The correlation failure in multi-axis tools

A breakdown is only informative if its axes can disagree. If proportion, symmetry, texture, definition, presentation and grooming all rise on your good submissions and all fall on your bad ones, in step, every time, then the tool has not given you six pieces of information. It has given you one, printed six times with small variations that look like independence but aren't.

What correlation looks like from the outside

You do not need statistics for this, only a handful of results laid side by side. Pull five or six submissions and their breakdowns and scan down each axis column. If a submission that scores high does so across every axis, and one that scores low does so across every axis, with the relative ordering roughly the same column to column, the axes are moving together. A genuinely independent set of axes would show at least one submission where, say, proportion is the strongest axis and texture the weakest, or the reverse on a different submission - some crossing of the lines, not six parallel ones.

Why it happens

Sometimes this is the decorated-total case, where one number was always doing the work and the axes are cosmetic. But it can also happen with a genuinely computed rubric, if the underlying model has learned an overall "quality" signal that leaks into every axis it scores, so that a submission it judges favourably on one dimension gets a small favourable nudge on all the others too. The distinction matters for what you do about it, but the symptom - full correlation across a handful of results - looks the same either way, and the practical consequence is identical too.

What it means for reading the breakdown

If your axes move together, stop reading the breakdown as six separate claims. Read it as one overall judgement with six labels attached, and treat the total as the only number carrying real information - the axis with the lowest bar this time is not necessarily "your weakest area," it may just be wherever the single underlying judgement happened to land lowest under a formula you cannot see. This does not make the tool useless. It means the tool is functioning as a single-number rater wearing a multi-axis interface, and it should be evaluated - and trusted - on those terms rather than on the promise of attribution the breakdown implies.

What genuine independence would look like

The useful contrast is a tool where the axes actually diverge submission to submission: proportion strong while presentation is weak on one result, the reverse on another. That divergence is what lets you attribute a change in the total to a specific cause, which is the entire value proposition of having a breakdown at all rather than one number. Testing for it takes the same handful of results this whole check is built on - no need for anything more elaborate than eyes and five or six rows of a table.

Where this fits

This correlation check is a symptom-level test - it tells you the axes aren't independent, not why, which is the narrower diagnostic covered in telling a real rubric from a decorated one. It is also a separate question from how many axes a tool reports in the first place - a tool can have twelve axes and still fail this test, and a tool with three that pass it is more informative than the twelve that don't.

The same failure mode shows up outside this category too. How reliable the underlying model actually is per judgement is one of the reasons axes correlate in the first place - a model less confident overall tends to fall back on one general impression across everything it's asked to score. Devices matter on the measurement side for a related reason: inconsistent gear introduces its own correlated error across whatever it touches, which is a different mechanism producing a similar symptom. A written human response can fail the same test in its own register - a review that praises everything in the same breath is giving you one overall reaction dressed as several specific ones, and it is worth noticing there for the same reason it is worth noticing here. Some tools make this easy to check because they show enough results to look at in the first place; Rate Cock publishes its six-axis breakdown on public entries, which is what makes a check like this possible without submitting a dozen photos yourself.