Research software, archived results, and a mathematical audit prepared for a Computers & Operations Research manuscript. Author: Michael Doane.
The paper develops a budgeted randomized selection formulation, its validity results, and a robustness–nominal-value frontier. The completed core v5 study passed its locked technical gate (GREEN).
An exploratory CTOD application is preserved in the archive. Its benchmark suitability for the intended decision setting remains unresolved, and its recorded YELLOW gate does not establish a general limitation of B-MERIT. See the positioning decision.
- Audit report and corrected interpretation
- Audited proofs
- Prior-art and novelty audit
- C&OR manuscript preparation plan
- Detailed paper outline
- Claim-to-evidence map
- GitHub and Zenodo instructions — graphical interfaces only
- Reproduction instructions for researchers
| Study | Included evidence | Scientific interpretation |
|---|---|---|
| Core v5 | 480 synthetic, 20 semi-synthetic, 12 exact-scaling and 18 scenario-scaling tasks | All locked technical gates passed |
| CTOD exploratory application | 69 retrospective portfolios, 10 methods, 27 scenario/budget cells | Recorded YELLOW gate for a different policy variant; benchmark suitability remains unresolved |
| Audit sensitivity | Recomputed gates and unrestricted hindsight normalization | Core top-set gains unchanged; pooled hidden-value gain remains positive |
No full PRISM or CT Open results are included. No prospective clinical, pharmaceutical investment, or causal performance claim is made.
experiments/core_v5/: the frozen 0.5.0 core source, configuration, original documentation, tests, and attributed public grant fixtures.experiments/ctod_v060/: the frozen 0.6.0 CTOD extension and its protocol.results/original/: complete unmodified core and CTOD result archives.results/audit/: newly computed audit outputs, explicitly separate from the original gates.audit/: archive verification and reproducible evidence-audit scripts.tests/: small independent checks of corrected mathematical statements.
The top-level audited documentation supersedes conflicting proof wording in the historical experiment documentation. The old files are retained to preserve the exact source identity of the completed experiments. In particular, the older FEV spending statement is false; see the counterexample in the audited proofs.
In the three primary synthetic regimes, canonical ON-B-MERIT improves hidden attainable top-set recovery over midpoint selection by 11.30–17.81 percentage points. Allowing a 0.02 loss of normalized robustness produces a pooled 4.43-point gain in normalized nominal value against the canonical endpoint. The audit's unrestricted-hindsight hidden-value gain is 3.37 points (95% interval 3.00–3.73).
CTOD's value-retaining mixture retains 98.879% of realized utility against the expected-value baseline, but achieves only 3.050% relative reduction in the across-portfolio q90 expected-regret statistic. Its modeled recovery is lower than deterministic maximin. Raw B-MERIT still matches or exceeds deterministic maximin recovery in every primary CTOD comparison; the mixture trades away that protection. These are observations within the constructed CTOD task. They do not identify whether the outcome arises from the dataset, task construction, policy variant, or their interaction. The manuscript assigns this application a supplementary role. External applicability remains an open validation question. See the audit for the complete gate and interpretation.
Original software and documentation are MIT licensed. Third-party data retain their own terms, including CC BY 4.0 for the Swiss grant-review fixture. See NOTICE.md. The repository includes research code and numerical artifacts, not a clinical or investment decision product.
MERIT was introduced by Alexander Goldberg, Giulia Fanti and Nihar B. Shah (NeurIPS 2025). This is an independent extension and does not imply their endorsement. See the novelty audit for attribution.
The author has confirmed the GitHub repository
and the v1.0.0 Zenodo record.
That release packages frozen algorithm versions 0.5.0 and 0.6.0 and their results.
Use CITATION.cff for its software citation.
This working-tree documentation revision follows v1.0.0 and corrects the CTOD interpretation. The existing DOI identifies the original archive; it does not identify these subsequent documentation edits. Archive the final manuscript support package as a new release when its contents are settled.