I don't know what we can do. You should be able to run a sweep now, later, tomorrow, the next day... and get the same, correct result over and over.
There won't be drastic differences if something significant hasn't changed. So that tells me in each of the very different graphs something has changed.
We have no way of knowing what that was.
You need to start from scratch with everything, reset to zero. Document every measurement in the notes area with signal chain and measurement rig and settings.
I would want to measure something that has a known response at this point. Like a driver close-mic. Get to the point where you can repeat the measurement setup and reading process to get the same result every time, know exactly how the final response is composed (calibration and correction) and have it consistently measure a known response so you have a baseline.
When your response changes you should be able to say, I know exactly why that is so different. I changed this, and it's exactly as I would expect.