The angle variable, isolated

Tilt the camera down and the image compresses along the line of sight, and a rating tool has no way to know that compression came from the angle rather than the subject. It scores what the frame shows, and a downward tilt changes what the frame shows without anyone touching the subject.

What a downward tilt actually does

Point a lens down at something and the parts nearer the lens get pulled toward the camera relative to the parts further away, along the same line the lens is pointing. The effect scales with how steep the tilt is, not with whether you meant to tilt at all - a phone held at chest height and glanced down at is already off level, usually more than it feels. The result reads to a viewer, human or model, as a shorter or more compressed subject than a level shot of the same thing would show.

Why "a little tilt" is not a controllable amount

The trouble with angle specifically is that it fails silently. Distance is easy to notice - you can see how close you stood. Angle is not: a five-degree tilt and a fifteen-degree tilt look nearly identical through a phone screen, and neither looks obviously off from level unless you are checking for it. That means angle cannot be reproduced by feel the way some of the other variables can be approximated by feel. It has to be set against something external - a level line, a mounted phone, a mirror held square - or it drifts a little differently every time, and a score that moved for that reason will look exactly like a score that moved for a real one. Image models are not immune to geometry either: Engstrom and colleagues (2019) found state-of-the-art classifiers "vulnerable to other, more natural classes of perturbations such as translations and rotations," and training on varied examples gave only "relatively small robustness."

Why level is the only repeatable choice

Level is not special because it flatters anything; it has no bias toward a higher or lower reading. It is special because it is the one angle you can define precisely enough to hit again. "Level with the subject" is a fixed target - a horizon line, a spirit-level app, the top edge of a door frame in the background. "A bit tilted down, like last time" is not a target at all, and a tool cannot separate an angle drift from an actual change in what it is scoring. That is the whole argument for treating angle the same way distance gets treated: pick a setting that can be specified, not one that merely felt right at the time.

Where this sits next to the other variables

Angle is one line in a longer list. The variables that move a score, ranked by how much each one matters, places angle alongside distance near the top, because both change proportion rather than anything cosmetic, and proportion is what several axes in a typical rubric are actually reading. Making a submission comparable at all starts from the same list - fix what is not the subject, angle included - before a second result can be read against a first one.

What this does not cover

This is about the score a model produces, not the geometry term for what a tilt does to an image - that gets its own definition elsewhere, and a rating tool has no interest in the word, only the effect. It is also not a claim about accuracy in the sense of a physical measurement: an actual length has its own method, independent of any camera angle, and that method is not what is being discussed here. What the underlying model is actually doing with a tilted frame goes deeper into the mechanics than a protocol note needs to.

The protocol answer

Set the angle against something fixed, not against how the last photo felt. A phone propped level, or a mirror checked square, beats "held about the same" every time, and this is as true when the recipient is a person rather than a model - a judge reading a tilted photo is looking at the same compressed proportions a tool would score. It costs one extra check before pressing the shutter and it removes an entire silent source of drift from every comparison made afterward. Rate Cock's photo guide walks through setting the angle for its own tool; the level-line habit it describes is the same one worth keeping regardless of where the file ends up.