When quality and score part ways

Yes, a better photo can score lower. Photo quality and rating score usually move together, which is why the exceptions surprise people, but they are two different axes measuring two different things. A clearer image shows more of what the rubric reads, and that can cut either way.

One mechanism, worked through

A soft, slightly underexposed photo hides detail. A sharp, well-lit photo shows it. If the rubric an axis is checking rewards a feature that only reads clearly in good light and sharp focus, a better photo raises that axis's score by revealing something the blurrier version was hiding.

But if that same detail is not flattering - texture, an asymmetry, a proportion the softer photo blurred into ambiguity - the improved photo does not hide it anymore either. The blur was doing quiet work in the rougher photo, smoothing over exactly the thing a sharper image now shows plainly, and the axis reading that detail can move down for the better photo rather than up.

This is not the tool malfunctioning. It is the tool doing what it always does - reading what is actually in the image - and a clearer image simply contains more of what was previously hard to see, for better and for worse. Models are also more sensitive to capture changes than intuition suggests: Azulay and Weiss (JMLR, 2019) showed that "small translations or rescalings of the input image can drastically change the network's prediction" in convolutional networks.

Why this is not really a contradiction

Photo quality axes and subject axes, where a tool separates them at all, measure different things by design: one is closer to "how well was this captured," the other closer to "what does the capture show." A photo can improve on the first and reveal something that costs it on the second, and both movements can be entirely accurate readings of what each axis is built to check. The confusion only happens when a single blended total obscures which of the two things actually moved.

The fix is the same one that applies to reading any result: look at the breakdown rather than the total where one is available, because a total alone cannot tell you whether a drop came from the subject reading differently or from the photo simply working better at its actual job. This is also a reason not to chase visible improvement as a goal in itself - what "better photo" and "better submission" actually mean are two different targets, and only one of them is reliably about the score going the direction you expect.

The same split exists in a cruder form outside this category. A tape measurement has no equivalent confusion, because a method applied correctly reads what is there regardless of how the room is lit - there is no "quality" axis for the instrument to improve on independently of the reading itself. A human reviewer can move in either direction too, and for a related reason - a clearer photo can shift a person's read of tone as easily as it shifts a model's reading of an axis, sometimes toward more favourable and sometimes away from it. What a model is actually responding to at the pixel level, and why sharpness changes what it can see at all, is worth a look at the mechanics behind these tools generally rather than assuming "better capture" and "better outcome" were ever the same claim.

A lower score after a technically better photo is not a bug report. It is the axis working, on a Rate Cock breakdown or any other, exactly as it was built to.