In trying to demonstrate for Tom how much the camera height needs to change for figuring a margin of error on "eye level," I realize I may be able to do that.
In the meantime, here are 3 stills from the same shot, with the only change being camera height. The top image is camera lower than water level, producing a sighting in which the near tube level appears 1px higher than the far tube level. I can see the misalignment with my eye.
In the middle image, I eyeballed the water to be level, but measuring from the bottom of the meniscus in each tube, the level of the near tube is actually 1px lower than the far tube level. That means the camera is higher than the water level.
In the bottom image, the camera is definitely too high since I can clearly (IMO) see that the near tube water level is lower than the far tube's.
The tubes are about 14" apart and the guideline is equidistant between them.
The camera is set back about 30" from the near tube.
These are un-resized stills from a video clip, and I can't remember what resolution my camera is set to record video in, so I can't yet work out the geometry to see what angular dimension 1px of image is equivalent to. But working that out, I should be able to calculate how much the camera was moved in the vertical axis to generate the differences in the top and bottom images.
The top and bottom images are what I would consider the boundaries for margin of error, and my guess is the camera height was adjusted by about 1" between 1px too low (top image) to 1px too high (middle image). The third image is 5px too high, and yet the guide string still hasn't reached the apparent horizon, so even with this margin of error, from about 800' in elevation, it doesn't seem to matter in answering a yes/no question of whether or not the horizon is at eye level. Even giving "eye-level the benefit of doubt at the outer range of error margin, there's still a gap.
I hope to reproduce this and actually take measurements of the change in camera height rather than deduce it through calculation, but in the meantime, have a look at the video clip:
Knowing the dimensions of the water leveler/camera setup up, and gauging the amount of camera height adjustment needed to bring the guideline in contact with horizon, we can, in fact, calculate the "dip" from level horizon. What will be missing is a target at a known distance that would be on the tangent point of the assumed globe.
I didn't notice it at the time, but there is a container ship that becomes apparent on the horizon just north of the setting sun. It might be possible to use that if we assume not too much of it's hull is obscured by the horizon (or convergence zone if you prefer) and we can estimate it's size/class.