Why laboratory values of G scatter beyond their stated uncertainties

Laboratory determinations of G (torsion balances in time-of-swing, angular-acceleration and servo modes, beam balances, simple pendulums, atom interferometry) spread over about $5 \times 10^{-4}$ relative, roughly ten times their typical quoted uncertainties of $10^{-5} \text{ to } 5 \times 10^{-5}$, near $G = 6.674 \times 10^{-8}\,\mathrm{cm}^{3}\,\mathrm{g}^{-1}\,\mathrm{s}^{-2}$. Identify the unaccounted systematic effects (or establish any genuine dependence of the measured G on method, material or geometry) that explain the excess scatter.

A consensus set of independent measurements with different methods agreeing within quoted uncertainties at the $10^{-5}$ level, together with an identified correction to the outliers.