Frame orientation as a variable

Yes, orientation matters a little: portrait and landscape frames fill a tool's fixed input shape differently and can trigger different automatic crops, changing how much of the image is subject. It is a small variable, and a protocol exists precisely to catch the variables small enough to dismiss on their own.

What orientation changes

Most rating tools resize an upload to fit a fixed input shape before anything else happens, and a portrait photo and a landscape photo of the same scene do not fill that shape the same way. A tall frame around a subject leaves less empty space around it than a wide frame of the same distance and angle, which changes the fraction of the image the subject occupies - the fill ratio - and fill ratio is one of the things distance itself already controls. Swap orientation without changing anything else and you have changed how much of the frame is subject versus background, which is a second variable riding along with the first.

Some tools also crop or pad automatically to reach their expected input shape, silently cutting the edges off a landscape frame or letterboxing a portrait one. Which one your phone defaults to, and which crop the tool applies as a result, is not something most people check before uploading. A crop or pad that shifts the subject within the frame is a translation, and a misread orientation flag is a rotation - both transformations Engstrom and colleagues (2019) found state-of-the-art image classifiers "vulnerable to."

Why it stays minor

This is not in the same league as distance, angle or light. Orientation does not foreshorten anything and does not change what the lens itself captures - it changes how much of that capture survives into the frame the tool actually reads, which is a smaller and more mechanical effect than the geometry aipenis.com covers on the modelling side. It is worth fixing for the same reason a small effect is worth fixing in any protocol: it costs nothing to hold constant, and an unfixed small variable is still a variable sitting inside a comparison you are trying to trust. A measurement taken with a tape does not have this problem at all, because there is no frame to fill or crop to apply - it is a variable specific to photographs, not to the subject.

A tool that shows a breakdown, the way Rate Cock does on public entries, at least gives a chance of spotting that one axis moved after a format switch when nothing else did; a total-only number gives no such signal.

What to do

Pick one orientation and use it for every submission in a series, the same way you would fix a distance or a crop. The four variables worth deliberately fixing do not list orientation on their own, but the same logic that applies to those applies here at a smaller scale, and it costs one line in a log next to the file to record which one you used. If you switch orientation partway through a series, treat it the same as any other single change - one variable per comparison, never two, and a format switch counts as one. A human reviewer would barely register a portrait-versus-landscape swap; a fixed input shape has no such flexibility, which is exactly why it is worth writing down rather than trusting to feel.