Is It the Entanglement? A Calibrated Interaction-Strength Ablation and Its Classical Control

Variational quantum classifiers are often evaluated on small clinical datasets, and the entangling layer is usually assumed to earn its place. Removing the CNOT ring provably reduces our circuit to an additive model in the original feature basis at an identical trainable-parameter count. On a synthetic family with an analytically defined interaction share $\iota=\alpha^2$ and fixed Bayes error, the entangled-minus-product AUC difference $\Delta_{\mathrm{ent}}$ tracks $\iota$ closely, from $-0.05$ at $\iota=0$ to $+0.29$ at $\iota=1$ (1,440 fits). That result is not about entanglement. The same contrast between two classical models on identical draws reproduces the curve with a comparable effect (Spearman $\rho=0.74/0.79/0.85$ classical versus $0.49/0.77/0.89$ quantum), so the dose-response measures function-class additivity. A second mechanism produces the same kind of false signal: sweeping the learning rate at $\iota=0$, the apparent cost of the ring runs from $-0.007$ (95% CI $[-0.031,+0.002]$) to $-0.088$ across a 40-fold range and is indistinguishable from zero at the smallest rate. On Pima the sign of $\Delta_{\mathrm{ent}}$ holds at every rate but its magnitude falls sevenfold. Entanglement ablations therefore require a matched classical control and a learning-rate sweep before any conclusion about entanglement.