Slop can be measured and the adjective still cannot be automated

Two design talks published from the same conference on September 10, and the first speaker apologizes in advance to whoever is on next. Paul Bakaus, who maintains a design skill called Impeccable, describes what they call Claude beige: instrument serif, italics, a competent page on which nobody decided anything. The fix is a vocabulary. Bolder, quieter, distill, harden, each word wired to a specific meaning inside the project, and the line for the whole approach is that an adjective with nothing behind it is just a nicer prompt. Bakaus gets asked about once a week to add an auto mode and says there will never be one. Taste, in this view, can be amplified but not lab grown.

Thais Castello Branco runs a lab. Taste Labs analyzed two million websites across about ten years, generated a synthetic set to compare against, and trained small classifiers, probes, for individual features. Combinations of those probes predicted slop better than asking a model to judge. The three markers are repetition, lack of fit, and low intent, and the word reached for at the end is not taste but judgment.

I do not think these two are as far apart as the stage made them look. The vault’s design notes sit in the same middle. Supply a concrete reference, make a compact plan before code, then screenshot and critique against the reference, with a separate verifier rather than the builder grading itself. The page also records a method for checking a design spec: withhold its reject-on-sight list, generate against the real brief, and see whether unguided output falls into those defaults. That is a slop probe by another name, and it shows the rules describe an external failure mode rather than one person’s preference.

So the measurable half is real and worth gating on. Repetition can be counted. Fit can be scored against a brief.

The adjective is what is left over. Somebody has to say bolder and mean something by it, and that is a decision the classifier was never asked to make.