REPRODUCIBILITY.md

Reproduction

The source, settings, raw observations, ground-truth models, derived results, and figure inputs are all included. No private files or external data are required.

Recorded environment

Python 3.12.14, NumPy 2.3.5, SciPy 1.17.0, Matplotlib 3.10.8. Full environment details appear in results/environment.json. PDF authoring used pdfTeX 1.40.25; Pandoc generates the Markdown reading copy. Standard TeX packages are listed in manuscript/main.tex.

Install Python dependencies in an isolated environment using requirements.txt. Exact version pins record what ran; platform wheels and BLAS implementations may affect low-order floating-point bits.

Verify the downloaded release before changing files

python verification/check_manifest.py

This checks each included file against SHA256SUMS.txt. It does not establish scientific correctness or independent authorship.

Execute the research workflow

From the extracted package root:

python -m unittest discover -s tests -v
python experiments/run_all.py
python experiments/run_nonlinear.py
python verification/verify_release.py
python verification/verify_nonlinear.py
python experiments/make_figures.py
python experiments/make_nonlinear_figures.py

The run_all script regenerates observations and results in place using fixed seeds. The verifier deliberately does not import ot.core. The figure script reads saved CSVs rather than hidden simulation parameters.

No GPU or network access is needed after installing dependencies. Runtime depends on the host; this release does not claim a cross-machine speed benchmark. The nonlinear verifier imports neither estimator module.

Expected numerical checks

• All 30 exact reconstructed ranks match their generating dimension.
• Maximum relative generator error is below 1e-9 on the released exact cases (observed 1.03e-13).
• All 120 local bounds hold, allowing the verifier's 1e-11 floating-point slack.
• All 199 support cases and 25,036 support values pass the separate verifier.
• The bounded two-clock branch example has relative error below 1e-9.
• Scalar-chain small-time fitted exponents are within 0.03 of 0,2,4,6 over the declared finite interval.
• Gaussian estimates vary with random draws; exact seeds reproduce them subject to library versions. The analytic error is the reference, not a guarantee that every Monte Carlo cell lies within a chosen confidence band.

The tolerances are implementation checks, not theorem assumptions for arbitrary physical devices.

Build the manuscript and reading copy

python scripts/build_manuscript.py

This runs pdflatex three times to stabilize the contents and references, writes EVE_Obstruction_Tomography_II.pdf at the package root, and builds manuscript/manuscript.md with Pandoc. When a TeX source is edited, review the generated pages again. The authoritative equations and theorem cross-references are in the PDF and TeX sources; the Markdown copy is provided for search and convenience.

Regenerate the release inventory

After intentionally regenerating or editing files:

python verification/check_manifest.py --write

This creates a new local checksum inventory. Do not describe regenerated hashes as the original release's hashes. The manifest excludes itself, its inventory JSON, Python bytecode, and temporary TeX build outputs. It includes the final PDF, all source code, raw data, results, figures, and documentation.

What is not reproduced

The package does not reproduce a real laboratory experiment, a historical novelty search over all literature, external peer review, proof-assistant verification of the general theorems, or total physical oracle acquisition costs. The unrestricted exact-data generator construction is a mathematical limit; the numerical code implements explicitly narrower logarithm branches.

Additional v2 checks and boundaries

All 19 tests pass. The new verifier checks 76,800 anchored loss observations, 99,846 grid observations, 1,280 support witnesses and 3,003 topology witness rows. All six grid certificates and 32 map/drift setting pairs hold. Noiseless map error is below 1e-12 (observed 3.51e-16 in the generating implementation; the separate formula can differ in the last bits).

The nonlinear code implements analytic shape-distance examples and exact anchored readout for a specified oscillator. It does not implement the general unconditioned nonlinear identification theorem. Shared timing error is modeled across each anchor's four probes; independently varying timing or uncertain anchors are outside this bound. The affine baseline is only a model-misspecification diagnostic, not a query-matched comparison.

Hugging Face distribution

This edition adds a dataset card, expert/agent guides, a lossless support-table concatenation and metadata. The scientific results and manuscript are unchanged. The original complete science ZIP under downloads/ is preserved byte-for-byte and has its own internal inventory. The root checksum inventory covers this expanded distribution, including that archive.