Resource map of quantum reservoir computing against tuned classical baselines

Abstract Quantum reservoir computing promises an exponentially large feature map at no training cost, but reported advantages rest on classical baselines of unmatched size or tuning, on exact expectation values, or on resource counts that omit measurement. Here we introduce a benchmark protocol that charges shots, experiment restarts and readout size together and apply it to 3451 exactly simulated cells (57 quantum reservoirs, 49 tasks) against three tuned classical reservoir families. No classical- or quantum-input task yields a win in both accuracy and resource scaling; the only winning cells are Forrelation estimators on oracle pairs drawn fresh at every step with no exposed structure, which beat every implemented classical estimator in every seed from field sizes of 16 thousand, with a matching classical query budget growing as the square root of the field size. Few-photon Kerr dynamics beats its mean-field twin only until the twin is tuned, spin readouts pay up to three orders of magnitude in shot noise where quadrature readouts pay threefold, and every structured oracle family falls to a classical shortcut. Quantum reservoir advantage therefore lives in the input access model, not in the reservoir; we give the superconducting-circuit parameters at which the protocol can be run.