Year: 2026
Type: report
Abstract: Coordinated manipulation can shift platform-level signals while leaving each participant locally inconspicuous. We study this aggregate footprint as a measurement problem: how can heterogeneous interval summaries provide calibrated evidence against a declared anchor-and-channel model, and when does persistent induced-summary excess above the matched baseline yield positive linear expected evidence growth? In each interval, the detector observes a released histogram, its count and resolution, and a platform-declared reference model. The summary is lifted to a distribution on a common bounded signal space and compared with the reference using Wasserstein–1 distance. Raw Wasserstein deviations are geometrically comparable across reporting regimes, but their null scale is generally nonzero and regime dependent because finite sampling and the reporting channel affect their distribution. We therefore subtract the expected deviation obtained by replaying the same reporting channel under the declared reference. The resulting signed increments form a cumulative evidence process. Undermatched-channel and non-anticipativity conditions, the exact process is a bounded-increment martingale; a valid conservative guardband yields a supermartingale, conditional on its guardband event when it is estimated. These properties provide finite-horizon and anytime-valid false-alarm control. Under the stated conditions, persistent induced-summary excess above the channel-matched baseline yields positive linear expected growth of the exact evidence process, whereas vanishing excess together with within-intervaltracking eliminates positive linear expected growth. We also characterize structural information loss. Non-injective reporting can make distinct latent distributions indistinguishable to every summary-based detector, and fixed-resolution histograms imposea geometric resolution floor. The framework extends to bounded multidimensional signals through either joint Wasserstein monitoring under a declared product-space metric or coordinatewise marginal monitoring with simultaneous error control. Finally, Wasserstein duality conditionally bounds shifts in induced expected values of Lipschitz outcome functions, supporting an integrity-budget interpretation of a justified induced-summary Wasserstein allowance. The guarantees are aggregate, anchor-relative, and channel-conditional: an alarm provides false-alarm-controlled evidence against the declared anchor-and-channel model; it does not establish that an observed departure is persistent, caused by coordination, attributable to particular identities, or adversarial in intent.
Author: Zhang, Rui
Institution: University of Delaware
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7218205994
