This may be similar to this issue I have been working on:
As I said in the github issue, I am not sure if this is a problem or not, because those CSV files are intermediary files used to calculate the final scores in results/final_scores.json. If you find that the final scores are incorrect, please report how you ran NAB so I can try to replicate the problem. But I am not very concerned with discrepancies in these intermediary CSV files.
thank you @rhyolight. i just wonder that Numenta created the corpus ( they labeled the 58 datasets) so they know exactly the amount of abomals and the time of every abnomal. Do you have any infomation about that ?