The wine scanner had been doing something I’d rationalized for months: pick one of the two parsers per image, trust it completely, and live with the fields it reads badly. Measuring both against the corpus made it obvious they’re good at different things. So instead of choosing, I merged the two parses and took the best of each. That’s the good half of the day.
The other half was undoing myself. A while back I’d let the crowd order results until A Wine App had enough of her palate to do it herself — reasonable idea, and wrong. Run against the real catalog, forty-one thousand wines, same six personas and the same fixture every time, precision at ten got worse. So I reverted it. Not a graceful retreat, just a straight revert and a note about why, which is the version of that I’d rather write.
Elsewhere it was quiet housekeeping: a nightly export refresh on Havnetryk that churned about as much as it added, the usual overnight content drop into The Cloudy Brain, plus a small reworking pass on the ideas repo where one idea got pruned.
Tonight the scanner reads better than it did yesterday and the ranking is back where it started, which counts as progress in both directions.