That's an understatement. Haha The docs are basically worthless. There is so many different versions, all scattered around the web. Not to mention, most of them are in the form of a .pdf that is hundreds of pages, with a terrible search function. hah
With that said, Palette  is an awesome company that builds exactly what you are describing. I've used there products for a bit now for PS work, and love it!
Also, you generally don't want potentiometers for these applications, but rotary encoders. Pots = physical min and max position; encoders = infinite rotation. So for example, when you're switching from one photo to the next, pots would annoyingly keep the old value, while encoders will switch to the correct value. Just a little detail, but it's nice to have to right term to search for.
I don't know what it's like now, but I'm glad I don't use Photoshop anymore to worry about it
There are still some features that I want which I think would need an exposed API. For example, setting an anchor point on the canvas, then triggering a quick jump back to it.
How much of this predictive power has to do with facial recognition, and how much has to do with compression artifacts, pixel value abberations, etc?
Edit: the paper addresses this question:
> One question is whether the
network is using “high-level” cues, such as shape and symmetry, or simply relying on “low-level” cues, such as resampling artifacts.
> although the network may be taking advantage of certain low-level cues, such as resampling, some performance is retained even when those are averaged away
Seems that it's both... but I'm not sophisticated enough to read this paper deeply.
So an unsharp mask would be a no, unless it was applied to background or everything except faces (for example).
Things like white balance don't count, unless you are selective according to the image content; so you could white-balance according to the whole-image statistics, but _not_ "recognise the is a face and use white-balance to, eg, lighten or reduce contrast in that area".
Fooled me one, shame on you. Fool me twice, shame of me. Fooled me with three perfect tacos, dammit!
This might have helped to detect taco clones: