No lens data reaches the score

A rating model sees a grid of pixels and nothing else. It does not receive the focal length of the lens, the physical size of the sensor, or the distance between the camera and the subject, even though a camera's own file can record some of that information alongside the image.

The pixels are all that travels downstream

Most scoring pipelines decode the image, resize it and hand the resulting tensor to the model. The camera and lens metadata, where it exists, sits in the file's EXIF block, which is a separate part of the file from the pixel data and is not part of the model's input. Even a pipeline that reads EXIF for other reasons - logging, orientation correction - has no reason to feed a focal-length number into a scoring head that was never trained to use one, and none does. The general handling of that metadata, what gets kept and what gets stripped, is a separate subject from what actually reaches the model.

Why this matters for the score

Focal length and distance together determine perspective distortion: a wide lens held close exaggerates whatever is nearest the camera, while a longer lens from further away compresses depth and looks closer to how the eye sees the scene. That distortion is a real, scored variable - the model reacts to the shape it produces in the image. What it cannot do is correct for it, because correcting for a distortion requires knowing the lens that caused it, and that number never arrives. The model reads the outcome of an unknown camera setup as if it were simply the subject's shape.

Sensor size works the same way from another angle. Two phones with different sensors and lenses can produce visibly different renderings of an identical scene from an identical position, and the model has no way to attribute the difference to hardware rather than to the subject. It scores what came out the other end of an entirely opaque optical chain.

What would need to change

A model could, in principle, be given camera parameters as an additional input alongside the image and trained to adjust for them, the way some computational photography systems already do internally for their own corrections. No general rating model does this today, because it would require a training set labelled with reliable camera data at a scale that does not exist for this task, and because most photos have no way to record accurate distance-to-subject regardless. A manual method sidesteps the problem differently, by fixing the distance and angle as part of the procedure rather than trying to recover them afterward - measuremycock.com's method is built that way, which is the trade-off a rating model does not get to make. Until that changes, treat any axis sensitive to shape or proportion as reflecting the photo's optical setup as much as the subject, which is one reason the same subject scores differently from one attempt to the next.

None of this is a case for hiding camera information rather than stripping it - what a file's metadata can reveal is a privacy question with its own answer, unrelated to whether the model can use it. Rate Cock scores the pixels it is given the same as any other tool in this position, and comparing how different services present that same blind spot is worth doing before trusting either one.