paper-yXqnyIvGAy
Paper: Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy (TimeRCD), Tian Lan et al., ICML 2026, OpenReview yXqnyIvGAy.
Verdict: all six registered claims VERIFIED.
This bundle is a CPU-only audit of the paper's own reported counts, metrics, architecture description, synthetic-data stages, and plotted values. It does not retrain the 350M/700M/2.5B-point models or run a full benchmark. The decisive evidence is preserved in pages/**/*.md; the small validator in code/validate_claims.py recomputes the arithmetic from outputs/paper_values.json.
The paper-native version used for the registered rank claims is the submission-era OpenReview/arXiv v1 text: 56 univariate cases, 41 first/6 second in the zero-shot comparison, and 28 first/5 second against full-shot baselines. The exact contextual-anomaly value is Standard-F1 0.827; the scaling plot prints VUS-PR 0.478, 0.487, 0.529 at 350M, 700M, and 2.5B points.
Start with the index, then the executive summary and six claim pages. No Space was created, uploaded, modified, or published by this work.