Preregistered analysis plan and dataset-selection screening for "Directional Hull Summaries for Budgeted SVM Training: Coverage Limits and Reconstruction-Guided Repair"

Abstract

This record registers the analysis plan for an expanded experimental study, together with the code and screening output that fixed the dataset suite. It was deposited before any subset selector was executed on that suite, and it contains no experimental results. The first round of experiments had two design faults, both identified by the authors. First, with six datasets the smallest attainable two-sided Wilcoxon signed-rank p-value is 2/2^6 = 0.031, while Holm correction over four contrasts requires 0.0125; power was therefore exactly zero for every effect size rather than merely low. Second, the repair mechanism under test was selected using the same six datasets on which it was subsequently evaluated. The registered plan addresses both. Datasets are admitted by a rule computed from a full-data linear reference fit alone - a size floor, convergence of the reference fit, a discriminative-headroom band on balanced accuracy, and a class-balance floor - so the rule never observes a selector and the suite cannot be tuned toward an outcome. Candidates failing the headroom criterion are retained in a reported SATURATED stratum rather than discarded. Screening admitted 13 datasets to the primary stratum, 4 to the saturated stratum, and excluded 1. The plan records why the suite was not padded to the intended target of 15: the remaining candidates are subsets of benchmarks already included, and counting them would constitute pseudo-replication. The deposit also contains the design analyses that motivated the redesign: attainable p-values and simulated power by sample size, the total sensitivity of the counterexample family used in the manuscript's scope result, and a base-rate-adjusted measure of how strongly directional sampling favours high-sensitivity points. Benchmark datasets are not redistributed. Retrieval URLs and checksums are included. Code is released under the MIT Licence; documentation, tables and logs under CC BY 4.0.