Limitations

Scope

QMint is a research platform. It is not money, not a payment service and not a deployable quantum currency. It reproduces and checks published results; it does not claim to have invented quantum tokens, S-money or quantum payments. The statements QMint must never make are listed in NOV-DISALLOWED.

Hardware

• No quantum hardware result exists in this build. Gate G6 (optional authorised IonQ run) is SKIPPED: no API key and no paid approval.
• The IonQ path is a dry-run cost and approval plan; submission is disabled. The Qollab integration has no public API and is blocked.
• Circuits are simulated with Qiskit Aer (ideal or with declared noise). Declared noise is hypothetical, not a device model.

Proofs and certificates

• Ideal bounds (W1, W1_CV, S1) are published theorems reproduced in code. Lane C's SDPs are certified only for n = 1 and n = 2; larger n is refused by a size guard.
• W2 (noisy/threshold Wiesner) has no bound. MVW Sec. 3.5's binomial tail covers a joint-success game, not per-copy thresholds.
• D1 (Schiansky 2023) is not certified by QMint. This is an unresolved reproducibility finding, not an invalidation of the original work: the published p_d = 5.9×10⁻⁴⁵ recomputes only with N = 4.2×10⁶ (the SI table lists 4,482,440), and we could not confirm that the Chernoff step applied to the weighted error-plus-loss sum meets that bound's stated hypothesis (a conservatively rescaled version gives 3.3×10⁻⁷). Coherent attacks are reduced by a de Finetti argument we did not formalise. These questions are open for review by the authors and independent experts.
• Kent et al.'s experimental tuple lies outside their Theorem 1 domain, as the paper itself says.
• Certificate compiler (lane I): it matches the exhaustive oracle on all small grids and every released certificate passes the independent checker, but its apparent photon savings (1.8–10.6× vs naive sizing on Kent Fig. 5 point devices) do not survive calibration uncertainty: under the uncertainty boxes every selector is infeasible. Only hypothetical devices clear the pre-registered bar, and the method is standard parameter optimisation, not a new technique.
• An independent red-team audit (lane M) found five high-severity defects in our own code; all are fixed with regression tests. Lower-severity items remain open and are tracked as strict expected-failure tests.
• Not done: no physical quantum hardware run (free IonQ noise-model simulators only), no S-money double-spend simulation in the network model, no more-than-two-site scenarios, no third-party reproduction, no manuscript yet.

Data

• The Zenodo 7979319 archive reproduces both published QBERs, but loss, detector efficiency and multiphoton probability cannot be observed from it; they stay null, and the paper's values are kept separately as PUBLISHED_AGGREGATE.
• Photon runs are overdispersed (χ²/dof ≈ 10): run-to-run heterogeneity driven by a few outlier runs, not temporal drift. A run-level Welch test of first vs second half finds no difference (p = 0.334 D/A, 0.587 H/V; an earlier pooled-iid p = 0.003 was invalid and was corrected after red-team review). Calibration intervals are widened accordingly.
• RIPE Atlas one-way delays are RTT/2, an assumption that is flagged wherever they appear.
• Several research entries were checked at abstract or metadata level only; their status says so.

Quantum vs classical

• The S-money network model covers timing only and says nothing about security. RIPE-based numbers are not a matched benchmark: trust model, privacy and processing differ.
• Classical baselines show that coordination alone is not S1's distinction; what differs is information-theoretic vs computational security. Rows that are not comparable are marked so in the comparability matrix.
• A measured quantum transaction-time advantage is Jiang et al. 2024's result, not QMint's.

Platform

• Where a lane is not wired into the API yet, the response says DEMO_ADAPTER or NOT_AVAILABLE with the reason, instead of a number.
• The uniform lane entrypoints are not merged yet, so qmint simulate and similar commands exit 3.
• The site is dark-only; LaTeX in research records is shown as source, not typeset.