Robust characterization of average gate set noise and cross-talk using gate-set shadows

Gate-set shadows extend randomized benchmarking toward practical noise tomography. On a 36-qubit superconducting processor, a single randomized-circuit experiment reconstructs local noise marginals and identifies correlated errors and cross-talk. Error-generator decompositions separate coherent, stochastic and correlated components, revealing patterns beyond physical nearest neighbours. Comparing these rates with standard mechanisms such as detuning, over-rotation and decoherence quantifies unexplained noise and suggests that familiar models can miss important cross-talk.