MPU-6050 thermal chamber results. Unexplained bias behaviour

I am attempting to characterise the MPU-6050 as a first step towards robust calibration. Part of this is achieved by logging the gyroscope zero-rate output over -20˚C to +70˚C (temperature change < 1˚C per minute). So far, I have only tested 2 devices of a larger batch but these results alone reveal an unexpected and extremely concerning behaviour.

I have posted plots with this identical post in a separate forum: http://forum.sparkfun.com/viewtopic.php?f=14&t=31902
It seems that the MPU-6050 incorporates a discontinuous process that constrains the gyroscope output within a 1 LSB range over a limited temperature range. This is seen in the attached plots for the x-axis of device A between -20˚C and 30˚C and for the x axis of device B between -20˚C and 0˚C.

The effect is that the bias temperature relationship that should be governed by an polynomial becomes discontinuous with what seems like an overriding heuristic asserted at some temperatures.

It just so happens that the x-axis of device B is approximately linear and this reveals that while the output is constrained within a 1 LSB range the gradient of the plot between ‘overflows’ is equal to the unconstrained output. See 'zoomed in' plot. This is what leads me to suggest that there is a heuristic process in place.

This discontinues behaviour would make robust calibration impossible. Please can someone explain what is going on and tell me how to 'turn off' whatever process is having this effect.