Datasets:
Reproducing ORYNTHRA-H6
This guide applies to mathematical core v1.0.0 and Hugging Face release edition v1.0.2. The manuscript and core algorithms are unchanged.
Environment
Use Python 3.10 or newer. The original and fresh verification environment uses Python 3.12. Set up a virtual environment and install the supplied dependencies:
python -m venv .venv
Windows:
.venv\Scripts\activate
python -m pip install -r requirements-tested.txt
Linux/macOS:
source .venv/bin/activate
python -m pip install -r requirements-tested.txt
requirements.txt gives minimum requirements; requirements-tested.txt records the exact versions used in verification. Other Python/dependency combinations are not certified by these checks.
Integrity before regeneration
python tools/hf_upload.py
Without --publish, this performs a local release-file hash check and prints a plan. It needs no Hugging Face package, token, account, or network connection. The manifest covers the distributed contents, not later generated output. Run reproduction in a separate extracted copy: it rewrites data, figures, environment records, and logs.
Computation and tests
From the repository root:
python code/demo.py
python code/reproduce.py
python -m pytest -q tests
python code/make_figures.py
Or use run_reproduce.bat on Windows and bash run_reproduce.sh on Linux/macOS after installing dependencies. The Windows launcher retains its window on completion or error; Windows execution is not claimed as tested in this Linux session.
Expected invariants:
| Check | Expected |
|---|---|
| Regression tests | 44 passing |
| Six-wing proper-subset maximum rank | 6 |
| Six-wing full-set rank | 36 |
| Generator subsets checked | 64 |
| H6 permutations | 720 |
| Symbolic full-Gram identity | True |
| Symbolic simplex-Gram identity | True |
| H6 coherent photon cutoff | 4 |
| H6 storage photon cutoff | 2 |
| H6 uniform mean photons | 10/3 |
| H6 simplex angle | Approximately 101.536959 degrees |
| S5 multiplicity permutations | 120 |
| Multiplicity matched dimension | 6 of 10 |
| Multiplicity normalized defect | 0.8, up to floating-point rounding |
| 64-label two-mode cutoffs | Storage 10; full permutations 62 |
| 4,096-label two-mode cutoffs | Storage 90; full permutations 4,094 |
Machine-readable results appear in verification/results.json. CSVs appear in data/. Floating-point residuals and eigensolver outputs should be compared with appropriate tolerances; wall time, platform identifiers, and PDF metadata need not match byte for byte. Numerical assertions establish particular examples, not general theorem proofs.
Fresh release records
The packaging session's checks are recorded separately in verification/hf_release/, preserving the original verification files. The report distinguishes original-source hash checks, test results, regenerated-result comparison, and uploader offline checks. It also records checks not performed.
Optional manuscript rebuild
The prebuilt PDF needs no LaTeX. To rebuild it, install LaTeX with latexmk and the packages named in the .tex preamble:
cd manuscript
latexmk -pdf -interaction=nonstopmode -halt-on-error ORYNTHRA_H6.tex
The figures are already supplied in PDF format. To regenerate the searchable Markdown, install Pandoc, return to the root, and run:
python code/make_markdown.py
This edition does not claim a fresh LaTeX rebuild or a new visual review of all 39 pages. The original package includes its document checks and render records.
Algorithmic boundaries
- The manuscript's field theory is more general than the implementation: code uses prime fields, not arbitrary extension fields.
- Exact packing is a reference solver for small examples and may take exponential time.
- Logical irrep dimensions/decompositions must be supplied.
- Optical matrices represent occupied sectors; they are not fabricated gate sequences or a complete foundry design.
- The H6 code intentionally fails arbitrary-state one-photon-loss correction; the limitation tests are part of the result.