> It looked to me like the fleet did consistently fall behind, which seems like thoughtful design

Thoughtful, but I think it isn’t necessary with ‘modern' NTP algorithms. They lose time by slowing down a clock derived from their hardware clock for a while, not by moving the hardware time back (https://serverfault.com/a/94691, https://community.hpe.com/t5/operating-system-hp-ux/ntp-atom...)