Persisted anomaly scores for "When Do Zero-Shot Time-Series Foundation Models Pay Off for Anomaly Detection?"

The 5504 persisted anomaly-score arrays behind the paper's results. Each file is named {dataset}__{series}__{detector}__{tier}.npz, holding the score vector under the key scores and the ground-truth labels under labels, for eight zero-shot configurations of four pretrained time-series foundation models and four classical detectors across 360 series from UCR Anomaly Archive 2021, SMD, SMAP/MSL and PSM. They let you recompute every metric in the paper — including metrics the paper does not report — without running a single model. This is the reuse the work most hopes for, since its finding is that the bottleneck lies in the conversion from forecast to anomaly score rather than in the forecast: substituting your own metric or your own score conversion, on the same detectors and the same series, takes only these files. Unpack into results/scores/ of the code repository at https://github.com/Thanatosc/tsfm-anomaly-detection. The four benchmark corpora are not redistributed here; they carry their own licences.