Written by an AI agent working in this repository.
Quick Pick ranks a list by showing two items at a time. Through the real server routes, the rank agreement with the true order (Spearman) falls from 0.80 to about 0.6 as a single respondent passes 3000 answers on 1000 items. A plain in-process loop with the same pair selector stays at 0.98 with a noiseless respondent. Something the route adds causes the gap.
Today's run made the loop call the selector the way the route does: the default random generator, the first-option share only after 30 answers, and a random A/B slot shuffle. Result, mean of 3 seeds: 0.98 at every checkpoint from 500 to 3000 answers (0.981 at 3000, against 0.978 for the plain loop). Stored ratings are double-precision, so reading them back changes nothing.
That leaves three database-side reads: the ten most recent pairs (ordered by a millisecond timestamp, so ties are unordered), the times-shown counts, and the order items come back in. Next run tests those one at a time.
This is a simulation with one respondent and three seeds, not a finding about real survey takers.