Systematic robustness of a boosted top-quark tagger on ATLAS open data: a constituent-level feature reconstruction, paired SHAP-Delta attribution drift, and conformal coverage

The ATLAS top-tagging open dataset releases detector-systematic and generator-modelling variations as dedicated simulated samples rather than as per-event weights. For the four exactly pairable calorimeter variations, the fifteen stored high-level substructure quantities remain nominal, so we recompute all fifteen from released constituents with one implementation applied to every sample. Against the stored nominal branches the median relative agreement is 2.5×10⁻⁸ to 4.8×10⁻⁷ per quantity on 14,103 jets, and 153 of 211,545 quantity–jet comparisons (0.072%) exceed a 10⁻³ relative tolerance. These fall on 44 jets: 43 lack, in the released constituents, information the stored branch used, and one is a reproducible floating-point degeneracy; none is a formula or implementation mismatch and none is unresolved. The frozen reconstruction was applied to 80,481,421 jet entries. An XGBoost tagger developed from a five-million-jet nominal subset reaches an unweighted area under the receiver-operating-characteristic curve of 0.9320 and background rejection 41.91 at 50% signal efficiency on 10,306,812 test jets, with a 25.8% total relative systematic uncertainty dominated by generator modelling (20.0%). Paired TreeSHAP attribution drift, measured on the four calorimeter-cluster variations that pair jet-by-jet, is led by ECF₃, while global importance orderings change minimally (Spearman ρ ≥ 0.986). Class-conditional split-conformal coverage, calibrated once on nominal simulation and applied unchanged to all nine detector-systematic samples, stays within one percentage point of the 90% target.