At the heart of copyright infringement cases is “substantial similarity” between the plaintiff’s and the defendant’s works. But while every circuit agrees on the centrality of substantial similarity, that basic agreement conceals surprising differences in what exactly we mean by substantial similarity and how it is to be proven in court. And the majority approach, defined by the Second Circuit in Arnstein v. Porter and the Ninth Circuit in Sid and Marty Krofft, has the analysis of proof exactly backwards – permitting analytic dissection of the works and expert testimony where the question is one that should be handed to the members of the jury, and falling back on the “ordinary observer” test on the very questions that require careful dissection by the court. I argue that the Arnstein and Krofft tests make no sense. A better model draws from software copyright cases, which give filtration and dissection of unprotectable elements a more central role.