How voting works
Anyone signed in can vote. What changes is how much a vote counts, and the rule is public: a vote weighs more the harder the person is to fake and the more often they have been right.
The vote
Two buttons on every picture: helps me remember (+1) and doesn't (−1). One vote per person per picture. You can retract it or flip it any time. Scores and voter counts are public on the word page.
Learners vote without knowing it, too. When a word has more than one picture in the app, the one a learner picks on the phone counts as a +1 for that picture. A phone paired to a community account counts as that person; an anonymous phone counts for half.
What one vote weighs
| Who is voting | Identity |
|---|---|
| A handle with no email (free to mint) | 0.5 |
| A handle with a confirmed email | 1.0 |
| A paying member (Plus or Studio) | 1.5 |
| A pick from a phone paired to an account | 1.0 |
| A pick from an anonymous phone | 0.5 |
Reputation starts at 1.00 and multiplies that. Someone who is usually right climbs toward 2 or 3; someone usually wrong sinks below 1; someone who never votes stays at 1.00. The cap of 3.0 means nobody, however trusted, is ever worth seven newcomers. How reputation moves.
The lever is weight, not access. Nobody is told they can't vote. A throwaway account's vote simply counts for half of a reachable one, and a fraction of a proven one.
How the leader is decided
A word page doesn't just add up votes. It asks how sure the community is, using a standard statistical bound (the Wilson lower bound of the weighted approval). Three votes for and none against is not more convincing than eighty for and twenty against, and the maths agrees. Every tally you see is computed this way, never by a plain sum.
When a word settles
Not on a calendar. A word settles when its leading picture has all three:
- At least 5 different people behind it, voters and phone pickers combined. People, not weight.
- At least 60% sure, by the bound above.
- At least 15 points clear of the runner-up.
Then the word waits in the settled queue. While the safety scanner is off for a language, a person gives it one click, a safety step rather than an editorial one. With the scanner on, a settled word whose winner scanned clean is ratified by the machine, and the click disappears.
From the queue to the phone
Ratified winners are published as part of the word pack. Phones check for a new pack when the app opens, show a now-or-later prompt, and apply it on a tap. Nothing reaches a phone by any other path; one exporter, one publisher, and every picture in the pack has passed both the scan and the vote.
Rings and pile-ons
"You vote mine, I'll vote yours" is detected: only mutual upvotes in the last 60 days count, and a flagged pair has its reputation capped at 0.25 until the pattern stops. Nothing is auto-punished beyond the cap; a moderator can clear it, and it clears itself once the mutual voting stops. A picture buried by a pile of downvotes is held, and if it's still at −5 or worse from five or more people after 48 hours, the system retires it as community rejected and the downvoters get paid for being right.
What is written down
Everything a moderator does, and everything the system does on its own, lands in one audit trail with a reason. Automated actions have no actor, so a person can find them and overturn them. A vote's shape never changes after it's cast; when a rule changes, a number changes in one place, not a row in the record.