Next, for AI papers, a reproducibility requirement. So much code and details are fudged and paper's cannot be reproduced. Ran the training with some other config, or other data, etc. to make their mechanism or intervention seem better.