Paper and code: links, not bundled copies
Abbas, A. et al., The power of quantum neural networks, Nature Computational Science (2021), DOI 10.1038/s43588-021-00084-1.
| Inspected item | Relevant location | SHA256 of inspected bytes |
|---|---|---|
| Author-hosted manuscript | PDF page 2: conditional Fisher/Hessian discussion; PDF page 7: discussion of trainability | d1c369f83c6dcda1ae2e9d296f66fb162616092135810e61cd12833f5313d9df |
| Publisher supplementary information, accessible from the paper page | PDF page 6 (printed page 5), §3.3: training initialization; PDF page 8 (printed page 7), §4: conditional connection | d2959b56e19a613e868e8d6891012160cdcc48f361ec22b6f1e3fbf92f0a127c |
| Zenodo archive v1.0.1 | amyami187-effective_dimension-5d9a9b6/Loss_plots/generate_data/classical_loss.py | Archive: 5ff68f6fb1316ff2308a498f5242e7bf02b3d17c07dfb70dc1644da03b2e7d20 |
| Pinned classical training file | Model architecture; uniform_(0,1) initialization; mean cross-entropy and ADAM | Member: e9dbd8f63801e392700283da31c9979c2f4ac84dd74e265c1d46874b19a1a642 |
The paper's code-availability reference points to v1.0.0. The inspected classical training member in that release has the same SHA256 as the v1.0.1 member above. This does not identify the historical execution or its random state.
The manuscript is an author-hosted version, not silently substituted for the publisher's main-text PDF. The supplement is publisher-hosted. Version distinctions remain part of the attribution.
Original paper PDFs, original figures, source screenshots, complete source archives and author result arrays are not included. A public download is not itself a blanket redistribution licence. The archived repository carries an Apache-2.0 root licence; that fact is not used here to assign the same licence to every research artifact. This companion uses a separately written scalar formulation of the model, with attribution, rather than copying the archived ANN class or its training program. We have not selected a new licence for LFR's own manuscript and scripts; this handoff does not grant third-party rights or imply author endorsement.
Exact input for our calculation
The input is scikit-learn 0.23.2 iris.csv, 2734 bytes:
f13ffa8fdd56fd8e6c8d16d4081a3fbd3114bcd0aae4256c43205169cd9d1449
The input file is not redistributed in the bundle. The optional fetch script requests that exact public URL and refuses a changed hash. Its use is explicit, not a hidden network step in the verifier.
The versioned scikit-learn description identifies its Iris source and distinguishes it from the then-described UCI copy. The scikit-learn 0.23.2 COPYING file provides the project's BSD terms. We do not relabel the exact input as a UCI download or attach a different dataset licence to it by name alone. This inventory records provenance and packaging choices, not a comprehensive legal opinion.
Our numerical evidence
The underlying investigations are identified by their immutable local result hashes, not claimed to be externally time-stamped or independently reviewed:
- New classical run
RESULT.json:8fd69bde547dc6528a0567c10f5eb4efad457a6022ebc41f35de11831ffa9564. - Saved-run decomposition
RESULT.json:5799e12cb4d7ffeddd926a94029a49cb864dfad52fb960b8ed71d7aad3e1c671.
checkpoint_0.json is an explicit projection of those results. It includes saved weights, the next update, matrices and per-row derived contributions, but not the raw flower measurements. PACKAGE_MANIFEST.json pins every public-bundle file other than itself. Hashes detect changes relative to that manifest; they do not establish scientific truth, author identity, independent custody, or publication permission.
If a source version, input hash, model definition, coordinate or interpretation changes, preserve this package and issue a named successor with the reason for correction.