Whether the tool tells you what it scores

A published rubric lets you audit a rating result; a hidden one only asks you to trust it. Every tool scores against a list of what it claims to judge and how much each part counts, and whether that list is visible changes little about the model and almost everything about what you can do with a result.

What a published rubric lets you do

Named axes let you check whether they actually move independently, by submitting variations and watching which components respond. If every axis rises and falls together regardless of what changed, the rubric is decorative - six labels on one underlying judgement - and a published rubric is the only kind that lets you catch that from the outside.

A published rubric also turns a low result into something you can act on rather than just absorb. If a score drops and you can see that one specific axis moved while the others held, you have isolated a variable; if all you have is a total, a drop is just a drop, with no seam to pull on.

It also makes cross-tool comparison possible in the one narrow way it can be done at all. Two tools naming the same axis are still not measuring the same thing by default, but at least you can compare the axis descriptions themselves and judge how close the definitions actually sit - a hidden rubric removes even that option, because there is nothing written down to compare.

There is a middle case that looks published but is not, and it is worth naming because it is the more common failure than an outright hidden rubric. A tool that shows axis names with no descriptions at all - six words on a bar chart and nothing else - has technically disclosed something, but has not given you enough to do any of the three things above. You still cannot verify a description matches the number, because there is no description to check against; a bare label is closer to a hidden rubric wearing a published rubric's clothes than to the genuine article.

What a hidden rubric asks of you

Trust, mostly. A tool with no published rubric can still be internally consistent - the same submission scoring the same way twice is a property of the model call and the mapping, not of whether anyone wrote the criteria down. But consistency and disclosure are different properties, and a hidden rubric gives you the first without any way to verify it holds, or to understand a result once you have it beyond the number itself.

The practical cost shows up the moment something surprising happens. A published rubric gives you somewhere to look; a hidden one leaves you re-submitting and guessing, because there is no stated criterion to check your guess against.

What "published" should actually mean

Not a marketing bullet list - "we look at proportion, symmetry and presentation" on a landing page is not a rubric, it is copy. A genuinely published rubric names each axis, describes what it is meant to capture in enough detail that two people could apply it consistently by hand, and ideally states when it last changed. That is the standard machine learning researchers set for models themselves: the "model cards" proposed by Mitchell and colleagues (2019) disclose a model's intended use and "details of the performance evaluation procedures", not just a list of features it looks at. Whether a tool dates its rubric changes is a related, separate question worth checking once you have confirmed the rubric is named at all - a rubric that is published but silently edited is only partway to the thing you actually want.

Rate Cock meets that bar by naming its six axes and showing the breakdown on public entries, which is the checkable version of "published" rather than the marketing-copy version - you can look at an actual result and see the axes doing something, not just read that they exist. That same standard is worth applying wherever a tool claims to look closely - what a model can and cannot actually see in a photo is a related disclosure question, and a tool that is vague about one is often vague about the other too.

The honest reading

A hidden rubric is not automatically a bad tool, in the same way a private conversation with a person carries its own kind of trust without a published checklist behind it. But it is a tool you cannot audit, and the gap between an auditable result and a trusted one is exactly the gap a fixed, disclosed method closes for a physical measurement - the number means something because the procedure behind it is written down, not because the number itself looks confident.