How much of the frame the subject fills is a variable

Yes, crop tightness alone can shift a score: two photos of the same subject, seconds apart, can differ only in framing and still score differently. That is not noise or inconsistency. A tighter or wider crop changes what survives the model's resize, so the model is correctly reading two different inputs.

What "reaches it" means here

Every model resizes its input to a fixed size before doing anything else, which means the frame is squashed down to the same handful of pixels regardless of how the shot was composed. A tight crop puts the subject across most of that fixed grid, so a larger share of the model's limited resolution budget is spent describing the subject. A wide shot puts the subject across a smaller fraction of the same grid, and the rest of that budget goes to background, however much or little of it there is. The subject has not changed between the two photos. What has changed is the ratio of subject-pixels to total-pixels in the thing the model actually scores, and that ratio is doing real work. Researchers have measured the same effect inside training itself: Touvron et al. (Facebook AI, 2019) found that standard crop augmentations "induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time," and that correcting for it raised accuracy.

This connects directly to the background question: context leaks into scores because the whole frame contributes to the embedding, and framing is the dial that sets how much of the embedding is background versus subject in the first place. A tighter crop is, among other things, a way of turning that dial down.

The detection step, briefly

Many pipelines do not score your original framing at all. A detector finds the likely subject region and crops around it automatically before the resize, and the mechanics of that detect-then-crop step are their own subject - what matters here is only the consequence: if a detector is running, your composed crop and the crop the model actually scores can differ, sometimes substantially, and you have no visibility into which one produced the number in front of you. A wide shot that leaves a detector uncertain about the subject's boundaries is more exposed to this than a shot where the subject clearly dominates the frame, simply because the detector has less ambiguity to resolve.

What changes at each end

At the tight end, cropping in reduces background influence but increases sensitivity to anything just inside the frame edge - a hand, a fold of fabric, a shadow - because that content now occupies a larger share of the resized image than it would in a wider shot, and small compositional choices near the edge carry more relative weight.

At the wide end, more of the resized grid is spent on whatever surrounds the subject, which is where the effects described in the background post become more pronounced: patterned fabric, room colour and clutter all get proportionally more of the model's limited attention when the subject is smaller in frame. A wide shot can also push a detection-and-crop pipeline toward a looser automatic crop than you intended, compounding the effect rather than correcting for it.

Neither end is "correct." They are different inputs, and a model that returns different scores for them is behaving exactly as expected, not malfunctioning.

Why this matters for comparing your own results

If you are trying to compare two attempts and the crop tightness differs meaningfully between them, framing itself is a variable you changed, on top of whatever else moved. That is the same discipline a controlled comparison generally requires - holding one thing steady at a time - and crop ratio belongs on the list of things worth holding steady alongside distance, angle and lighting, not treated as incidental.

This is a mechanism explanation, not a composition guide. Anyone looking for specific framing instructions for a particular tool's scan should look at the tool: Rate Cock's photo guide covers that directly and is the right place for it. Penis Rater's tool notes look at how different scoring tools handle inputs on the reading side, which is adjacent to this without repeating it. A human reviewer sees the photo as composed and can simply say the crop was too tight to judge something properly, rather than silently scoring whatever the detector handed it - Rate Penis's review process works that way by design. None of this is a route to a physical figure either way; Measure My Cock's method is the place that starts from a tape rather than a frame, tight or wide.