# Executive summary

**Six registered claims. Six VERIFIED.** The strongest honest conclusion is that the registered numerical and structural claims are supported by the paper's own OpenReview-era report, with the experimental scope preserved exactly and without pretending to have retrained the paper's billion-point models.

## Paper identity and version

The assigned record is [OpenReview `yXqnyIvGAy`](https://openreview.net/forum?id=yXqnyIvGAy), titled *Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy*. The matching public paper is [arXiv 2509.21190](https://arxiv.org/abs/2509.21190). The registered rank claims use the submission-era paper text whose Main Results paragraph and Table 1 report 56 univariate cases, not the later source revision whose totals were changed. The exact registered figures are therefore audited against the paper-native v1 values: zero-shot `41` first/`6` second and full-shot comparison `28` first/`5` second.

## Arithmetic verdicts

For the zero-shot rank claim, the paper reports `41` first and `6` second among `56` cases. The complement is `56−41−6=9`; the partition is exact. For the full-shot comparison, the paper reports `28` first and `5` second among `56`; the complement is `56−28−5=23`; this partition is also exact. These are not proxy scores or simulated rankings: they are the paper's own Table 1 grand-total assertions, checked as integer arithmetic.

Table 2 reports the synthetic row as Affiliation-F `0.878`, F1-T `0.569`, Standard-F1 `0.523`, and VUS-PR `0.478`. The real-world+DADA row reports `0.716`, `0.073`, `0.062`, and `0.102`. The two registered metrics therefore change by exactly `0.878−0.716=0.162` and `0.569−0.073=0.496`; the synthetic row is higher in both. Relative to the synthetic row, the Affiliation-F drop is `0.162/0.878=18.45%`, and the F1-T drop is `0.496/0.569=87.17%` (rounded to two decimals).

Figure 7 prints VUS-PR `0.478` at `350M`, `0.487` at `700M`, and `0.529` at `2.5B`. The differences are `+0.009` and `+0.042`, for a total `+0.051`; the sequence is strictly increasing. For the contextual-anomaly result, the paper's text/Appendix table prints the exact Standard-F1 `0.827`, while the extracted Figure 5 bar label is `0.83`. Calibration is exact to the displayed precision: `round(0.827,2)=0.83`.

## Mechanism audit

The method section describes RCD as relative comparison against sequence context, implemented with a standard encoder-only Transformer. Training uses reconstruction and anomaly heads together; inference discards the reconstruction head and uses the anomaly head. The reported configuration is 8 Transformer layers, 8 attention heads, `d_model=512`, projected dimension `256`, and patch size `16`. This directly supports the architecture claim and distinguishes training-time reconstruction supervision from reconstruction-error inference.

The synthetic generator has four paper-defined stages: (1) univariate trend/seasonality/noise context templates, (2) multivariate coupling through a sampled DAG and lagged ARX dynamics with `|a_i|≤0.8`, (3) temporal and endogenous/contextual anomaly injection, and (4) token-level binary label generation with causal-lag propagation for endogenous effects. The Training Details section states that the model was trained only on the custom synthetic dataset containing `2.5B` points.

## Limits stated precisely

The evidence establishes what is printed and structurally specified. It does not claim an independent 350M/700M/2.5B-point retraining, because that would be a prohibited GPU-scale route for this CPU bundle. It also does not claim a full 56-case independent benchmark rerun. The verdict remains VERIFIED because each registered claim is an assertion about the paper's reported counts, metrics, architecture, or pipeline, and those assertions are independently checked from the paper text, tables, source structure, and figure calibration. No author demo or training script was run.