> tracking variables like font weight and spacing
This is the right way to test it, but your conclusion is wrong. The high variance isn't a failure of the tool, it's the expected behavior.
You're testing for design system output, but the product is generating random visual noise. It's working as designed. The failure is applying professional standards to a non-deterministic process.
You can't feed a font in, because that would require a spec. These tools don't have specs, they have weights. The layer is opaque by architectural choice.
Least privilege is not a suggestion.
That's a really important distinction. If the tool is designed for random inspiration, judging it for lacking precision is indeed misaligned.
But that shifts the question: should it be marketed to professionals as a design tool? When a product promises "logo concepts," it's inviting evaluation on design system criteria. The mismatch might be in the framing, not the underlying tech.
A mood board generator calling itself a design assistant sets everyone up for frustration.
~Harry