v1.1.6 — curated final submission artifact

v1.1.6 — curated final submission artifact

Version 1.1.6 is the final curated artifact for “Sharp Target-Domain
Certificates for Quantum-Kernel Advantage under Distribution Shift”, prepared
for manual submission to npj Quantum Information.

Relative to immutable v1.1.5, this release changes no dataset, split, seed,
candidate pool, prediction, label, result, table, figure, theorem, method,
bibliographic claim, or scientific conclusion. It removes unreferenced internal
planning and obsolete venue-conversion material, makes the network-flow driver
portable, clarifies the public artifact structure, and updates release identity
and packaging metadata.

Identity

• Version: 1.1.6
• Git tag: v1.1.6
• Version DOI: https://doi.org/10.5281/zenodo.21776862
• Concept DOI: https://doi.org/10.5281/zenodo.19147649
• Primary entry point: python scripts/reproduce_v11.py --stage all

The Zenodo deposit contains the tagged source archive, final manuscript and
Supplementary PDFs, release notes, manuscript-source bundle, and cryptographic
checksums. The GitHub release contains the final PDFs, manuscript-source bundle,
release notes, and matching asset checksums.

v1.1.5 — Bibliographic and Submission Closure

v1.1.5 — bibliographic and submission closure

Version 1.1.5 is the closed submission artifact for “Sharp Target-Domain
Certificates for Quantum-Kernel Advantage under Distribution Shift.” It
contains the manuscript, Supplementary Information, sources, evidence chain,
tests, and integrity records prepared for manual submission to npj Quantum
Information.

Relative to immutable v1.1.4, this release adds no experiment, dataset,
analysis, theorem, or changed result. The final bibliographic pass positions
the fixed-batch certificate within partial identification; distinguishes it
from minimax transductive aggregation; and separates it from active,
multi-model, and out-of-distribution risk evaluation. It also records the
finite-shot complexity context and recent favorable QML evidence without
changing the study's claims. Bibliographic metadata was checked against
publisher, proceedings, and preprint records.

The central claim remains the sharp and minimal interval for the finite-batch
functional comparing a fixed candidate with the best member of a prespecified
fixed family under unrestricted completions of unaudited labels. The update is
pathwise exact after any realized audit subset, including adaptively selected
subsets.

Identity

• Version: 1.1.5
• Git tag: v1.1.5
• Version DOI: https://doi.org/10.5281/zenodo.21764577
• Concept DOI: https://doi.org/10.5281/zenodo.19147649
• Full test suite: 120 tests on Python 3.11 and 3.12
• Primary entry point: python scripts/reproduce_v11.py --stage all

The Zenodo deposit contains the tagged source archive, final manuscript and
Supplementary PDFs, release notes, manuscript-source bundle, and cryptographic
checksums. The GitHub release contains the final PDFs, manuscript-source
bundle, and matching asset checksums.

v1.1.4 — Final Submission Artifact

v1.1.4 — final submission artifact

Version 1.1.4 is the final submission artifact for “Sharp Target-Domain
Certificates for Quantum-Kernel Advantage under Distribution Shift.” It
contains the complete theory, evidence chain, manuscript, Supplementary
Information, tests, and integrity records prepared for submission to npj
Quantum Information.

Relative to immutable v1.1.3, this release adds no experiment, dataset,
baseline, theory, or changed result. It applies the closed final editorial
pass: precise identified-interval terminology; positioning relative to
Slattery et al., Phys. Rev. A 107, 062417 (2023); consistent Gate-2 scope
against the prespecified classical-kernel reference family; factual reporting
of prospective corroboration; neutral figure and public-language labels; a
conservative main-text prune; corrected bibliographic metadata; and the
documented ethics determination.

The study falls outside the scope of institutional ethics review because it
consists exclusively of secondary analyses of publicly available,
deidentified datasets, without recruitment, interaction, intervention, access
to direct identifiers, or any attempt at reidentification. Ethics approval and
additional informed consent were therefore not required. No ethics approval
or exemption is claimed.

