Reproduction and integrity checks
From the dataset root:
git lfs pull --include='matrices/**'
sha256sum -c manifests/SHA256SUMS
python -B scripts/verify_dataset.py .
The verifier uses only Python's standard library. It reads the materialized
compressed Matrix Market files, checks their content and file hashes, verifies
H_X H_Z^T = 0 over GF(2), recomputes both ranks and k, checks the strict
FOM arithmetic, and validates every included logical witness against the
published matrices. The Stage 3 typed-exact verifier additionally reconstructs
the twisted Hx/Hz, validates the complete Gx/Gz logical basis, and
replays the weight-16 upper witness under the stated pinned-solver boundary.
The exact public pipeline parameters and source bindings are recorded in config/run_config.public.json and manifests/selection.json. Full stochastic search replay requires the source revision and matching dependencies; this compact expert package intentionally omits model transcripts, internal prompts, process state and checkpoints.