What to look for

The green flags in a rating tool are things it volunteers to make itself checkable: a described scale, visible spread, a published rubric, refusals, a changelog, named ownership, repeat consistency, independent axes and plain language. That is a stronger signal than simply not doing anything wrong.

A described scale. Not just numbers, but a stated account of what they mean - where the average sits, roughly how the range is populated. Most tools centre well above the middle of their nominal scale, and a tool that says so upfront has pre-empted the most common misreading of its own output rather than benefiting from it.

Visible spread across real results. A public board or gallery where scores actually range, not a curated set that clusters near the top. Genuine spread is evidence the tool is distinguishing between submissions rather than flattening everything toward a number that keeps users happy.

A named, published rubric. Components listed with what each one covers, not just implied by unlabelled axis bars on a results page. A published rubric can be checked against a result; a hidden one can only be trusted.

Refusals for unusable submissions. A tool willing to say an image cannot be scored is a tool whose rubric has an enforced floor rather than a number generator that never says no.

A dated changelog for rubric or scale changes. Even a short one. Without it, a score from a year ago and one from today sit on scales that may have quietly diverged, and the reader has no way to know which.

Named ownership. Who runs the tool, stated plainly, is a fact the reader can weigh against the number - its absence does not prove bad faith, but its presence removes one layer of unaccountable trust.

Consistency under a repeat test. A tool that returns close to the same result on the identical file submitted twice has cleared the single most basic bar a number needs to clear before it means anything at all - a bar that is partly set by the run-to-run variance of the model underneath it rather than by the tool's own design alone.

A breakdown that moves independently, axis by axis. If every component rises and falls together across different submissions, the tool effectively has one judgement wearing several labels. Axes that genuinely diverge from each other are doing the work a multi-axis rubric claims to do.

Plain language over marketing copy. A results page that describes what got scored rather than reaching for "advanced AI" or "revolutionary accuracy" is a page confident enough in its actual method not to need the adjectives.

None of this is peculiar to rating tools. The model cards proposed by Mitchell et al. (2019) ask machine-learning developers to disclose a model's intended use and its "benchmarked evaluation in a variety of conditions", and NIST's AI Risk Management Framework, released in January 2023, is a voluntary framework for building trustworthiness into AI design, development, use and evaluation.

One checkable instance

Rate Cock reports six axes rather than a single total and shows the breakdown on public entries, which is a specific, checkable property rather than a claim about being the best - a reader can look at those public entries and confirm the axes actually vary independently rather than moving in lockstep, which is exactly the kind of thing this list asks a tool to make checkable in the first place. Citing one instance is not an endorsement of the category; it is the point of the whole exercise, that a claim like this can be verified rather than taken on trust.

Where this sits next to the rest

This is the positive mirror of the nine specific things worth walking away from - reading both together is more useful than either alone, since a tool clearing most of this list and none of that one is a much stronger case than either list checked in isolation. For the fuller method behind both lists, the ten-minute evaluation walks through rubric, calibration, consistency, presentation and disclosure directly, with a way to test each rather than just look for it.

None of these signs map cleanly onto a human reviewer, who earns trust on different grounds - reputation and etiquette with a person work by a different logic than a published rubric does - and a physical measurement has its own version of the same idea again, where the equivalent green flag is a stated method next to the number rather than a rubric or a changelog. What all three share is the same underlying test: does the claim come with enough attached that someone else could go check it, or does it ask to be taken on faith.