Ranking twenty things at once is hard. Choosing between two is easy. That is the whole idea behind Quick Pick, and until today you had to deploy a Next.js app and a Postgres database to use it.
Now there is a page: rank a list by picking pairs. Paste up to 60 options (features to build, job offers, holiday spots, grant applications), answer "this or that?", and get an order with a range for each place: "rank 2, plausibly 1 to 4". Send the list to your team as a link, everyone answers on their own phone, and pasting their result codes back gives one combined order, an agreement score, and the options people split on. Nothing is uploaded: the list rides in the part of the link after the #, which browsers never send to a server.
Same model as the app
The page does not reimplement anything. A small build script strips the types from the app's own btRefit.ts (a Bradley-Terry refit with a Laplace posterior) and favouriteSet.ts and serves them as browser modules, and a test fails if the copy goes stale. Each question is the pair whose answer is expected to tell the model the most. Sides are a coin flip, so screen position can't creep in.
The stop rule that never stopped
The obvious way to stop is "every neighbouring pair in the order is at least 90% likely to be the right way round". I built that first, and it never finished. Even a simulated person who never makes a mistake, asked each neighbouring pair 4 times, left those pairs at 87 to 92%: the model allows for lapses, so a few consistent answers never make anything certain. With ordinary noise it ran to the question cap every time, at every threshold from 75% to 85%.
The rule that ships asks a different question: if you drew orders the answers still allow, how closely would they match the one on screen? It stops when that average Spearman correlation reaches 0.9. For simulated people with ordinary noise, the order shown at that point matched their real order at about 0.94. A scripted consistent respondent now finishes 8 options in 25 picks instead of 48. It is a little optimistic for very noisy answerers on short lists (0.77 read against 0.66 true), and the measurement notes say so.
"Fewer questions" was wrong
I also offered a "just find the favourite (fewer questions)" mode. Measured, it asks more: 49 picks against 37 for 10 options. But it named the true favourite 12 times out of 12, where the full-order stop managed 8 (and 6 of 12 at 20 options, where full order often leaves the top two close). So it is now called "be sure of the favourite", which is what it does.
Try it
marisstella.app/quick-pick-survey/rank.html is free, MIT licensed, and keeps your list in your browser. If you run surveys with many respondents and want demographic splits and an admin view, the full app is a one-click deploy from the project page.
A ranking from a limited number of picks is a statistical estimate, not a fact; don't use it as the only basis for a hiring, financial, medical or relationship decision.