a pilot with a thousand of the most reviewed steam games asked whether game quality is predictable from store text. old-fashioned ridge regression on embeddings, five fold cross validation, held out predictions. the answer is a solid yes, the correlation between predicted and actual review positivity sits around 0.6 depending on the feature set. store text carries a real quality signal, not a rumor of one.
the interesting split: gameplay embeddings built only from tags, genres, and categories beat full concept embeddings built from the whole description. the store prose is fine, but the mechanics view predicts quality better. neighborhood features, how close a game sits to similar games in the space, push it to about 0.67. there's mild leakage in that last set, but the embeddings-only numbers are clean and still strong.
the blunt interpretation: this is a usable prior, not an oracle. it tells you a game reads like good ones, which is exactly the filter the most-different-game idea needs so novel doesn't just mean lonely. wallpaper engine, beat saber, holocure, inscryption, the novelty ranking at the other end came out coherent, real neighbor clusters, not junk. the pieces fit together: a quality gate and a diversity search are the same project now.