Almost sixty thousand lines landed on A Wine App yesterday, and nearly all of it was one question: can a crowd of strangers tell you what a bottle tastes like before you’ve tasted it? I’d tried this once already and reverted it — that version spent the crowd’s opinion on every tier at every confidence level and made all six of my test personas worse. The second attempt measures against people instead of categories. Held out half of a million-rater dataset, built the producer table on the other half, and checked whether the numbers predicted the half that never saw them. They did, at four different splits, consistently enough that I stopped looking for the flaw.
The honest result came at the end of the day: I’d been carrying two ideas about how to lean a recommendation toward a persona, and the measurement says one of them does nothing at all. Wrote that down rather than quietly dropping it. Elsewhere in the same project, scanning now decides a row is a wine because it recognises wine, not because we happen to sell it — the catalog question comes later, and the answer is usually no.
Plus a nightly data export on Havnetryk, a content pass on The Cloudy Brain, and an idea got pruned.
Tonight: 61% of the catalog carries a crowd signal, and one of my two leaning strategies is officially dead.