Still settling
- Clock rate not yet measured. Until it is,
nothing takes out the receiver's sample-clock drift — around 24 ppm,
or 1.4 ms a minute — so the upstream NTP offset climbs steadily and
the difference between the classes goes with it.
- NTP filter keeps eight polls; at a 64 s poll that is about
8 minutes just to fill once.
- Sample-clock fit regresses over a 5-minute window, per source.
- Sources confirm a first decoded time against their own history
before counting, and one that recovers waits 180 s before it is taken
back, so the consensus can change shape for a while.
Nothing below is wrong, but the difference between radio and NTP
is not worth reading yet. Give it about an hour, then use the 24-hour history.