G Metrology Reproducibility Package

This repository contains the data, software, analyses, tests, and manuscript-output generators required to reproduce the computational results of the accompanying study "Covariance scaling and a dual-mode torque diagnostic in measurements of the gravitational constant".
arXiv2610.04707

Scientific scope

The study evaluates the adequacy of the quoted covariance model for a historical archive of G measurements, tests its predictive implications for a later lineage-related NIST configuration vector, using a BIPM-14-excluded historical training set as the primary prospective configuration, and analyzes free-versus-servo differences within the NIST apparatus. The full historical archive is retained as an explicit prospective robustness configuration.

The study does not claim a new value of G or identify a universal physical mechanism for the observed discrepancies. Statistical parameters such as lambda, tau, and nu are model parameters and are not interpreted as uniquely identified physical causes.

Install

Create a clean python environment and choose between the two installation methods.

Option A — Standard installation

Install G-metrology and its test dependencies. Pip selects compatible dependency versions.

python -m pip install ".[test]"

Option B — Pinned core/test environment

Use the dependency versions recorded in requirements-core.lock, then install without changing those dependencies.

python -m pip install -r requirements-core.lock
python -m pip install . --no-deps

A pinned environment reduces variation caused by dependency upgrades. The lock file pins the packages it lists; it does not by itself guarantee identical results across every operating system, Python version, or hardware configuration.

Choose one method, not both. Option A is simpler for ordinary installation. Option B is intended for reproducing the documented core/test environment with specified dependency versions.

Additional dependencies for NUTS and SBC

For NUTS and full SBC stages:

python -m pip install -r requirements-bayesian.txt

Reproduce the results

The reproducibility workflow has three cumulative execution modes. The default is the lightweight --core mode.

• --core runs the primary analyses, integrity checks, and regression tests needed for the core reproducible results.
• --full extends --core with the full simulation-based-calibration (SBC) stage.
• --archive extends --full with fresh NUTS sampling and archived-reference generation.

python reproduce.py --core
python reproduce.py --full
python reproduce.py --archive

--skip-nuts skips the fresh NUTS sampling stage in --archive. --skip-sbc skips the SBC stage in --full or --archive. Both flags affect execution only and do not delete existing saved results. On legacy hardware, combine them with --archive to shorten runtime.

Targeted commands are available for generating the paper figures/tables and a supplementary predictive band, respectively:

python reproduce.py --paper
python reproduce.py --gaussian-predictive-band

Predictive artifacts

The primary predictive band is the Gaussian proportional common-scale model. The historical Student-t predictive band is preserved only as an archived artifact. Neither is a numbered manuscript table; see CLAIMS.md and docs/paper_reproduction.md for their status and provenance.

Repository layout

• data/ contains the authoritative scientific input datasets and their integrity manifest.
• src/g_metrology/ contains the shared scientific implementation.
• analyses/ contains executable analyses organized by scientific purpose.
• results/machine_readable/ contains generated numerical results.
• results/nuts_reference/ contains completed NUTS computational reference artifacts.
• tests/ contains the standard pytest regression and integrity tests.
• paper/ contains generators and outputs for manuscript figures and tables.
• docs/ contains data provenance, model specification, reproducibility, limitations, and publication mapping.

Traceability

CLAIMS.md links each manuscript result to its input data, analysis script, numerical result, and manuscript output. docs/paper_reproduction.md describes the paper-reproduction workflow, execution modes, and generated outputs.

Data

Each active input dataset occurs exactly once under data/, with SHA-256 values recorded in data/MANIFEST.json. Generated results and manuscript artifacts are downstream products, not alternative input sources. The software is released under the MIT License; the curated scientific input datasets originate from external published sources and retain their own licensing, attribution, and reuse terms. The MIT license does not relicense those external datasets.

Terminal and progress display

Successful stages report only completion status and elapsed time; detailed output is shown only on failure. SBC is the only stage with a live progress display. Fresh NUTS runs have the native sampler progress bar disabled.

Citation

The software citation metadata are provided in CITATION.cff.