This bundle is a manuscript and a one-point numerical companion. It does not reproduce the original paper's historical benchmark. No quantum simulation or training is run.
1. Check the saved arithmetic, offline
In a terminal opened in the extracted bundle directory:
python -B verify.pyOnly the Python standard library is required. The command checks bundle hashes, matrix and row identities, sums, direction normalization, and the numbers displayed in the case. It explicitly reports that input-based reconstruction has not run.
2. Recompute the model at the saved point
Provide the exact input described in SOURCES.md. You may download it yourself, or explicitly run:
python -B fetch_input.py iris_input.csv
python -B verify.py --input iris_input.csvThe second command requires NumPy and PyTorch. It was tested with Python 3.13.5, NumPy 2.2.6 and PyTorch 2.7.1+cpu. No dependency installation happens automatically. The exact CSV hash is checked before calculation; a different Iris copy is refused.
The command recomputes F, the fixed-label Hessian, the residual decomposition and the first-update direction. It checks a model-probability-target control and a branch-preserving finite difference. It prints actual runtime versions and tolerances are documented in the methods. No saved result is overwritten.
Reading failures
A hash mismatch means the requested input or package bytes are not the pinned ones. A numerical mismatch means the stated check did not pass in the current execution. Neither justifies silently relaxing a tolerance, substituting another dataset, or rewriting this package.
Preserve the output, file hashes and runtime version; investigate the failing comparison. Any corrected package should be a named successor, not a replacement that hides the failure.