Beyond the centre: learned uncertainty in AIS position-consistency checks, its detection gains and its attack surface: derived results, analysis scripts and extended version

Abstract A shore-side AIS position-consistency check alarms when the displacement between two reports leaves a calibrated ellipse around a kinematic prediction. Rather than refine the prediction, we learn the ellipse's width as a function of the motion context, with split-conformal thresholds fitted on held-out ships. We introduce the reach of a check, the distance a position-only drifter covers in 30 minutes before half of its drifts trigger a CUSUM alarm, and evaluate gradient-boosted, recurrent and transformer width models on 94 million reports from four feeds, with ship-disjoint splits, a model frozen before the final test set, and a further test on an independent day and feed. A learned tree width more than halves the reach, from 582 m to 264 m, and the gain holds on the independent day; a learned centre does not raise detection. An adversary can inflate a learned width before jumping. Under inflate-then-jump, delayed-report and adaptive attacks, widths conditioned on the motion context retain detection while widths driven by recent residuals collapse, although at matched false-alarm rates the margin over the baseline is small. Capping the width at a multiple of a history-free width gives a provable bound on the decision region against position-only attackers. Contents Manuscript PDF: the paper as submitted. Extended-version PDF: the full model lists, the stateful-filter results, the repeats on test set A and the table of criteria and outcomes. Results-and-analysis zip: the result files (aggregates only, no AIS rows), the scripts that generate every table and number of the paper from them, the dated analysis plan and its addenda, and the LaTeX source of the extended version. Not included AIS rows, caches and fitted model weights. The model-training code, which belongs to a commercial product under development; it is available to the editor and reviewers on request. Data The runs use an AIS archive of 10 to 26 October 2026 (aisstream, OpenWaters, Digitraffic, BarentsWatch). Only the Digitraffic feed (Fintraffic, CC BY 4.0) permits redistribution.