How it works
Every number in this game comes from somewhere you can check.
This page explains where the cards come from, how Attention is measured, how a rank is computed, what protects the standings, and what doesn't exist yet.
Sources and attribution
All content comes from English Wikipedia, and only from it. Card text is under the Creative Commons Attribution-ShareAlike licence, and every card links to its source article. Pictures come from Wikimedia Commons under free licences only, with author and licence on the back.
Wikilore is not affiliated with, endorsed by or approved by the Wikimedia Foundation. Wiki is a generic word, as in wikiHow. The Wikipedia and Wikimedia names, logos and puzzle globe appear nowhere in the game, only in the attribution you're reading now.
Wikilore reads Wikipedia and never writes to it. There is no reward, ever, for editing an article, voting, donating or doing anything at all on Wikipedia, and no feature that would encourage it.
Where the pool comes from
It doesn't come from a random draw, which would land mostly in Wikipedia's long tail of stubs. The pool is built from public editorial lists, since being on such a list already means a subject is notable:
| Source | What it is | View floor |
|---|---|---|
| Popular pages | The monthly reports of Wikipedia's Community Tech bot, one per WikiProject: the 500 to 1,000 most-read articles of each subject area. 134 projects are read, and which ones is an explicit editorial choice | 20,000 |
| Vital articles | Wikipedia's own list of its most important subjects, levels 3 to 5, by theme. It brings the concepts, events, works and animals that popularity lists under-weight | 10,000 |
| Monthly top 1,000 | The 1,000 most-read articles of each month, over 24 months. The 1,000th of a month sits at 230,000 views | 20,000 |
| Memes | The "Internet memes" category tree, used as the editorial list of memes because there is no popularity report for them | 10,000 |
| Hand-checked lists | The titles chosen by hand for the Concept and Place types. The list says what to measure; the floor decides what gets in | 10,000 |
| Curiosities | Wikipedia's own list of unusual articles, read line by line by us: no recent tragedy with real victims, racism judged case by case. It is the only way into the Curiosity type, whose rarity is measured among its own cards | 1,000 |
A title found in several sources takes the lowest of their floors, so a card needs importance or strong popularity. The floor also has to hold over time: the twelve-month median and two separate 30-day windows must all clear it, so a spike in the news isn't enough.
Each candidate goes through three filters. It must be a real subject (no disambiguation pages, lists or season pages), it must have what a card needs (short description, Start class or better, a 150-character opening), and it must clear its floor.
Pages about explicit sexual content are not cards: sites, genres, practices, pornographic works. People are never excluded for their profession. A well-known adult-film actor is a card like any other well-known person, and removing them would go against the notability rule.
No domain is capped. Only the view floor decides.
How Attention is measured
Attention is the number Wikimedia publishes: pageviews for an article, per day, "user" traffic only. The interface removes the traffic it identifies as automated. Wikilore reads and stores that number and never counts anything itself.
- 60 days are fetched per subject, which gives the current 30-day window and the 30 before it. That's where the change shown on the card comes from.
- Twelve monthly values are stored as well, which is where the yearly median comes from.
- A day's figure appears the day after. Cards and ranks use the last published day. The Trend Call reads the hourly files Wikimedia publishes during the day, hour by hour, as they come out.
- Nothing is fetched during a game. The pool is built offline and served from a file. A player opening a booster never triggers a call to Wikimedia.
The card shows the rank views, rounded ("1.2M", "48k"), plus the 30-day change as a percentage. It never shows the exact number.
How a rank is computed
- Rank views = the median of three measurements: the current 30-day window, the previous one, and the median of the last twelve months. A single month in the news can't make a card Legendary.
- Cuts by percentile, inside each type: Common below the 55th, Rare to the 85th, Epic to the 97th, Legendary above. The two types with fewer than 500 cards use 50 / 80 / 95.
- Validated by hand. A script proposes the cuts from the real distribution and writes them to a file with the date and a validated: false flag. Nothing uses them until a human sets it to true, type by type.
- 20% hysteresis. A card rises only at the higher cut plus 20%, and falls only below its own cut minus 20%. Around 17% of the pool sits within 20% of a cut and would otherwise flicker on noise.
- Refreshed weekly, every subject, on a new window. Once a month the sources are read again and new subjects are added and ranked.
- After a change of cuts, the movement is reset. A "▲" never reflects a change of rule.
In most types the step between neighbouring ranks is about ×2 (from ×1.8 to ×2.7), and between ×1.5 and ×1.8 in the smallest types. That margin keeps a rank from moving on statistical noise.
The bot check
Some articles carry artificial traffic that Wikimedia still counts as human, and it shows in the device split: real readerships are mostly mobile, robot pages are almost all desktop or almost all mobile web with no app readers.
Every heavily read subject gets its desktop / mobile-web / mobile-app split measured and is flagged when that split looks like a robot's. A flagged subject stays in the pool but never appears in a Trend Call pair, a League list or an Up or Down list, and a flagged card is capped at Epic until a human reviews it.
Pictures and licences
A free licence covers copyright, but it doesn't cover everything, and the card says so.
A lead image is kept only if it's on Commons under public domain, CC0, CC BY or CC BY-SA (never NC or ND), is a still image, and has a readable author when the licence needs one. It never goes on a Fiction card, except for a hand-written list of 33, almost all figures of ancient myth.
