Scores
Getting from six numbers to one
Equal weights, fixed weights, and non-linear combinations all produce a total out of ten; each one makes a different submission look best.
Six axis scores do not become a total by themselves. Something has to combine them, and the combination rule is a choice with consequences, made once by whoever built the tool and almost never shown to whoever reads the result.
Three ways to get from six numbers to one
Plain average. Every axis counts equally: add the six, divide by six. It is the simplest rule, the easiest to reverse-engineer from outside, and the one that implicitly says the tool has no opinion about which axis matters more than another.
Weighted average. Each axis gets a fixed multiplier before the sum - proportion might count for twice what colour does. The tool has an opinion here, it just usually declines to publish it. A weighted average produces a total that looks like a plain average and is not one, and there is no way to tell the two apart from a single result.
Non-linear, minimum-dominated. The total tracks the lowest axis more than it tracks the mean - one weak component drags the total down further than an equal-weighted rule would. This rule is rarer, harder to spot, and it changes which submission "wins" in a way neither of the linear rules would.
The same six numbers, three totals
Proportion 8.0, symmetry 7.5, presentation 6.0, texture 7.0, colour 8.5, composition 7.0.
| Rule | How it combines | Resulting total |
|---|---|---|
| Plain average | Equal share to each axis | 7.3 |
| Weighted average | Proportion and symmetry double-weighted | 7.6 |
| Minimum-dominated | Weighted toward the lowest axis (presentation, 6.0) | 6.7 |
Same submission, same six numbers, and the printed total moves by nearly a full point depending on which rule sits underneath the display. A submission with one visibly weak axis looks noticeably worse under the third rule than under the first, and a submission with no weak axis at all barely changes between any of the three - which is the tell: minimum-dominated rules mostly punish unevenness, not low averages. Swap which axis is lowest and the ranking of two submissions can flip between rules even though neither submission's six numbers changed at all - the total moved because the arithmetic underneath it moved, not because anything about either submission did.
What each presentation implies
A tool that shows six axes and a total inviting you to check the arithmetic is usually running something close to a plain average, because a weighted or non-linear rule would visibly fail that check on the first few results anyone tried. A tool that shows a total with no breakdown at all could be running any of the three, and there is no way to know which from the result page alone - checking whether the numbers a tool does show actually reconstruct the total is the only way in. What a total keeps and what it discards compared to a breakdown is a related but separate question from which rule produced it - a tool can show a full breakdown and still hide a non-linear rule behind it, if nobody checks the arithmetic.
Why tools rarely disclose the rule
A published weighting is a public statement about which physical or aesthetic property the tool's owner considers more important, and that statement invites disagreement in a way a silent number does not. It is also a lever the tool can retune without telling anyone - nudge one weight, and every future total shifts in a chosen direction without a single axis definition changing. Neither incentive points toward disclosure, so most tools leave the combination rule unstated and let the total arrive as if it simply is what it is.
Where this leaves you as a reader
If you can see the axes, you can approximate the rule yourself by comparing the average of the axes shown to the printed total across a handful of results - a consistent gap in one direction across several submissions is a weighting, not noise. Rate Cock reports all six axes alongside its total on public entries, which is exactly what makes that check possible; a tool that only ever shows the final figure gives you nothing to compare it against. None of this touches whether the axes themselves measure something real in the first place - what a vision model is actually doing to produce a per-axis number sits underneath the combination question, not on top of it - and it has no analogue at all for a person: a human reviewer does not run six numbers through a formula, they just decide what to say, which is a different process with a different kind of opacity. Measurement has the opposite property to all of this: a length in centimetres is not a combination of anything, so there is no weighting question to ask of it in the first place - the whole idea of a combination rule is specific to tools that score more than one thing at once.