A rubric written in English, read by a model

Some scoring systems have a rubric baked into weights: a head trained on thousands of labelled examples that has no separate description of what it is looking for. Others carry the rubric as text, handed to a language model at the moment of scoring, and that difference is bigger than it looks. When the rubric is a prompt, changing the words changes the scorer, with no training run in between.

Two ways to build the same rubric

A trained scoring head learns its criteria implicitly, from examples. Ask it what "symmetry" means to it and there is no sentence to point to - only weights that respond to certain patterns because those patterns co-occurred with high labels during training.

A prompt-based system does the opposite. Someone writes "score the symmetry of the subject from 1 to 10, where 10 is perfectly even and 1 is markedly uneven" and that sentence, verbatim, is the entire definition the model has to work with at inference time. Nothing was learned about symmetry specifically. The model is applying general language and vision understanding to a definition it is reading for the first time in that request. Some prompt-based rubrics go further and embed worked examples directly in the instruction rather than only a description, a technique covered in few-shot examples in a scoring prompt.

Both approaches produce a number. Only one of them has a rubric you could read.

Why wording moves the score

Because the instruction is natural language, it inherits every property natural language has: ambiguity, connotation, and sensitivity to emphasis. "Rate the proportion" and "rate how well-proportioned this looks" are not the same instruction, even though a person would treat them as interchangeable, because the model is responding to the literal request rather than to an underlying stable concept it already holds.

Ordering matters too. A rubric that lists six axes in one order can produce different individual scores than the same six axes listed in a different order, because each instruction is processed in the context of everything that came before it in the prompt. This is not a defect specific to any one tool. It is a property of asking a system to define its own yardstick fresh each time, rather than giving it one baked in during training.

The practical result is that two tools which describe the same rubric in their marketing copy can behave differently, because "the same rubric" understates how much the exact phrasing carries. It also means a single tool can drift if an engineer rewrites the prompt to fix an unrelated bug, and the score distribution shifts as a side effect nobody intended.

What stays constant and what does not

The image itself, and its encoding into a vector, are the same regardless of which kind of scorer reads it. What changes is the layer that turns that vector, plus a rubric, into a number. A trained head has one fixed rubric, for better and worse - stable, but expensive to change and opaque to inspect. A prompt-based system has a rubric anyone with access to the prompt can rewrite in a minute, for better and worse - inspectable, but only as stable as the text is.

This is one reason multi-axis breakdowns matter more for prompt-based tools than for trained ones: when the rubric is a sentence per axis, seeing which axis moved after a prompt change tells you exactly which sentence is responsible, in a way a single blended number never could. It also connects to why the number itself can vary run to run even with an unchanged prompt - a separate cause, covered in how sampling randomness enters a score, and worth keeping distinct from wording sensitivity, since one is about what you asked and the other is about how the model answers.

Where this shows up for a user

Nobody sees the prompt. What a person sees is that the same photo, submitted to the same tool on two different days, returned two visibly different numbers, and the explanation is sometimes as mundane as a backend update that reworded one line of an instruction the user never had access to. Rate Cock documents version changes for exactly this reason - a rubric change is a real change to what "score" means, not a bug, and a tool that tracks it is doing something most competitors do not bother with.

The comparison to a fixed method is worth naming rather than assuming. A written rubric read by a language model is still an inference from pixels; if you want a figure with an actual measurement behind it, that is a different discipline entirely, with its own tools and its own error bars, and Measure My Cock is the property that covers it properly. Penis Rater's tools coverage looks at the practical side of choosing between rubric-based and trained scorers, if you are deciding which kind of tool to use rather than how either works internally. And a human reading the same rubric brings something a prompt cannot: a judge who can ask a follow-up question rather than apply a fixed sentence once and move on.

None of this makes prompt-based scoring unreliable in some absolute sense. It makes it a different kind of system than a trained head, with a different failure mode - not drift in the weights, but drift in the words - and knowing which kind you are looking at is most of what you need to interpret a surprising result.