Artifacts for: An ablation measured twice -- small component effects can change sign across repeated training runs in breast-cancer histopathology segmentation

Run artifacts, analysis code, frozen data splits and manuscript source for a study that runs a seven-configuration segmentation ablation on BCSS twice and measures which component contrasts survive retraining. Contents. 77 per-run directories (21 second-measurement ablation runs, 24 runs forming twelve same-configuration replication pairs, 15 baselines, 6 full-data runs, 11 single-seed pre-experiments), each with resolved configuration, environment capture and test metrics, with per-patch predictions where they survive; the generated tables and figures; the frozen train/validation/test split identifiers; the environment lock file; the code and the shared statistics package it depends on; and the manuscript source. Verification. SMCA-HistoSegNet/paper/mbec/verify_paper.py reads every reported number out of the manuscript source and re-derives it from the tables here. It runs unmodified from the extracted archive. Not included. The first ablation measurement's per-patch predictions, lost before reaching version control; its inferential statistics were committed before the loss and are included as summary files, which the manuscript states. Model checkpoints, which are not needed for any reported number. The BCSS images, redistributed by their original authors (Amgad et al., Bioinformatics 2019, doi:10.1093/bioinformatics/btz083). On "reproduce". These runs are not bit-reproducible, which is the paper's subject. Reproducing the paper means regenerating every reported number from these artifacts, not recovering identical weights.