When the score is 'how well does this match a sentence

Not every scoring model was trained on photos labelled with numbers. Some were never shown a score at all during training. Instead they learned to compare an image against a sentence, and a score gets built from that comparison afterward, using a prompt instead of a labelled example.

What contrastive pretraining actually learns

A contrastive image-text model is trained on a large number of image-caption pairs, and the training objective is simple to state: push the internal representation of an image close to the representation of its matching caption, and push it away from the representations of captions that do not match. Do this across enough pairs and the model ends up with a shared space where images and pieces of text that describe similar things land near each other, regardless of whether either one was seen during training in that exact combination.

Nothing in this process ever taught the model to output a number. It taught the model to measure similarity between an image and arbitrary text, which turns out to be a surprisingly flexible building block.

Turning similarity into a score, with no scoring data at all

Because the model can measure how well any image matches any sentence, you can build a scorer without a single labelled score by writing sentences that describe the ends of a scale - something like "an excellent, well-composed photograph" against "a poor, badly composed photograph" - and measuring which one a given image sits closer to. The gap between those two similarities, or where the image falls on a line between them, becomes the score. This is genuinely elegant: no scoring labels were collected, no fine-tuning happened, and the whole system can be repointed at a new task just by writing different sentences.

The failure mode this creates

The elegance has a cost, and it is a specific one: the score now depends on the exact wording of the prompt sentences as much as it depends on the image. "Excellent photograph" and "outstanding photograph" are close in meaning to a person and can sit at meaningfully different points in the model's text space, because the model learned statistical associations from captions rather than a human sense of synonymy. Two tools built on the same underlying contrastive model, differing only in the prompt sentences their designers happened to write, can produce noticeably different scores for the same photo, and neither is more wrong than the other - both are answering the question their specific sentence actually asked.

This also means the scale itself is only as stable as the prompt. A designer who tweaks the wording to fix one problem can shift the scores of everything else without meaning to, since the two ends of the comparison moved and every image's position relative to them moved with it. It is a different fragility from the one that comes from training on comparisons between images, since here the axis being measured against is a sentence rather than another photograph, and the sentence is a single point of failure the designer controls directly.

Where the line sits with a full vision-language model

A system that has a language model actually write out a description of the image before producing a number is a further step beyond this, with its own separate mechanics and its own failure modes, and is worth its own treatment rather than folding in here. What is specific to a contrastive scorer is that the comparison is a single similarity measurement, not a generated explanation - there is no reasoning step to inspect, only a number describing how close two points are in a shared space.

What this means for reading a result

A score built this way is trustworthy to the extent that the prompt sentences were carefully chosen to describe what the tool claims to measure, and largely untrustworthy the moment the wording drifts from that. Rate Cock uses a trained scoring head rather than a bare prompt-matching setup for this reason, which keeps the scale tied to labelled examples rather than to a sentence someone wrote once and may not have revisited. How training-set composition sets a model's baseline sense of normal, separate from anything prompt-related, is covered in more depth here, and what the training objective itself does to a score's shape is worth reading alongside how the loss function shapes the numbers you get. Text-to-image matching is a general-purpose tool well beyond scoring, and Measure My Cock sits firmly outside it, relying on a tape and a repeatable method rather than any comparison to a sentence. A human reader interpreting a photo brings context no prompt can encode, which is part of what Rate Penis offers as an alternative. Reading a number for what it can and cannot support, prompt-based or not, is covered practically at penisrater.com.