Under-represented bodies get thin-sample scores

A model does not treat every input with the same amount of evidence behind it. Some regions of the space it learned are dense, built from thousands of similar examples, and some are sparse, built from a handful. A score for a body type that was rare in training comes from the sparse end, and a score built on little evidence is closer to a guess than a read, however confident the number sounds.

What "rare in training" actually means

A training set for a scoring model is assembled from whatever images and labels were available, and availability is never evenly distributed across the full range of human bodies. Some combinations of build, proportion and presentation appear constantly in whatever sources fed the dataset; others appear occasionally; a few appear barely at all. This is not a claim about which specific types are common or rare - that varies by dataset and nobody outside the team that built a given tool actually knows the breakdown for it - it is a claim about the shape of the problem: any dataset built from whatever was available has an uneven distribution, because "whatever was available" is never a random, representative sample of anything.

Why sparse data produces less reliable scores, not just different ones

When a model is trained on many examples clustered around a particular type, it learns a stable, well-supported mapping from that kind of image to a score - the labels it saw agree with each other often enough that the pattern is real and learnable. When it is trained on very few examples of another type, the pattern it extracts from those few is more sensitive to whatever those specific examples happened to look like, including any noise or quirks in that small sample. Two models trained on the same overall data but with slightly different examples in the sparse regions can produce noticeably different scores for a body type they each saw rarely, while agreeing closely on the body type they both saw constantly. That instability between essentially similar models is itself evidence of thin support - a well-supported score does not swing that much based on which handful of examples happened to be in the training run.

This connects to, but is more specific than, the general point that a model can only be reliably calibrated on the kind of input it saw during training - that piece covers distribution shift generally, across any unusual property of an image. This one is about body type specifically as one axis along which that general problem shows up, and it is worth separating because body type is not something a photographer can fix with better lighting or framing the way other distribution-shift causes can be partially mitigated.

What this does not mean

It does not mean an unusual score for an under-represented body type is necessarily low, or necessarily unfair in a specific direction - thin data produces unpredictable results, not systematically harsh ones, and which direction a sparse region skews depends on exactly what was in the sparse sample, which nobody outside the model's builders can audit from the outside. It also does not mean a mainstream body type gets a "more correct" score in any meaningful sense - it gets a more stable one, which is a different property. Stability is not the same as being right about a fundamentally subjective judgement; it is only a claim about how much a given model's number can be trusted to repeat itself and to reflect a real pattern rather than sampling noise.

Why this is hard to detect from a single result

There is no visible flag on a score that says "this came from a thin region of training data." The interface returns a number with the same formatting and the same apparent confidence regardless of how much evidence supported it. The only practical signal available from outside is instability under repetition: if a score for the same photo moves more across repeats than seems typical, thin training support is one of several plausible explanations, alongside ordinary retest variance that affects every photo to some degree. Distinguishing "thin data" from "normal noise" from the outside is not reliably possible without access to the model's training composition, which most tools do not publish.

What is worth doing with this

Treat any single score, for any body type, as a sample from a distribution rather than a verdict, which is sound practice regardless of representation, and it is the same reasoning behind why a rubric broken into axes beats a single blended figure. For a body type more likely to sit in a sparse region of a given model's training, that practice matters more, because the distribution the single score is sampled from is wider and less trustworthy to begin with. Rate Cock publishing a chart of scores across axes rather than one number at least exposes some of this: if one axis moves a lot between repeats while others hold steady, that unevenness is more informative than a single blended figure would ever be. None of this is a measurement question - actual physical data, where it is relevant at all, belongs to Measure My Cock rather than to a scoring model's judgement, and the two should not be conflated. A human reviewer's judgement has its own version of this problem, shaped by that reviewer's individual exposure rather than a training set's, and how human judges are selected and calibrated is a genuinely different mechanism worth reading before assuming a person avoids the issue entirely. Reading a single number sceptically, especially an unusual one, is consistent with how Penis Rater recommends treating any AI result generally.