A common claim, tested

Usually not in any way you can verify: "corrects for angle" is a claim, and a two-photo test settles it better than trust does. It would mean a shot taken from a slight downward tilt scores the same as a level one - a hard computer-vision problem, and an easy sentence to write in marketing copy.

What the claim would actually mean

Correcting for angle means the tool detects the geometry of the shot and adjusts its internal reading before scoring, so that a foreshortened image is treated as equivalent to the level one it is a distortion of. That is a nontrivial piece of computer vision - estimating camera pose from a single image and undoing perspective effects is a real, hard problem, not a footnote. How a model actually processes an image is worth understanding in general terms before deciding whether a specific correction claim is plausible for a given tool; this site is not the place to adjudicate the internals, only the outcome.

Some tools may do a version of this. Many that use the phrase loosely mean something much weaker: a note in the results copy, a disclaimer, or nothing measurable at all. The phrase "corrects for angle" is doing a lot of work in a sentence that costs nothing to write.

How to test it

Take two photos of the same subject under otherwise identical conditions - same distance, same light, same crop - varying only the camera angle: one level, one tilted down by a noticeable amount. Submit both to the tool. If the correction is real, the two scores should land close together, within whatever noise floor a same-file resubmission produces. If the tilted shot scores meaningfully lower, the correction either does not exist or does not reach that magnitude of tilt.

Run it more than once. A single comparison is not evidence of anything by itself - a two-photo test carries the tool's ordinary noise plus whatever true angle effect exists, and one pair cannot separate them. Three or four angle pairs, and a look at whether the gap is consistent in direction, is a much stronger basis for a conclusion than one.

Why the default assumption should be no correction

Foreshortening from a downward tilt is a real geometric distortion of the photographed subject, not an artefact a tool can always distinguish from an actual difference in what it is scoring. Image models are also known to be fragile to geometric changes far smaller than a tilt: Azulay and Weiss (2019, Journal of Machine Learning Research) showed that "small translations or rescalings of the input image can drastically change the network's prediction," and that data augmentation alone does not guarantee invariance. Absent a specific, verifiable claim and evidence that the tool holds up under the test above, the safer working assumption for your own protocol is that angle is not corrected and has to be controlled at the point of capture - which is exactly the standardisation approach covered in making a submission comparable.

This is a different question from how a human reviewer handles an off-angle photo, where a person can mentally adjust for perspective in a way a model may or may not replicate, and where the correction, if it happens, happens in someone's head rather than in code. It is also different from a measurement context, where angle is controlled at the point of capture precisely because no software correction is trusted to substitute for a level camera in the first place.

What a real answer would look like

A tool making an honest correction claim has an easy way to demonstrate it: publish the test result, or at least describe what the correction covers - a specific range of tilt, a specific method, a specific limitation. A rating service that shows its own working, rather than asserting a feature and leaving the reader to trust it, is doing the reader a favour that costs the tool nothing except the discipline of writing it down. Absence of that kind of detail is not proof the claim is false, but it removes the one thing that would make trusting it on faith reasonable.

If a tool passes the test, relax the protocol on that one axis and keep the rest fixed. If it does not, angle stays on the list of variables you control yourself, same as distance and light. Either way, the two-photo test costs almost nothing and replaces a marketing claim with an actual answer, which is a better trade than most protocol decisions get to make.