Statistical Engine Verification

Independent reference implementations. Every statistical engine is validated against an independent reference implementation written from scratch, following the published mathematical formulation of each algorithm. The reference implementations do not share code with the platform's engines: they serve as an independent validation oracle. Where multiple algorithms exist for the same computation, different ones are used in the reference and the platform (e.g., Jacobi eigendecomposition for reference PCA vs. power-iteration deflation in the platform), strengthening the independence of the comparison and ensuring that agreement is not an artifact of shared numerical paths.

Equivalence criteria. Equivalence is assessed against three criteria matched to the output type. For continuous outputs (p-values, fold changes, ordination scores), a Pearson correlation coefficient r >= 0.999 between the platform's output and the reference implementation's output is required for a PASS. For deterministic computations (PCA scores, distance matrices), the maximum relative error must not exceed 10^-6. For discrete outputs (cluster assignments, pathway membership), 100% exact match is required (with label permutation permitted for clustering, since cluster labels are arbitrary). These thresholds are conservative: r >= 0.999 means the platform and reference explain 99.8% of each other's variance, far exceeding the level at which downstream biological interpretation would be affected.

Agreement assessment. The use of Pearson correlation as the primary equivalence metric follows established practice for method-comparison studies. Bland-Altman analysis (Bland & Altman, 1986) provides a complementary view of agreement by quantifying the bias and limits of agreement between two methods; the correlation threshold r >= 0.999 ensures that the two methods are effectively interchangeable for all practical purposes. The combination of correlation, maximum relative error, and exact-match criteria covers the full range of output types encountered in omics analysis.

Datasets. Deterministic, gold-standard datasets modeled on published reference designs (SEQC/MAQC, MetaboLights, LIPID MAPS, Olink/ProteomeXchange, CAMI/QIIME2, 10x Genomics pbmc3k). All data is embedded in the platform: no download or server call required. Datasets are downloadable as CSV for independent reproduction in R, Python, or any statistical environment.

Execution. The benchmark runs entirely client-side in the browser. It uploads nothing and needs no server call: the datasets and reference implementations are bundled with the platform, so any reviewer or user can re-run the benchmark at any time. (The platform's separate server-side analysis paths are described in the Security Architecture Report.) This page does save the summary of a run as a platform record so results persist between visits.

Reproducibility. All datasets are deterministic (no random number generation), so benchmark results are reproducible across runs, machines, and browser sessions. Results are displayed on the platform's Validation page and are downloadable as CSV and PDF for inclusion in supplementary materials. This level of validation transparency is designed to meet the code availability and reproducibility requirements of journals such as Nature Methods.