What the decimal place in a score is actually carrying

The decimal in a rating score usually means less than it looks: it is the same single judgement rounded one place later, not extra information. A result reads 7.4, not 7, and the extra digit looks like the tool found something a whole number missed. Mostly it did not.

Where the digit comes from

Almost nothing inside a rating tool outputs an integer. The model produces a continuous internal value - a float with far more precision than the display shows - and the interface rounds it to one decimal place because that is the convention the category settled on. The decimal is not a second measurement layered on top of the first. It is the same single judgement, truncated at a different point.

What that decimal can and cannot carry

A tape and a stated method can genuinely support a decimal - that is a measurement claim, and it belongs to a different property, not this one. A rating tool's decimal is closer to a percentile than a length. It says the internal value landed slightly above or below a whole number on this particular pass, which is a much smaller claim than it looks.

The test that exposes this is repetition. Submit the same file twice and the total will often move by a few tenths without anything in the image changing. If the noise between two identical submissions is larger than the digit you are reading, the digit is not carrying information about the submission - it is carrying information about the tool's own instability that day. Measurement science has a rule for this: the Guide to the Expression of Uncertainty in Measurement (JCGM 100:2008) says results should not be given "with an excessive number of digits" and should be rounded to be consistent with their uncertainties. Rounding compounds the same problem from the other direction: a value can sit on either side of a rounding threshold and produce two different displayed totals from two internal values that were almost identical.

What produced the number in the first place

None of this is a comment on whether the underlying estimate is any good. How precisely a vision model estimates something at all is a question about the layer beneath the display, and it is a different subject from what the display does with the answer once it has one. A human reviewer, for what it is worth, rarely hands back a decimal at all - a person says "solid seven" and stops, which is arguably more honest about the resolution actually available.

How to read the digit that is there

Treat it as a tie-breaker, not a claim. Between two submissions of your own, a 7.4 next to a 7.6 is weak evidence that one edged ahead - useful only alongside a few repeats, never on its own. Between two different tools, the decimal means nothing at all, since each tool's ten-point scale is built and calibrated on its own terms and there is no reason their tenths should line up.

The same caution applies at the far end of the display choice. Stretching the same judgement onto a longer scale does not change what produced it - a 74 out of 100 carries the same amount of information as a 7.4 out of ten, no more. On a site that shows its breakdown, like Rate Cock, the decimal at least has something to check itself against: the total's tenths should move roughly the way the six visible axes suggest, and if they do not, that mismatch is more interesting than the digit itself.