This is and interesting observation and thought experiment. And I'd agree that controlling a light source's intensity while monitoring the camera's response is a logical approach to look for a defined correlation. However, I think "the fly in the ointment" is that under normal spectrometer measurements, there is very little light in the spectral band within the camera's image field -- especially when there is no clipping of R/G/B channels. I've been guessing that any "AGC effect", when the image field is very near to black, would place the AGC control loop at one extreme with full gain. This suggests that the only intensity correlation that would be available as a correction factor is at the end of the gain range which is normally very non-linear. (Medium light and mid-range for the camera image as a whole, i.e. just low-light conditions' might be much more linear.) The data suggests some randomness to the low frequency "drift" I observed which is what leads me to guess that the drift is just the near-DC noise component of the AGC effect. If that were correct, then there'd be no way to develop a correction from the signal. I'd considered the use of background variation as a substitute, but all I've observed so far is just noise as that "signal" level is essentially zero so noise is the only thing left. However, averaging the noise from the dark field, taken at the same time as the signal data, could be subtracted from the signal as a background offset. This is easily done and I've found benefits in reducing the low-level residual noise of the spectral data.