What does it cost to trust a quantum model? Shot-priced certification for hybrid QML

Hybrid quantum-classical models are increasingly compared through reported Fisher geometries, kernels, and capacity measures. However, every entry of such a report is estimated from finitely many projective measurements, and the shot budget behind it is almost never stated. We price that budget with an $\varepsilon$-matched certification protocol: both arms are grown to the same relative Fisher error and billed in their native units, circuit executions for the quantum arm against per-sample gradient evaluations for the classical arm. Under coordinate-wise parameter shift, certification costs $B_Q\approx p^{2}V/(\varepsilon^{2}G)$ circuit executions up to a measured gradient-alignment factor, independent of the minibatch size until a bias floor sets in; whether a joint strategy such as classical shadows pays less remains open. Once readout correlation $V$ and signal $G$ are measured, two circuit families whose fitted cost exponents differ by a full power of $p$ give a single measured constant at fixed minibatch, $0.338$ and $0.349$. In a nine-design-point mirror-circuit case study on ibm_marrakesh, attenuation inflates the certification budget by a measured $2.07\times$ (cluster-robust CI $1.41$-$3.02$); the same grid on ibm_fez gives $2.49\times$ ($1.13\times$ the seed-matched Marrakesh mean). The same accounting changes conclusions: matching feature budgets reverses our own early unmatched encoder comparison, and the identity of the winning classical arm then changes across four synthetic targets and a Fashion-MNIST split while the fixed, untrained noise encoder never wins. Four pre-specified controls in our own stack failed or were demoted, among them the $V$-scaling check that demotes the two-class readout picture to a pair of limiting regimes, and two more narrowly missed. The protocol is released as the sned-certify tool, with code and locked result files in the accompanying archive.