OK, what you were asking for before is not the test you have been doing, as nowhere in your procedure you were asking to subtract a master.
What I gather from your experiment is that:
First line is the stats of a single bias
Second line is the stats of same bias with a constant value subtracted
Last line is the stats of same bias – a masterbias.
Fair enough, it makes “more sense”. At least if the discussion from the start had been about calibrating biases… which is not at all the point.
So here is the test that we have been talking about. Made 200 darkflats of 0.9s, matching the flats done for a session from last summer. ZWO ASI294MC, gain 120, offset 30, T -10C for reference.
Stacked 1, 10, 25, 50, 100, 150 and 200. Yes, of course, the std decrease with increasing subs…
Then checked the stats of a flat (1st CFA channel only, it would not make sense on a color sensor to measure all channels altogether) from which I have substracted a synthetic offset of 1920 and masterdarkflats made with 50, 100, 200 stacked frames respectively.
And here-below the results:
I’m not even going to claim that there is significant reduction in noise in the calibrated flat, because the stds are probably within the uncertainties. But to be honest, this kind of measure is the basis for questioning the usefulness of using darkflats (or bias) to calibrate the flats.
I hope this will encourage others to test with their setup.
So now I think we've done enough testing (here and with the post). We're not forcing anyone to adopt synthetic bias, but we've shown in different ways that it's worth a try.