Persisted test-set probabilities for "Classical Class-Imbalance Remedies Do Not Transfer to Tabular Foundation Models"

The 8,600 persisted validation and test probability arrays behind the family × remedy interaction study: one .npz per (dataset, base model, remedy, seed) cell, each holding validation and test probabilities together with the corresponding labels. These are what make the post-hoc arm of the study free to reproduce. Every threshold rule, Platt and temperature calibration, and class-conditional conformal result in the paper is recomputed from these arrays offline, without refitting a single model — and so is any metric the paper did not report. Split into parts for transfer; extract every part into results/ and they merge into a single results/proba/. The code that consumes them is in the companion software record.