Simulated scores and code for Local information recovery by matching-averaged AUC e-values
Description
Simulated scores and preserved code accompanying "Local information recovery by matching-averaged AUC e-values". The frozen diagnostic uses two continuous score models (overlap_uniform and row_gate), balanced per-class sample sizes 64, 256 and 1024, baseline and local-alternative sampling laws, and 64 repetitions per condition: 12 conditions and 768 datasets. PCG64 seed 20261021 generated 688128 uniform draws; no models were fitted. The archive contains raw score arrays, per-repetition metrics, all condition summaries, frozen protocol, code and dependencies, environment and hash manifests, and a bounded computational audit. All 768 direct-metric and summary records were checked; matching moment-generating functions were independently recalculated for 48 predetermined rows, while 720 were not independently recalculated. Noisy, nonmonotone null-law likelihood-distance results are retained. The source uses original absolute paths and exclusive output creation; extraction alone is not a verified independent-machine reproduction. All data are simulated. AI research assistance and internal checks do not constitute external peer review.
Files
Steps to reproduce
Read README.md and the frozen diagnostic protocol first. Inspect the retained raw arrays and replicates.csv/summary.json for all 12 conditions. Python 3.12.14 and NumPy 2.3.5 were used in the recorded run. Original code, dependencies, input/output hashes, seed and scientific conditions are preserved unchanged. The original scripts contain absolute workspace paths and intentionally refuse to overwrite outputs. An independent rerun requires an explicit path mapping and a new output location; preserve original source and records, and report any relocated execution separately. Do not delete retained outputs to rerun. The bounded cache audit independently recomputed 48 predetermined matching MGFs and checked all 768 direct-metric/summary records; it did not independently recompute the other 720 MGFs. No independent-machine full rerun is claimed.
Institutions
- Guangdong University of FinanceGuangdong, Guangzhou