Cross-expert Jacobian alignment

Measures functional (not merely representational) similarity between two experts in a mixture-of-experts layer by comparing the cosine similarity of their local input-output Jacobians at shared inputs, rather than comparing their weights or output activations directly — a near-zero value indicates the experts implement decorrelated functions even if their routed representations occupy overlapping subspaces.