Identity

• Version: 1.1.4
• Git tag: v1.1.4
• Version DOI: https://doi.org/10.5281/zenodo.21759058
• Concept DOI: https://doi.org/10.5281/zenodo.19147649
• Full test suite: 120 tests on Python 3.11 and 3.12
• Primary entry point: python scripts/reproduce_v11.py --stage all

The Zenodo deposit contains the tagged source archive, final manuscript and
Supplementary PDFs, release notes, manuscript-source bundle, and cryptographic
checksums. The GitHub release contains the final PDFs, manuscript-source
bundle, and matching asset checksums.

v1.1.3 — Sharp Target-Domain Certificates

v1.1.3 — sharp target-domain certificates

Version 1.1.3 is the submission release for “Sharp Target-Domain
Certificates for Quantum-Kernel Advantage under Distribution Shift.” It
contains the complete theory, evidence chain, manuscript, Supplementary
Information, tests, and integrity records prepared for submission to npj
Quantum Information.

The central contribution is an exact, attainable target-batch identified set
for the advantage of a fixed quantum classifier over the best member of a
prespecified classical reference family. The framework includes exact partial
label updates, an information-optimality result, a reference-breadth–target-
supervision evidence frontier, balanced-accuracy sensitivity, and both
retrospective and internally locked prospective audits. Earlier benchmark
contributions remain preserved: protocol reversal, equal search budgets,
factorial and shortcut analyses, external TableShift shifts, product and ZZ
maps, circuit resources, geometry diagnostics, and finite-shot experiments.

Relative to the immutable v1.1.0 predecessor, v1.1.3 changes no scientific
result, prediction, label, analysis, theorem, figure, or reported number. It
only makes the frozen-manifest check byte-portable, invokes the validator as a
module, and permits the public reproducibility suite to omit the private,
Git-ignored cover letter while still validating that letter in the local
submission workspace.

Identity

• Version: 1.1.3
• Git tag: v1.1.3
• Version DOI: https://doi.org/10.5281/zenodo.21751137
• Concept DOI: https://doi.org/10.5281/zenodo.19147649
• Full test suite: 120 tests on Python 3.11 and 3.12
• Primary entry point: python scripts/reproduce_v11.py --stage all

The Zenodo deposit includes the tagged source archive, final manuscript and
Supplementary PDFs, release notes, manuscript-source bundle, and cryptographic
checksums. The GitHub release provides the final PDFs, manuscript-source
bundle, and matching asset checksums.

v1.1.0 — Sharp Target-Domain Certificates

v1.1.0 — submission release

This release consolidates the complete reproducible artifact for “Sharp
Target-Domain Certificates for Quantum-Kernel Advantage under Distribution
Shift.” It is the intended submission snapshot for npj Quantum Information.

Central advance

• Generalizes the finite-target identified set from zero-one accuracy to any
  additive bounded loss on a finite label space.
• Gives the exact lower endpoint and an exact finite max-min/MILP upper
  endpoint after any partial target-label audit.
• Proves endpoint attainment, monotone contraction, and information
  optimality: no assumption-free interval computed from the same fixed loss
  tables and audited labels can be uniformly smaller.
• Retains the closed-form sharp accuracy certificates used by the SVC and
  Laplace-GPC experiments.

Empirical evidence retained and clarified

• Preserves the controlled protocol reversal, equal-budget comparisons,
  factorial and shortcut analyses, external TableShift transfer, quantum-map
  strata, circuit resources, geometry diagnostics, and finite-shot study.
• Reports the frozen nested 30/60/115-candidate reference-breadth frontier
  without introducing a post-outcome family ordering.
• Preserves the retrospective eight-shift audit and the internally locked
  prospective Gate-2 replication, including its two-task technical scope,
  product-map limitation, prediction locks, separated labels, and failure
  records.
• Positions disagreement acquisition, active evaluation, partial
  identification, QML surrogates, geometric screening, and concurrent work as
  prior art rather than claiming those ingredients individually.

