HCMDP — Evaluation Code and Numerical Reproducibility Data

Abstract

This archive accompanies “HCMDP: A Multi-Criteria Evaluation Protocol for Selecting Open-Weight Large Language Models in Air-Gapped Institutions.” It contains recovered evaluation code, configuration records and derived numerical data for run 20260202_153016_b62488db: six models, four deployment scenarios and 1,202 common test questions (28,848 predictions).Version 1.1.0 reproduces the original dimension values and 72 composite scores across the HCMDP v2.0, v2.1 and v3.0 scoring profiles. It retains the original components, weights and saved measurements. A separate command exports the original stored statistical results without recomputing inferential tests. Split/source identifiers, numerical auxiliary-test records, provenance, declared transformations and file checksums are included.The archive distinguishes numerical reproduction from measurement validity. The constant-confidence calibration fallback, study-specific HHEM input and score interpretation, S4 retrieval metric assembly and statistical dependence limits are documented in MEASUREMENT_SCOPE.md and SOURCE_SCOPE.md. Composite scores are interpreted within a specified scenario and implemented profile. This release contains no replacement composite analysis.The package does not regenerate model answers or guarantee a complete historical training/inference rerun. Full question, source, reference and generated-answer text, trained adapters, a standalone retrieval cache and some exact historical assets are not included. The split shares source content with training, and the evidence comes from one production run. Automated support measurements are not human-validated factuality estimates.Code is licensed under MIT; derived numerical data and documentation under CC BY 4.0. Metadata are public. Files are restricted during peer review, with confidential access for the editor and reviewers, and will be made public when the article is published.