I doubt that you have actually tested the response with sustained lateral acceleration. With magneticcorrection, you have two axes of alignment with which to correct the orientation drift. Even so,you will get a consistent dipping effect if there is sustained lateral acceleration, because the orientationplane formed by the magnetic field vector and the apparent "down" vector will become tilted.Since most cars and aircraft do not perform tight continuous turning manoeuvres, this may bea non-obvious problem.Anyhow, to get back to the OP's problem. There seems to be clearly something wrong with theOP's device, or the way that the information is being processed. If the offset and scale of the different axes is out of spec, this can be corrected for provided youknow what the behavior is, and provided the device is still linear. At the moment, this is beingobscured by the processing using code which we can't see.The next thing that I would suggest the OP do, is try to obtain the actual A/D values which are beingobtained from the device. Then, imagine the device is a cube, and turn the cube so that each of thefaces of the cube is facing upwards, and write down what the readings are. In other words, takethe readings with the flat side of the module up, and down. And the the short end pointing up, and thendown. And the the long edge of the module pointing up, and then down. It will then become moreclear what the problem is.The other thing I would do, is try actually observing the voltages with a multimeter.