Reproduce and audit the collection
1. Obtain a fixed edition
Use the v1.8-open.1 release assets. ORSH_AIU_v1_8_Finalized_Complete.zip
is the unchanged original archive; ORSH_AIU_v1_8_Open_Publication_1.zip
is the separately licensed public repository bundle. Reproduction_v1_8.zip
contains the curated source/data/code and these instructions. Compare downloads
with the release's SHA256SUMS.txt (PowerShell: Get-FileHash -Algorithm SHA256 FILE;
Linux: sha256sum FILE). Extract into a new directory.
The normal checkout intentionally omits bulky historical bundles and review
correspondence. provenance/source-inventory.json lists every retained and
archive-only file. Original archive checks must run on the complete archive,
not the curated publication/ subset.
2. Install an isolated numerical environment
Use Python 3.13 (the original snapshot used 3.13.5). A local publication check also uses Python 3.12; exact local versions are recorded in the new receipt.
python -m venv .venv
# Linux/macOS
. .venv/bin/activate
# Windows PowerShell, instead:
# .\.venv\Scripts\Activate.ps1
python -m pip install -r requirements-numerics.txt
Set OPENBLAS_NUM_THREADS=1, OMP_NUM_THREADS=1 and PYTHONUTF8=1 before
running. On PowerShell use $env:OPENBLAS_NUM_THREADS='1' and likewise for
the other two names. Run all commands below from the repository/bundle root.
3. Fast, scoped checks
python scripts/verify_publication.py
python scripts/check_numerics.py
The numerical check recomputes the factorial analysis from retained arrays and compares all output fields; it also runs selected four- and six-qubit propagation checks using exponential action and density matrices. It writes fresh outputs to a temporary directory, preserving the published research files. The numerical check is not a replay of every original trajectory or external replication.
4. Full pilot and reviewer replay
Choose output directories that do not already contain results.
python publication/code/run_injection_pilot.py --config publication/pilot/config.json --out fresh-pilot
python publication/code/replay_reviewer.py --root publication --out fresh-reviewer --task all
The full reviewer replay can take many minutes or substantially longer on slower
machines. Inspect fresh-reviewer/comparisons.json: every match must be true.
Its comparison uses rtol=1e-8, atol=1e-10; a completed process alone is not
proof of agreement. The manually dispatched Full research replay workflow
performs these runs and asserts agreement. It is separate from pull-request CI.
5. Complete original package and document checks
Extract the original archive to a separate directory. Install Pandoc 3.1.11.1
and the dependencies in requirements-site.txt (including lxml and
python-docx). With that directory as ROOT:
python ROOT/code/verify_release.py --root ROOT --report original-structure.json
The original verify_archive.py uses platform-native path separators; its
unlisted-file comparison is Linux-oriented. The maintained verifier's
--archive PATH.zip option performs a portable complete ZIP/hash audit:
python scripts/verify_publication.py --archive ORSH_AIU_v1_8_Finalized_Complete.zip
PDFs are preserved from the source package's Pandoc + XeLaTeX build, not
the older LibreOffice process mentioned in publication/code/README.md.
Full PDF rebuilding requires TeX Live and the fonts recorded in
publication/verification/environment.json; byte-identical rebuilt PDFs are
not promised. DOCX files contain native Word math. Rebuild only in a working copy.
6. Build the accessible reading site
This section requires the full Git checkout or open-publication bundle; the smaller reproduction bundle omits the website tooling.
Install Pandoc 3.1.11.1, Python dependencies and Node 22.12 or later:
python -m pip install -r requirements-site.txt
npm ci
python scripts/build_site.py
python scripts/check_site.py
npx playwright install chromium
npm test
python -m http.server 4173 --directory _site
The site uses locally bundled KaTeX, local full-text search, no analytics and no
account requirement. Generated _site/ is deployed by GitHub Pages Actions.
The full replay workflow retains its fresh outputs as a downloadable Actions artifact for 14 days. These run artifacts are temporary; the versioned research release assets are the permanent publication copies.
View automated verification runs · Open-publication check receipt