Extend drift diagnostics to multi-week collection windows

Extend the Kolmogorov–Smirnov, Population Stability Index, and negative-binomial drift diagnostics to the multi-week collection windows of the S2, S4, and S5 brand-recommendation auditing studies.

Background

The paper tests temporal stationarity in the S1 Valentine’s Day data, collected over approximately two weeks, and detects no statistically significant drift. However, the S2, S4, and S5 datasets were collected over longer windows and were not subjected to the same diagnostics.

Applying the drift battery to those datasets is necessary to determine whether the S1 stationarity result extends to longer collection periods and to evaluate whether model-version or provider-side temporal changes affect the audit measurements.

References

Drift diagnostics for the multi-week collection windows of S2, S4, and S5 are deferred to future work.

— The Dice Roll Method: A Standardized Protocol for Repeated-Query Auditing of Large Language Model Brand Recommendations  (2609.04047 - Żatuchin, 3 Sep 2026) in Section 7.8, subsection “Stationarity and Drift Diagnostics”