Weakening in Real Time: The Association Between Artifact Reuse and Citation in Artifact-Dense Research, 2015–2024

Preprint of a manuscript submitted for publication. Data and analysis code are archived separately (see Related works). Research artifacts, chiefly public code repositories, are how computational research is used; citation remains the standard impact proxy. Do they still track each other? Linking 20,529 arXiv papers to their author-designated GitHub repositories, we estimate the rank association between annual fork and citation flows, for the 8,864 with both series, in every cohort-by-period cell, 2015–2024. Across cells the association peaks at 0.45 in 2018 and falls to 0.26 by 2024. The structure surprises: it falls as steeply within cohorts as across them, and no cohort gradient survives conditioning on calendar year. Across 55 cells period explains 43.2 percent of variance against 17.8 for cohort, and with period included the cohort coefficient is -0.0017 (p = 0.74). The three linear trends are not separately identified; we argue for period on parsimony and on identified nonlinear structure. Two candidates fail on the period axis: no discontinuity at general-purpose language model release, and no gradient in field-level exposure, though that proxy is weak. Nor does the citation mix degrade. In the full frame of 160,150 papers, no trend appears among heavily reused and cited works (-0.0031 per year, 95% CI [-0.0156, +0.0094]). This is a property of the population, not its high-impact tail.