The blind test
Which model wrote which?
The four untouched apps, shuffled into slots A–D with nothing to identify them. Play with each one, assign a model to every slot, then reveal. If the thesis holds, you're mostly guessing on style — because on correctness they're indistinguishable.
Hints if you've read the study: one build has a warm editorial look, two went clean and clinical with green traces, and the biggest one has a purple gradient header. Knowing that is the only way to score 4/4 — which is the point.