In short
The results come from a perfect, noise-free simulation on one dataset with 2–12 qubits. The VQCs were probably under-trained, so their scores are lower bounds. And the Random Forest alone does as well as the full pipeline.
RF alone performs as well. The most obvious threat to validity is that RF by itself performs as well as the RF + VQC pipelines. The votes arm matches RF by itself because the RF does the feature learning. A per-sample test on a single fold (3,109 rows) showed that the votes 4-qubit arm agreed with RF by itself on 99.77% of the samples. Nothing in these results supports a reading that the quantum part helps. The claim remains that the RF pipeline is better than PCA as a pre-stage for a VQC; the question was always how this compares to PCA, not to RF by itself.
COBYLA budget. The final configurations used all 500 evaluations in 100% of the splits, except the leaf 8-qubit arm (98%). The VQCs were likely under-trained, and their scores are lower bounds. A fixed budget hurts larger models more. However, at 2 qubits PCA used the full budget in only 32% of splits and still scored only 0.760, so the budget alone does not explain PCA’s gap. Only COBYLA was tried.
Model selection. Configurations were chosen on 10 validation splits, so small differences within an arm are within noise. Other ansätze and feature maps were never tried. Every arm was searched over the same grid.
Leakage was handled: the RF, PCA and min–max scaling were all fitted on the training fold only, model selection used only inner validation splits, and the votes VQC was trained on out-of-bag votes.
Reproducibility across machines. Floating-point rounding differs between CPUs. Per-split metrics differed by up to 0.055 between machines, while means stayed within 0.002. Fixed seeds (42–51) were used, and the conclusions rest on the means and the paired tests, not on individual splits.
Timing cannot be compared between arms because machine speeds differed by about 5–7 times. A single-split test on one machine found PCA cheaper for classical preprocessing (about 0.7–1.9 ms per sample end to end against 5.2–5.4 ms for the RF arms), but these times are tiny next to the sandbox run needed to collect API calls. The accuracy is the claim, not the time.
Noise-free simulation. All experiments used an exact statevector with no shot noise, no gate noise and no mapping onto real hardware, so the results are likely an upper bound for current (NISQ) hardware.
One dataset, 2–12 qubits. Only GBAC was used (15,542 samples, API-call n-grams hashed to features, binary labels). The 12-qubit condition was run for only 10 splits and is exploratory. Trainability at larger scale, including barren plateaus, was outside this study’s scope.