Online calibration of Token Service Share under bounded live perturbations

Investigate online calibration of Token Service Share with bounded perturbation to the live serving path, so that calibrated service-deficit profiles can adapt to model updates and workload-distribution drift during production operation.

Background

Token Service Share (TSS) is a calibrated proxy whose cross-model comparability depends on profiles that remain fixed during online execution. The paper notes that production Model-as-a-Service platforms experience frequent model updates and shifts in workload distributions, while the current slow-path recalibration assumes that such changes occur infrequently relative to the control-loop timescale. The unresolved direction is to develop an online adaptation mechanism—such as Bayesian updating from streaming latency feedback or lightweight canary replay—without introducing unsafe perturbations into the live serving path.