About nine cards in ten have a usable photograph. The others, Fiction by rule or articles with no image at all, show the illustration of their type.
Some files carry Commons' "Restrictions" warning (trademark, personality rights). That's outside copyright: the subject of a photograph may hold rights the licence doesn't cover. A licence filter can't settle that, so a reported picture is hidden at once and reviewed by a person within 24 hours, and the game is built to run entirely without photographs.
Cropping already counts as a modification. A cropped CC BY-SA photograph is still CC BY-SA, and the card says so. Every card has a report button: one click hides the picture, the card falls back to its illustration, and a human decides within 24 hours.
What Wikidata adds
Each subject's Wikidata entity is read once, after the pool is built. It adds birth and death dates, nationality, occupation, coordinates, and the number of Wikipedias that have the article, which is the best single measure of worldwide fame. Deaths are detected from it, so a death can neutralise a pair or a pick.
Anti-cheat
The two families of modes face different threats.
Real-time modes
Face Down has nothing to look up: both players see the same cards and the same trophies. What a script can do is read the other player. The strongest possible bot wins about 55% of games against a person, but a script that reads a predictable player wins far more. That's why the duel carries INK and prestige, and its weekly standing gives boosters, never anything outside the game. What protects the standings:
- A server clock, never a client one: thirty seconds a round.
- Moves stay secret. Both cards are played face down at the same moment, and the server holds them until both are in. The duel is only ever played live: a replay has no opponent reacting, and two accomplices could trade their moves in it.
- Statistical detection, computed weekly before any standing reward is given: pairs of accounts that meet often, wins without a contest (the opponent left, or had most of their cards played by the clock), and losses that keep going to the same two accounts. The thresholds aren't published, but the principle is.
- Leaving costs. Whoever leaves a duel loses it, and it counts in their five of the day; the one who stays gets everything the win gives. Leaving three duels in a day puts a wait before the next one.
- Sign-in with Google, with no minimum account age to play. An X account is linked later, from the game. Every week counts in the standings, X or not; a reward booster needs an X account linked before the rewarded period starts, and one X account serves one Wikilore account only.
- Rewards held on a flag. A flagged account's standing rewards are held until a human checks it, and a second flag excludes the account from standings and their rewards.
- A daily cap per account on counted duels: five a day, so screen time doesn't decide the weekly standing.
- Victory boosters only from server-matched duels: never between two accounts on the same device or address, never for the same two players again within 24 hours or more than three times a week, and a fixed ceiling of ten a day, inside the shared limit of 11 boosters a day.
Prediction modes
The answer doesn't exist yet, so the threats are multi-accounting and view inflation.
- Every pick is judged against a published expected figure, the same for everyone and frozen when picks close. Picking whatever is simply big or simply rising doesn't work, because the expected already knows it.
- Reward boosters need an X account linked before the rewarded period starts, so linking X after a good result gives nothing, and one X account serves one Wikilore account only, plus cluster detection on identical picks, timings and devices.
- Leagues and Up or Down: rolling four-week standings, at least three weeks played, and the same reward for the whole top tenth, which a random account practically never reaches.
- Trend Call: a bar that chance doesn't clear. A week's Codex takes 25 correct out of 35, which chance reaches about one week in 120.
- A view floor on every listed subject (inflating a subject at 2,000 views a day takes tens of thousands of requests), plus a desktop / mobile check on that subject's week, measured and shown to the team. Nothing is excluded automatically.
- Deaths neutralised, with the formulas and data published, so there's nothing to argue about afterwards.
Boosters are protected by design rather than by statistics: a booster has no monetary value, its tier is drawn on the server from a secret and can never be re-rolled, the free booster is three in reserve at most (ten with PRO) and shares a daily limit of 11 with victory boosters, and Trend Call takes five picks a day.
Accounts
Sign-in with Google, with no password ever stored. Sign-up asks three things: an age declaration, a nickname (pre-filled from your Google name), and a declared country. No email address is stored. An X account is linked later, from the game, when it's needed: to trade, to message a player who isn't your friend, and to receive reward boosters.
You can look at the game without an account: the cards, the standings and a guided first duel against The Archivist are open to everyone. Signing in is only needed to own something.
Check it yourself
You don't have to take any of this on trust. Everything the game uses is public:
- The views of any article are published by Wikimedia through its pageviews interface, day by day. The numbers on a card are that data, read through the rules above.
- The dates, languages and coordinates on a card come from Wikidata, which is public.
- The licence and author of a picture are on the file's Commons page, linked from the back of the card.
- The Open League list is published with its week number and each subject's expected growth, everyone's picks are revealed when picks close, and the final figures are published with the calls that produced them.
- Trend Call results are published daily at 12:00 UTC, with the figures and each subject's expected views, for the day whose picks closed at 12:00 UTC the day before.
- The formulas behind "expected" are frozen, dated and published. How expected is computed →
- The drop rates are published and are the ones the server applies. See them →
If a calculation turns out to be wrong, it is corrected and the correction is published within seven days.
What does not exist yet
We list these here because this site never describes something unbuilt as if it existed.
- PRO, the optional subscription. It is presented, not open.
- Guild chat and private messages, which are switched off until someone reads the report queue every day.
- Timeline and Rank Them, two short modes that are planned.
- Browser notifications. In-game badges and an inbox come first.
- A mobile app. The site is built for phones first; an app comes later, if the game works.