The pattern
A team announces quantum advantage on benchmark X. The press release runs hot. A week later somebody runs benchmark X with a stronger classical kernel and matches the result. A month later somebody else uses a different feature representation and beats it. The original team quietly publishes a follow-up with narrower claims. Coverage moves on.
This has happened enough times that the right default response to a viral quantum-advantage claim is interest plus skepticism, not interest plus retweet. Six questions get you most of the way to a defensible read.
1. Was the classical baseline strong
The most common deflation pattern is a weak baseline. RBF SVM with default hyperparameters is not a strong baseline. Hyperparameter-tuned RBF, plus polynomial, plus a small MLP, plus a gradient-boosted classifier, plus a deep model where data permits, that is a baseline. If the paper does not report that suite, the quantum result has not been stress-tested.
2. Was the quantum implementation honest
Two specific tells. First, did they run on real NISQ hardware or only in noise-free statevector simulation. Simulation is fine for theoretical claims, but a "we ran on quantum" headline should be backed by hardware execution, with reported circuit depth, gate count, and fidelity degradation. Second, did they transpile and report the actual circuit, or only the abstract one. The abstract circuit is not what the hardware runs.
3. Was the metric correlated with usefulness
A 0.001 increase in cross-entropy on a synthetic toy dataset is not the same as a 10 percent AUC improvement on a real anomaly-detection benchmark. A speedup on a contrived sampling task is not the same as a speedup on a useful one. The metric should connect to a downstream decision someone actually cares about.
4. Was the run reproducible
Code released. Hardware backend named. Calibration date pinned. Random seeds reported. If it is not reproducible, it is not a result. Quantum hardware is noisy enough that run-to-run variance can swamp the claimed effect, and the only way to defend against that is to make the run easy to repeat.
5. Was the dataset realistic
Synthetic Gaussian blobs do not generalise to industrial control system telemetry. MNIST does not generalise to anything in 2026. Look for cross-testbed validation: same algorithm, different domain, similar gain. If the gain only appears on the dataset the algorithm was designed against, you have a fitting story, not a generalisation story.
6. Did anyone re-test on hardware that exists today
Many advantage claims rely on extrapolation. "If our circuit could be run on a hypothetical 1000-qubit error-corrected device, we would beat classical by N." That is fine as a research direction, not as a present-tense headline. The honest version is: did the team run the circuit on a 156-qubit ibm_fez or comparable, and how did the hardware result compare to the simulation result.
The honest case
An honest quantum-advantage claim looks like: small, specific, hardware-validated, reproducible, with a strong classical baseline, on a dataset that is not engineered for the algorithm, and a metric tied to a real decision. My IEEE Access 2026 paper on 8-qubit ZZFeatureMap kernels for ICS anomaly detection tries to meet that bar. Cross-testbed validation across SWaT and HAI. RBF SVM as the classical baseline. AUC as the metric, since false-positive cost matters in SCADA. Hardware execution on IBM ibm_fez (156 qubits, depth 76, 28 CNOTs after transpilation). Statistically significant gap on the harder testbed (p = 0.003). No "advantage" claim, just a reproducible structural improvement.
Whether the gap survives shot-noise scaling is the next experiment. I do not know yet. That is a feature, not a bug. The paper says so. The advantage-claim deflation pattern stops the moment researchers stop overclaiming and reviewers stop letting them.
The reviewer's job
If you are reviewing a quantum-advantage paper this year, those six questions are your job. Do not let "we beat classical on benchmark X" land without each one answered. The literature gets stronger when ambitious claims ship with their limitations made visible, not when they are excluded and reappear elsewhere with the limitations omitted. That is the actual peer-review contract, quantum or classical.