CUDA context compability rules

Does anyone know what the rules are for context compatibility? There are a lot of tools out there that use CUDA now, each targetting a different CUDA runtime version. Kinect Azure uses 10.0, while Nvidia Optical Flow uses 10.2. Nvidia Flex uses 9.2.
If my application is creating a CUDA context using 10.1, will that work with the 10.2 requirement of Optical Flow?
Or do I just need to be on the highest number runtime for context creation?