So your total test time is 5 hours?
its more 12-15 hours.
ramping down and up and with 1 hour rest at minimum and maximum temperature + some overlapping because at the beginning the ramp speed starts at zero and behaves different when finally ramping down.
Hey Andreas,
Very interesting results. If I understand correctly:
+ no hysteresis between ambient and the outside of the ref
+ some thermal hysteresis between the outside and inside of the ref measured with Vbe
+ apparently some unaccounted for hysteresis in the ref itself
+ some non-linear ref tempco but approx -0.03 ppm per K (assuming the meter is at a stable temp)
Is that right? 6K per hour is quite a rapid temp rise for a lab in practice of course. Does the hysteresis loop change much if you increase/decrease the ramp speed? Could the apparent hysteresis in the ref just be an artifact from time-based meter drift?
p.s. a left scale in ppm nominal ref voltage would be easier to read 
I think you are mis-interpreting the diagrams.
They are both the same except for the x-axis.
In one case (tPCB) the NTC near the LTZ on the PCB is on the x-axis.
Since I did not remove the Y-axis for this sensor he shows no hysteresis but the other sensor.
In the other case (tAMB) I have a second NTC outside of the aluminium case of my reference as x-axis.
So this sensor shows no hysteresis but the PCB-sensor lags against the ambient temperature.
Vbe is never measured in the diagrams (This would change the setpoint of the LTZ).
The hysteresis on the reference appears mainly because the temperature is neither the same of the one or other NTC.
The other part could be some effect due to temperature gradient.
I think you did not mention the 2:1 divider in the measurement (so you should not divide by 7200 mV but by 3572 mV
when calculating T.C. (around 0.06 ppm/K).
My "meter" is a LTC2400 based ADC with temperature compensation in this case.
It is rather noisy (the ADC itself + the around 0.1ppm quantisation noise of temperature compensation giving around 0.25ppm standard deviation) but has almost no drift over time.
Changeing the ramp speed has it´s limits.
Increasing is not possible for the whole temperature range (the peltier cooler is at its limit with 0.1 K/minute).
Slowing down would need running over night or splitting the measurement over several days.
With best regards
Andreas