The Thunderbolt, Star-4, Trueposition, and HP GPSDOs/Smartclocks are all known to do this. Typically the learning period is around 24 hours before the compensation is allowed to run. My HP GPSDO seems to take a week for best performance. Separating tempco and aging drift seems to me to be a rather thorny problem to do well. --------------- > I never measured the temperature of oscillators and used the information to compensate holdover but it makes sense with a specific oscillator and enough run time to collect the data to categorize the oscillator for temperature and ageing.