Different parchment color through Herbal section?

[Koen G]

Stolfi: about the two sides of the parchment, this is another thing Claude came up with spontaneously. Apparently it is confident that it can tell one side from the other, no problem. I'm actually willing to give it the benefit of the doubt here, since detecting such differences is precisely what one may use an AI for. So assuming it tells flesh side from hair side properly, it claims that the effect is more outspoken when we look at the flesh sides only. The charts I posted do not reflect this, because I chose r+v or full bifolios. 

Exposure and uneven handling between folios is definitely a factor. F2 is much darker than f7, although they're the same bifolio. But this does not explain why B-folios are still identified in the mixed quires, which is the surprising part. You can use page brightness fairly well to separate currier A from B leaves in a mixed-quire context, which is not normal (readers don't decide to thumb B-pages less than A-pages, on average).

Uneven handling of different parts of the same page does not affect the results. I got the AI to compare the lightest samples, as you suggested, and this intensifies the split. The same is true for comparing different panels from the same page: the difference holds. In other words, contamination of part of a page can be ruled out as a driving factor. In fact, it may even look like the "cleaner" a section of parchment is, the more obvious the difference becomes.

I think you make a good point about camera's automatic processing. This can be checked to some extent against the blackness of the background. I asked Claude to do this, and indeed there are differences, but not enough to affect the results. I'll just paste the explanation here:

Herbal B after correcting for the background: the Herbal B pages do sit mostly at the higher background values (average 2.08 against 1.87 for Herbal A). But within Quires 4–7, correcting for background level, Herbal B is still about 1.2 L* lighter (p ≈ 0.001), and the background adds nothing significant itself. This rules out the most damaging version of the camera concern, a per-photo exposure boost. It can’t rule out a finer colour or tone correction, which is why a calibrated target in each frame, or the raw multispectral data, would still help.

Here is a chart sorted by brightness of "clean" samples (still hundreds per page), only flesh sides. It pulls Herbal B sheets out of their current bindings.
Interestingly, two of the exceptions are from Quire 8, which is weird and "transitional" in many ways. That orange bar in the middle section is not truly a B-bifolio. The plant drawings are strongly A-style, and the parchment is dark like regular A-sheets. But text-wise, it sits in a weird place between Currier dialects and writing style.