Reference-prevalence SoftMCC profiles: public replication repository

Public replication material for the manuscript Reference-prevalence SoftMCC profiles: Exact characterization and threshold-risk non-identification (Canay).

This repository was replaced on 2026-09-20 with the artifacts of that study. The artifacts of the earlier cost-analysis study remain reachable in the git history at commit 99827b5.

Contents

• code/scientific_redesign.py: theory checks, registered units, controls, aggregation, and resource capture.
• code/posthoc_sensitivity_20260826.py: common-model, exact-manifold, residual, near-match, and monotone-placebo checks.
• code/generate_redesign_figures.py: manuscript figure generation.
• config/scientific_redesign_20260826_v2.json: locked design and dataset manifest.
• results/canonical_results_20260826.tar.gz: canonical main and control results, run manifests, the locked configuration as executed, checksums, theory verification, and figures.
• results/posthoc_sensitivity_20260826_v2/: post-hoc audit outputs with their checksums.
• information_resolution/: the separately frozen resampling extension. Its design configuration, run manifest, admission freeze, descriptive outputs, verification records and audit scripts are published here; the raw run directories are about 920 MB and are not.
• requirements.txt: package versions recorded on the run host, plus the local plotting dependency.
• checksums.sha256: SHA-256 of every tracked file in this repository.

What the runs produced

The canonical main run completed 720 of 720 registered model-fit units and the negative-control run completed 84 of 84. The information-resolution extension completed 120 real panels and one synthetic panel. The post-hoc common-model, prevalence-shrinkage and monotone-placebo analyses are safeguards, not superiority experiments: the mathematical results rest on the exact rational construction and the sharp moment bounds, not on empirical pair counts.

Verification without downloading data

From the repository root, the theory layer verifies on its own:

python code/scientific_redesign.py verify-theory --output verification/theory --project-root .

Data boundary

The locked configuration names OpenML dataset identifiers and does not redistribute raw third-party rows. A full rerun acquires the named public datasets through fetch_openml, so network availability and the upstream dataset versions remain external dependencies. See DATASETS.md.

License

MIT, see LICENSE.