A reliability-tested changepoint benchmark for published velocity-dispersion profiles of Milky Way globular clusters

ABSTRACT

Published velocity-dispersion profiles of Milky Way globular clusters combine heterogeneous line-of-sight and proper-motion products, so apparent radial slope changes can reflect source, component, or sampling effects. When reliable, such radial structure can motivate physical tests of a compact central mass or of tidal and escape-driven effects in a cluster’s outskirts. We present a reproducible changepoint benchmark whose primary unit is a cluster, kinematic component, and literature data product rather than a pooled cluster profile. The reconciled diagnostic set contains 48 clusters, 1593 canonical radial bins, and 94 primary source/component units. We compare a single logarithmic slope with a continuous segmented model, counting the selected breakpoint as a fourth parameter. Each fit is assessed using the corrected Bayesian information criterion (BIC), absolute goodness of fit, 1000 bootstrap and 1000 parametric-null realizations, and leave-one-bin-out, outer-bin, and minimum-side-bin sensitivity tests. Radial and tangential proper-motion components are treated separately using official tables or derived dispersion and anisotropy-ratio products; an independent Gaia EDR3 star-level reconstruction is retained as a caveated audit. Although 36 primary units have corrected $\Delta {\rm BIC} \ge 6$, only 16 satisfy the complete supported gate, including eight high-confidence units. Nineteen additional units retain formal evidence before a severe reliability veto, and 16 of these undergo outer-leverage evidence collapse. Cluster-level synthesis gives six Tier A, five Tier B, and 37 audit-only systems. These tiers measure changepoint identifiability, not data quality, dynamical state, or population occurrence rates. The machine-readable catalogue, complete fit atlas, reason codes, and deterministic pipeline provide a pre-modelling and posterior-predictive audit layer.