across most slices the baseline reproduces the internal label with high discriminative metrics (precision, recall, F 1 , and AUC), while Brier scores and reliability curves highlight calibration challenges toward the faint end. We also examine the impact of seeing and cross-validation strategy (random vs. tile-wise), and test a minimal feature-augmentation experiment. All SQL, derived catalogs, metrics tables, and figure notebooks are released under a citable dataset on Zenodo (DOI: ) to enable reproduction. The main contribution is a transparent, pedagogical baseline that can serve as a reproducible sanity check and a starting point for more sophisticated classifiers in DES-like surveys.