Reproduction package — Finite-noise route polarization: coarse-graining bounds and compatibility gaps in nonequilibrium branch selection
Description
This dataset provides the complete computational reproduction package for the manuscript *Finite-noise route polarization: coarse-graining bounds and compatibility gaps in nonequilibrium branch selection*. The manuscript studies information loss and polarization under finite route coarse-graining of matched forward–reciprocal path experiments. The package reproduces the numerical results supporting the finite-noise KL decompositions and bounds, Fisher-information decompositions, high-resolution and high-noise asymptotics, weak-noise compatibility results, and the thermodynamic-identification examples discussed in the manuscript and Supplementary Material. The archive contains: * Python source code for all numerical calculations and independent consistency checks; * source-hashed CSV and JSON data used in the tables and figures; * scripts for regenerating the manuscript figures; * independent numerical audits of the KL chain rule, sharp likelihood-ratio bounds, Fisher decomposition, high-resolution coefficients, weak-noise finite-family limits, and model identities; * three executable nonequilibrium models illustrating the physical-defect, same-endpoint additive-state, and topological-holonomy constructions; * exact dependency and build-environment information; * SHA-256 manifests for source and retained-input verification; * LaTeX sources and build scripts for reproducing the manuscript and Supplementary Material. The numerical outputs distributed in this dataset are generated deterministically by the supplied source code, subject to standard floating-point tolerances. Independent checks reproduce the principal algebraic identities and numerical results without importing the corresponding production routines. The associated preprint is archived on Zenodo: * Version 1 DOI: 10.5281/zenodo.21481831 * All-versions DOI: 10.5281/zenodo.21481830 Code is released under the MIT License. Manuscript text, figures, and data are released under CC BY 4.0 unless otherwise stated.
Files
Steps to reproduce
The package requires Python 3.10 or later with NumPy, pandas, Matplotlib, and SciPy. Reproduction of the PDFs additionally requires a LaTeX installation with latexmk and the packages listed in the README. 1. Extract the archive Extract the complete reproduction package and open a terminal in its top-level directory. 2. Install the Python dependencies python -m pip install -r requirements_finite_noise_route_polarization_submission_v34.txt The exact environment used for the audited release is recorded in BUILD_ENVIRONMENT.txt. 3. Run the complete reproduction workflow On a Unix-like system with Bash: bash build_all.sh The build verifies the retained source hashes, executes the main numerical calculations and independent checks, regenerates the CSV/JSON outputs and figures, compiles the manuscript and Supplementary Material, and produces the merged PDF. For concise output: QUIET=1 bash build_all.sh 4. Alternative two-stage execution If the outer shell delays returning after the numerical subprocesses have completed, run the numerical and LaTeX stages separately: BUILD_STAGE=numerics bash build_all.sh BUILD_STAGE=latex bash build_all.sh Both commands should return with exit status zero. 5. Verify the results The build verifies the files listed in SOURCE_SHA256SUMS.txt and then executes the production calculations together with independent and import-independent audit routines. Generated numerical outputs use the naming convention: submission_v34_*.csv Independent validation outputs use: independent_v34_*.csv FIGURE_SCRIPT_MAP.md identifies the source script and numerical data associated with each generated figure. The principal reproducibility checks cover: - directional KL and Jeffreys chain-rule decompositions; - route-refinement information loss; - sharp bounded-likelihood certificates; - Fisher-information chain rules and integral identities; - high-resolution coarse-graining coefficients; - finite-noise and high-noise polarization behavior; - weak-noise finite-family compatibility limits; - detector mismatch and sufficient-statistic checks; - physical-defect and connector-covariance identities; - same-endpoint additive-state lifting; - circle-holonomy and winding-sector calculations. Floating-point comparisons are tolerance based. Last-bit numerical differences may occur across BLAS implementations or compatible library versions, while PDF file hashes may differ because of build metadata. These differences do not affect the audited numerical identities or reported manuscript values.