Reproducibility and identity

• Version: 1.1.0
• Git tag: v1.1.0
• Version DOI: https://doi.org/10.5281/zenodo.21750725
• Concept DOI: https://doi.org/10.5281/zenodo.19147649
• Full test suite: 120 tests
• Consolidation contract:
  docs/V11_CONSOLIDATION_SPEC.md
• Primary entry point: python scripts/reproduce_v11.py --stage all

Raw benchmark source records are not redistributed and remain subject to the
providers' terms. Derived predictions and target labels remain physically
separated in the archived artifact.

v0.8.0 — circuit-aware npj Quantum Information submission artifact

Circuit-aware submission artifact for npj Quantum Information and the Collection “Quantum machine learning: understanding capabilities, limitations, and perspectives for quantum advantage.”

This release corrects the quantum-map characterization and adds matched entangling-ZZ versus separable-product strata, a complete within-v4 2×2×2×2 SVC evaluation-protocol factorial, a frozen port/protocol/service-field ablation, logical circuit and finite-shot resource accounting, and the complete 360 fixed-C plus 270 shortcut-ablation campaign outputs. The controlled fixed-case result is not produced by pooling product maps with entangling circuits: source-dataset-equal effects are −0.00928 for entangling ZZ and −0.00547 for product maps at matched 30-candidate budgets. Within v4, the fixed-C/same-test-oracle/customary/native corner is +0.01308 and the train-CV/ID-validation/extended/equal-count corner is −0.00467; these are descriptive finite-design endpoints, not causal or population effects.

The canonical reproduction and historical integrity gates pass, 79/79 tests pass, and GitHub CI passes on Python 3.11 and 3.12. The Nature-style main manuscript (44 pages) and Supplementary Information (10 pages) compile cleanly and were inspected page by page. No equivalence, hardware-advantage, computational-speedup, or population-wide ordering claim is made.

Immutable Zenodo DOI: 10.5281/zenodo.21717074
All-version concept DOI: 10.5281/zenodo.19147649
Tagged commit: 275add1

v0.7.0 — npj Quantum Information submission artifact

Corrective submission artifact addressing the second strict pre-submission review. This release unifies the confirmatory 60-versus-60 candidate-budget description, clarifies the associational scope of the specification curve, reports complete sampling-scheme and small-cluster sensitivities, and extends the repeated finite-shot analysis with both an explicitly heuristic train-Gram intervention and a coherent Nyström out-of-sample PSD extension. The full reproduction and artifact audit pass; 69/69 tests pass; the Nature-style main manuscript (43 pages) and Supplementary Information (6 pages) compile cleanly and were inspected page by page. Immutable Zenodo DOI: 10.5281/zenodo.21676563.

v0.6.0 — npj Quantum Information submission artifact

npj Quantum Information submission artifact

This release contains the fully revised submission package for “Evaluation Choices Determine Apparent Quantum Kernel Robustness under Distribution Shift”.

Scientific revision

• Unifies the confirmatory endpoint and specification curve on three source datasets: EMBER, UNSW-NB15, and ToN-IoT.
• Places every reproducibility-critical method in the main manuscript.
• Reports all three candidate-budget sampling schemes and designates kernel_blocked as primary.
• Completes effective-rank matching diagnostics, caliper and one-to-one sensitivities.
• Adds 30-replicate finite-shot Monte Carlo analysis before and after PSD projection.
• Retains prospectively frozen external corroboration on three TableShift domain shifts.

Validation

• 66 tests pass locally and in CI on Python 3.11 and 3.12.
• All v0.6.0 artifact gates pass.
• The 41-page main manuscript and 5-page Supplementary Information compile without overfull boxes or unresolved references and have received final page-by-page visual QA.

