arXiv:2604.15931 Code and Data
Description
This dataset contains the code, numerical inputs, completed outputs, and validation records supporting the manuscript “Robustness of Starobinsky inflation in a minimal two-field scalar-tensor completion” (arXiv:2604.15931). The calculation studies a two-field scalar-tensor extension of Starobinsky inflation across 246 trajectories in the sampled parameter domain. The seven-stage analysis includes reduced slow-roll backgrounds, stability tests with the complete constraint-reduced quadratic operator, scalar and tensor spectra, Bunch-Davies convergence, entropy-transfer diagnostics, and the numerical data behind the manuscript figures and tables. Two independent calculations test the derivative-coupling reduction used for the production scalar spectra. The direct calculation evolves the complete scalar perturbation operator at beta = -1, 0, +1 on every stored background and evaluates Delta_beta = P_R^complete(beta)/P_R^complete(0) - 1, where P_R is the dimensionless curvature-perturbation power spectrum. It includes a global scan, 501-mode refinements, solver and matrix-grid convergence tests, vacuum-surface tests, and a separate beta = 0 complete-versus-reduced regression. The coefficient calculation bounds the beta-dependent kinetic, gradient, physical mixing, and mass matrices in a common beta-zero kinetic-normalised basis for beta = -1, -0.1, 0, 0.1, +1. Both tests use fixed backgrounds generated from the reduced slow-roll system. They test the perturbative reduction and do not represent a derivative-coupled homogeneous-background evolution. The adaptive SciPy/DOP853 CPU calculation supplies the reference result and numerical-error estimate. A separate double-precision JAX/Diffrax calculation on an NVIDIA A30 supplies dense 501-mode coverage. The GPU result is supplementary because a deliberately stringent CPU-GPU parity condition was not satisfied; its full discrepancy is included in the conservative error budget. The reference scan gives max |Delta_beta| = 1.98407957e-12 with numerical allowance 1.75004455e-12. Including the dense-GPU result and full CPU-GPU discrepancy gives the conservative bound 3.89488834e-07, below the 1e-3 accuracy relevant to the manuscript’s conclusion. The maximum common-basis differences are ||Delta K-hat||_2 = 2.81645764e-08, ||Delta D-hat||_2 = 2.80264147e-08, ||Delta Omega_phys||_2/H = 1.76090623e-10, and ||Delta M_phys||_2/H^2 = 1.58572738e-11. The reproduction archive contains the manuscript outputs, shared inputs, validation results, figures, tables, machine-readable summaries, provenance, and SHA-256 manifests. The source archive contains the code, configurations, tests, and documentation. The workflow supports Linux, Python 3.11 or newer, 1-8 CPU workers, resumable calculations, and optional CUDA 13 acceleration.
Files
Steps to reproduce
Use the reproduction archive to verify the published results. Commands below assume Bash on Linux and Python 3.11 or newer. Do not overwrite the deposited result directories. 1. Verify and extract the archive: sha256sum -c starobinsky_modification_numerics-0.6.1-jcap3-reproduction.tar.gz.sha256 tar -xzf starobinsky_modification_numerics-0.6.1-jcap3-reproduction.tar.gz cd starobinsky-numerics-portable/project python3 scripts/build_portable_bundle.py --verify /absolute/path/to/starobinsky_modification_numerics-0.6.1-jcap3-reproduction.tar.gz 2. Create the CPU environment and run the tests: python3 -m venv .venv ./.venv/bin/python3 -m ensurepip --upgrade ./.venv/bin/python3 -m pip install --upgrade pip setuptools wheel ./.venv/bin/python3 -m pip install --no-cache-dir -r requirements.txt ./.venv/bin/python3 -m pip check PYTHONPATH=src ./.venv/bin/python3 -m pytest 3. Inspect the deposited validation without recomputation: PYTHONPATH=src ./.venv/bin/python3 -m starobinsky_modification_numerics.cli status --analysis-dir ../results/cpu-validation cat ../results/cpu-validation/validation_summary.md cat ../results/coefficient-bound/coefficient_bounds.md cat ../results/gpu-benchmark-501/gpu_benchmark.md cat ../results/dense-gpu/dense_validation_summary.md cat ../results/referee-summary/referee_validation_summary.md Each archive has a complete MANIFEST.json; the result directories also contain independent SHA-256 manifests. 4. To repeat the reference CPU validation into a new directory: mkdir -p ../recomputed/cpu-validation/logs nohup env PYTHONPATH=src OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 ./.venv/bin/python3 -m starobinsky_modification_numerics.cli complete-validation --source-analysis ../source-analysis --analysis-dir ../recomputed/cpu-validation --workers 8 > ../recomputed/cpu-validation/logs/nohup.out 2>&1 < /dev/null & Monitor it with: PYTHONPATH=src ./.venv/bin/python3 -m starobinsky_modification_numerics.cli status --analysis-dir ../recomputed/cpu-validation The recorded eight-worker run took about 28 h 10 min on THEOR4. Recompute the coefficient bounds with: PYTHONPATH=src ./.venv/bin/python3 -m starobinsky_modification_numerics.cli coefficient-bound --matrix-run ../source-matrix-scan/results.h5 --source-analysis ../source-analysis --output-dir ../recomputed/coefficient-bound The complete seven-stage manuscript rerun, GPU installation, 501-mode benchmark, dense GPU scan, interruption recovery, and final consolidated-summary commands are documented in docs/reproduce_manuscript.md, docs/server_portability.md, and docs/publication_archive.md. The recorded GPU run used Linux x86-64, Python 3.13, CUDA 13, JAX 0.11.1 with float64/complex128, and an NVIDIA A30.
Institutions
- Joint Institute for Nuclear ResearchMoscow Oblast, Dubna