Abstract
Systems and methods are described for detecting and correcting cascading proxy drift in multi-layer optimization pipelines. A directed acyclic graph represents proxy metrics across layers, including a ground-truth metric and downstream proxies. For each adjacent layer pair, a drift state is computed using sliding-window Spearman correlation, a trend estimate, and volatility, with adaptive window sizing via change-point detection. Cascade onset is detected based on a product of per-layer correlations and/or concurrent negative drift across multiple layers. Upon cascade detection, a drift source is localized using intervention-based tests including sequential freeze, counterfactual correlation estimation, and Shapley-value attribution (optionally via permutation sampling). Layer-specific corrective actions are initiated, including proxy recalibration using periodically collected ground-truth, proxy switching, optimization budget reduction, and optimization pause. A cascade risk score may be computed to trigger preventive control actions and to adjust ground-truth sampling frequency.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Cascading Proxy Drift Detection and Corrective Action System for Multi-Layer Optimization Pipelines", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10736