Identity

• Tag commit: e010cc63534d64c817a36f20391de300de4ebe03
• Immutable Zenodo DOI: 10.5281/zenodo.21672470
• Concept DOI: 10.5281/zenodo.19147649

SHA-256

• kernel_shift_framework-v0.6.0.zip: FCF341E9104FC3D80C2F35180B26DF21DE9F276F3EE78CFF451FED213072590E
• kernel_shift_framework-v0.6.0-main.pdf: 1A161065821597D76F4E823142358F8645C64FDB6DE17F057A1B3AE1C40FE3FC
• kernel_shift_framework-v0.6.0-supplementary.pdf: 8AAC822E02C99AE4E5AAA897271E50FD3CB409601C3FC974D68C3598A7AACE45

v0.5.0 — Audited external validation and Q1 manuscript revision

Audited external validation and Q1 manuscript revision

This release freezes the complete v5 artifact used for the revised manuscript, Evaluation Choices Determine Apparent Quantum Kernel Robustness under Distribution Shift.

Scientific additions

• Prospective TableShift protocol frozen at commit e3253b0 before the external outcomes were generated.
• External validation on three tasks, two sample sizes, and five seeds: 30 complete experimental units and 37,800 summary rows.
• Fail-closed audit confirms no OOD-validation access, no target-dependent preprocessing, and no population-level p-value computed from reused splits.
• Frozen interpretation: replicated_protocol_sensitivity.
• Task-equal SVC estimate changes from −0.0119 ROC-AUC under honest validation to a +0.0208 apparent advantage attributable to validation-oracle selection.
• Specification-curve and oracle-budget analyses distinguish kernel behavior from evaluation-induced conclusions.

Release contents

• Springer Nature-compatible main manuscript and Supplementary Information.
• Frozen protocol, manifests, split hashes, aggregated external results, audit records, and figure-generation code.
• Reproducible Python 3.11/3.12 environments and CI-tested reporting pipeline.

Run the complete v5 reproduction and audit with:

python scripts/reproduce_v5.py --stage all

The release preserves the protocol-freeze commit history and is associated with the project DOI 10.5281/zenodo.19147649.

v0.4.0 — Fair comparison: no robust quantum-kernel advantage under shift

v0.4.0 — Fair comparison: no robust quantum-kernel advantage under distribution shift

A major methodological revision answering a second external review. The headline verdict changes, and this release supersedes the v0.3.0 "geometry governs robustness" story.

What changed

• Under a fair comparison, fidelity-based quantum kernels show no robust out-of-distribution advantage. When every configuration's regularization C is tuned by cross-validation on the training split alone, configurations are selected on an in-distribution validation split with no OOD labels, and both families search an equal candidate budget, the OOD difference collapses to near zero against both the customary linear+RBF baseline and a geometry-matched extended classical family (dataset-equal-weighted ΔOOD ≈ −0.004 under SVC for both references), classical-favoured or tied in 7 of 8 scenario-groups.
• The apparent advantage was an evaluation artefact of test-peeking (Best-by-OOD) selection and a fixed C=1. At matched effective rank, classical kernels are at least as accurate in every scenario-group.
• Two flaws in v0.3.0 are corrected and retracted here: the "hierarchical permutation test" flipped signs per-setting over correlated units (the p = 2×10⁻⁴ headline is withdrawn), and the 115-vs-60 candidate budget favoured the classical family in the "reversal." v0.4 uses matched budgets and reports effect sizes with conditional intervals over fixed case studies — no population p-value.
• Geometry is an association, not a mechanism: effective rank and OOD alignment predict performance in a regime-dependent way and favour neither family.
• New finite-shot fidelity-estimation model: effective rank inflates under finite measurement, so the exact geometry is a statevector idealization; alignment and accuracy survive.

Artifact

• Full v4 recompute (1080 runs): ID-validation split, internal C CV, 23 classical + 12 quantum geometries per run, and the finite-shot subset.
• New analyses: budget_matched_selection, hierarchical_effect_estimation, honest_family_comparison_v4, mechanism_robustness_v4, shots_stability_analysis; frozen spec in docs/ANALYSIS_SPEC_V4.md.
• scripts/reproduce_v4.py regenerates the confirmatory analysis, tables, and figures; CI runs the unit-test suite and a terminology guard.
• Manuscript rewritten to the F1 verdict (compiles clean, 59 pages).

Full changelog: v0.3.0...v0.4.0