4.7x fewer numbers, 6.7% error

error = np.linalg.norm(T - recon) / np.linalg.norm(T)      # 0.067
ratio = T.size / (core.size + sum(u.size for u in Us))      # 4.71

4.7× fewer numbers, at 6.7% error. 480 numbers become 102.

That is a real compression on a real tensor, and it is the part everyone quotes. It is also not the interesting part.