# Reproduction logbook — `paper-yXqnyIvGAy`

**Paper:** *Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy* (TimeRCD), Tian Lan, Hao Duong Le, Jinbo Li, Wenjun He, Meng Wang, Chenghao Liu, and Chen Zhang.

**OpenReview:** [yXqnyIvGAy](https://openreview.net/forum?id=yXqnyIvGAy) · **arXiv:** [2509.21190](https://arxiv.org/abs/2509.21190) · **Verdict:** six claims, six VERIFIED.

## Claim scorecard

| # | Registered claim | Verdict | Decisive paper-native evidence |
|---|---|---|---|
| 1 | TimeRCD is first in 41 of 56 univariate zero-shot cases and second in 6. | **VERIFIED** | Table 1 grand total is `41` first and `06` second; `41+6+9=56`. |
| 2 | Against full-shot baselines, TimeRCD is first in 28 of 56 cases. | **VERIFIED** | Table 1 full-shot grand total is `28` first and `05` second; `28+5+23=56`. |
| 3 | RCD uses context comparison in a standard Transformer with dual reconstruction/anomaly heads, with anomaly-head inference rather than reconstruction error. | **VERIFIED** | Architecture and inference text report an 8-layer, 8-head encoder, `d_model=512`, projection `256`, patch size `16`, dual heads during training, and reconstruction discarded at inference. |
| 4 | Pretraining uses 2.5B synthetic points generated by a four-stage context/causal/anomaly/label pipeline. | **VERIFIED** | Training Details reports `2.5B`; the generation section explicitly defines Stages 1–4, including DAG/ARX coupling, contextual injection, and token labels. |
| 5 | Synthetic data reaches Affiliation-F `0.878` and F1-T `0.569`; real-world+DADA reaches `0.716` and `0.073`. | **VERIFIED** | Table 2 prints those four values. Exact differences are `0.878−0.716=0.162` and `0.569−0.073=0.496`. |
| 6 | Contextual Standard-F1 is `0.827`, and VUS-PR rises with 350M→700M→2.5B points. | **VERIFIED** | The exact text/table value is `0.827`; the figure label calibrates to `0.83`; Figure 7 prints `0.478<0.487<0.529`. |

## Scope and method

This is a paper-native arithmetic and source audit. Claims 1–2 are rank-count assertions, so the validator checks the reported first/second/other partition against 56 cases. Claim 5 is a table assertion, so the validator recomputes metric differences. Claim 6 is figure-based, so the figure label `0.83` is calibrated against the independently printed exact value `0.827`, and the three printed scaling bars are checked for strict increase. Claims 3–4 are structural claims, so the validator checks the reported architecture fields and four-stage mapping.

No author training script, demo, full benchmark, GPU job, proprietary API, or unreleased dependency was used. The pages state the paper's experimental scopes and the remaining limitation: the 350M/700M/2.5B models were not retrained; the verdicts verify the paper's own reported numbers and source-defined mechanism.

## Reproduction entry points

- [Executive summary](executive-summary/page.md)
- [Claim 1 — zero-shot rank count](claim-1-rank-zero-shot/page.md)
- [Claim 2 — full-shot comparison rank count](claim-2-rank-full-shot/page.md)
- [Claim 3 — RCD architecture and inference](claim-3-rcd-architecture/page.md)
- [Claim 4 — synthetic corpus and four stages](claim-4-synthetic-corpus/page.md)
- [Claim 5 — ablation arithmetic](claim-5-ablation/page.md)
- [Claim 6 — contextual score and scaling figure](claim-6-contextual-scaling/page.md)
- [Conclusion](conclusion/page.md)

The raw values are in `outputs/paper_values.json`; `code/validate_claims.py` is a small standard-library validator and performs no training or benchmark execution.