The ranking does not survive the meeting
Nothing records which candidate each person preferred, so the next run starts the argument from zero.
Preferences by Wisent
Pick the one you like more
Every choice ranks the items and the generators behind them with Bradley-Terry, writes down the taste that made the call, and exports text choices as ster contrastive pairs.
- Bradley-Terry rankings
- No account to vote
- ster pair export
- Private team arenas
Four candidates, one room, and no record of who preferred what
A generator hands back a run of logos, taglines or banners, and the decision between them belongs to whoever happens to be looking.
One reviewer's taste stands in for everyone's
The candidate that ships is the one the loudest reviewer liked, not the one people would have picked in front of a pair.
The taste itself is never written down
The reasons behind a choice stay in someone's head, in a form no model and no colleague can reuse.
One pair at a time, and the ranking follows the votes
A visitor sees two items, picks one, and that single choice moves three things: the ranking of the items, the ranking of the generators behind them, and the voter's own profile.
Serve a pair
Two items from one category appear side by side, sides randomised and generators hidden until the choice is made. No account is needed: the first pair creates an anonymous voter held in one cookie.
Record the choice
Left, right, tie or skip, with the milliseconds it took. Skips and sub-400 ms answers never reach the fit, streaks of twelve identical sides are down-weighted, and more than sixty choices a minute are dropped.
Refit the ranking
Every category is ranked with Bradley-Terry, over the items and over the generators that produced them, with a 95% interval per subject, rebuilt inside the transaction that records the choice.
Export the taste
Choices in a text category export as ster contrastive pairs, the chosen side positive and the rejected side negative, ready for ster train to turn into a steering vector.
The same product from a terminal, with exit codes a script can trust
preferences doctor tells a deploy gate whether the deployment is serving and ready, and preferences rank refuses to name a winner until the vote has earned one.
$ PREFERENCES_URL=http://localhost:3000 preferences doctor
url http://localhost:3000
operator token missing
account token missing
healthz ok
readyz ready
$ preferences rank --submission 01M1G1S8ZCA1E9S6VQ7W15ADA4
# CANDIDATE SCORE COMPARISONS
1 codespy 1.505 6
2 lem 1.499 6
3 jeden 0.346 7
4 brama 0.304 6
5 echo -1.184 5
6 las -2.471 6
18 qualified comparisons, 30 required
submission 01M1G1S8ZCA1E9S6VQ7W15ADA4 has 18 qualified comparisons; 30 are required before a winner is declared
$ echo $?
2What people ask before the first pair
Can a bot or a bored clicker swing a ranking?
Choices that are not evidence never reach the fit: skips and sub-400 ms decisions are dropped, streaks of twelve identical sides are down-weighted, and more than sixty choices a minute are dropped. A subject with fewer than 30 comparisons is marked provisional.
Who writes the sentence that describes my taste?
Nobody but the arithmetic. Profile summaries are deterministic sentences built from the recorded leans; no model writes them in this release.
Is my team's unreleased work visible to the public?
Not in a private arena. A private arena is bound to a Wisent organization and is invisible to everyone outside it, items and rankings included.
What does it cost?
Voting in public arenas and every public ranking are free. Private arenas for a team and the personal model path are the paid surface.
Start with one pair, or bring a whole run
Vote at app.preferences.wisent.com without an account, or hand a generator run to preferences submit, let people vote, and let preferences rank tell the generator which candidate won.