The verdict beat

A review verdict is a performed beat that resolves the clip's register, not a finding produced by analysis. Buyers remember the ending as the judge's conclusion, the way a report has one, and that is where the misreading starts.

A beat, not a finding

A verdict is a structural part of the clip, the same way a greeting is. The greeting sets the register up; the verdict resolves it. Both are performed choices about how to land the piece, not steps in an analytical process that produced a result. The ending also carries more weight in memory than its length suggests: Do, Rupert and Wolford (2008) found strong evidence that the peak-end rule - judging an experience largely by its most intense moment and its final one - applies to material goods as well as to pain.

This is why a judge working an honest register lands a plain, direct verdict, and a judge working worship lands an enthusiastic one, on the same material. The verdict follows the register the clip was performed in. It does not sit above the register as an objective summary of it.

Why it varies so much

Because it is rhetoric, not measurement, the same photo gets a different verdict from every judge who reviews it, and a different verdict again from the same judge in a different register. None of these are more "correct" than the others. Two reviews from two judges are two performances, not two data points to average or compare, and the verdict is the clearest place that shows.

A verdict with a number attached is a further layer of performance on top of this, not a different kind of statement - whether and how judges give numbers is its own question, and the short answer is that a number inside a verdict is closer to a flourish than a score.

What it is not a substitute for

A verdict is not the place to look for something an automated score is actually built to give: a consistent, comparable number. A numeric score exists to be compared against other numbers; a verdict exists to close a performance, and reading one as the other produces disappointment on both sides. It is also not a measurement, however it is phrased - a stated figure from an actual measurement is the only reliable way to get a number that means the same thing twice. And it is not generated the way an automated read is generated: a model produces a fixed output from an image with nothing rhetorical about it, where a verdict is written to land, for one person, in one register, once. Choosing which of those two kinds of ending you actually want is worth doing before you commission either, and Rate Cock's comparison of a human and an AI judge covers exactly that choice.