Frequently asked questions

Entering a tournament and putting a probability on each of its real-world events. You are not picking winners — you are saying how likely each outcome is, and the scoring marks how well calibrated those numbers turn out to be.

About four weeks, with events resolving across it. A tournament's own standings and THE CALL do not carry into the next one. What does carry is your track record and the standings — each tournament's Best predictor / Most winnings, and the daily, monthly and yearly boards — which rank every call that resolved in that window.

Per event, just before that event's resolution window opens — never all at once for the whole tournament. Until an event locks you can move its probability as often as you like, and only the last value counts. Different events lock on different days.

On Scroll, each story opens on a swipeable headline rail (photo, publisher, an AI summary of that article, and Read the story), then THE CALL ON THIS STORY — the question, the slider, and an Events row of every live tournament that wording sits on — then THE ARGUMENT: an Overall take, then the case for YES and the case for NO — then THE MODEL, a fixed-height AI forecast with Gemini, GPT, Claude, and DeepSeek pinning their Yes percents from that coverage (tap one to read its rationale), with a My Prompt tab if you want to prepend an instruction. You never leave the feed.

At /agents, several language models each put a probability on a story from the news articles — the same mini slider as a comment. Click through to read one agent's rationale; underlined names like RBA article open the piece. We do not pass them the Polymarket or Kalshi price. When the agents and the market disagree, we flag it — the news pack may be thin. They are ranked on mean accuracy (how close the first call was to the outcome), plus variance and correlation to the pinned market, so copying Polymarket does not look like independent skill. They are a benchmark, not players, and they do not enter tournaments. On Scroll they pin on THE MODEL rail and on the field histogram with their own icons (Claude, Gemini, GPT, DeepSeek) so you can see a zero-prompt Gemini percent next to GPT on the same story. That is not the robot mark used for operator test bots. We link other public benches (ForecastBench, Prophet Arena) as peers; we do not ingest their questions.

On Scroll and on your tournament slate, once you release the slider, a histogram animates in under the slider. Until a question has a hundred calls, those bars mix a shape drawn from a pinned public-market price with the real callers, labelled as a citation of that market — not a count of players. After that they are an anonymous histogram of everyone else, the one live peek at the field, and they do not name anyone. Your first release stays on the card as First; later moves show as Now. After you lock yours in, friends who have called that event pin as named avatars on the same bar. House models (Claude, Gemini, GPT, DeepSeek) pin there too, with their own icons, so a zero-prompt Gemini call can sit apart from GPT — they are not operator test bots and they are not mixed into the anonymous bars. A Friends toggle on Scroll walks you through every live question they have already called (their number still waits until you save). Everywhere else waits until the event locks: the field's average and the named distribution stay hidden, otherwise the first few numbers would anchor everybody else's. Once an event locks, the field's average appears on the tournament slate.

You can enter and forecast every event that has not locked yet. But if any event has already resolved before you join, you cannot be fairly ranked against people who answered it, so you appear in the standings as incomplete and unranked for that tournament. Fixing that properly — crediting missed events at the field's average score — is designed and not yet built. Until then, joining before the first resolution is the way to be ranked.

They submit at 50%, which scores 75 points either way — the fence. Nothing is skipped and nothing is scored as zero; you just get the flat outcome. The app warns you once before locking in a slate with untouched sliders.

Each event pays 100 × (1 − (p − o)²), where p is your probability and o is the outcome (1 for YES, 0 for NO). This is the Brier score, expressed as points. Called 100% and right: 100 points. Called 100% and wrong: 0. A hedged 50%: 75 points, whatever happens. Your tournament total is the sum across every event, with THE CALL counted twice.

Only when you are genuinely unsure. The scoring is proper, which is a mathematical property meaning your expected points are highest when you report exactly what you believe. Shading a 90% belief down to 70% to feel safe costs you points on average, and so does inflating a 60% belief to 90%. There is no clever strategy hiding behind the honest number.

Yes — equal weight, with the single exception of THE CALL, which counts double. Hard events are not worth more than easy ones, because the points already handle it: nobody gains much on a question the whole field gets right.

No, and that is deliberate. A contrarian bonus of the obvious kind turns out to change nothing: adding a term for how far you sat from the field is the same ranking as leaving it out, because everyone's crowd term is identical and it cancels. Being different from the field only pays when being different is also being right — which the plain score already rewards.

They are not. Two players on the same total share a rank, and the next rank down skips accordingly. Scores are carried to one decimal place in the standings.

A second ranking, sitting on top of the topical tournaments rather than replacing them. Standings sums the Brier points of every event you called that resolved that Sydney calendar day, month or year. Missing an event is just a skipped score — you are not marked incomplete. THE CALL doubles on Most winnings only when that event itself resolved in the window; Best predictor ignores THE CALL. There is no separate slate and nothing extra to enter: every call you already make counts. The Tournaments tab on the same page ranks each topical tournament the same two ways. Agents ranks house LLMs with no reader prompt. THE MODEL Default tab pins Gemini, GPT, Claude, and DeepSeek on the story; My Prompt stays locked to DeepSeek. The Agents board scores all four. Prompts ranks reader instructions plus the model and an optional creator name. Agents shows Accuracy, Brier, variance, and market correlation; Prompts shows Accuracy %, Brier (hover the ?), confidence from a p-value, and evaluation count. Rank on Agents is by accuracy; rank on Prompts is by Brier.

An instruction you prepend before DeepSeek forecasts a story — ignore an outlet, weigh a source. The Default tab is the house evaluation with no instruction, ranked on Agents. Change the instruction over time — each wording keeps its own accuracy, Brier, confidence, and evaluation count on Prompts, with the model in its own column. Your display name is the creator only if you tick that box on your account page; otherwise Anonymous. Accuracy is how often the Yes/No lean was right; Brier is the 0–100 score (same rule as Best predictor; hover the ?). Rank is by Brier. There is no model picker; it is locked to DeepSeek. You can edit the instruction on a story or on your account page.

Every tournament — and the daily, monthly and yearly boards — has two rankings. Best predictor is pure accuracy: Brier points with THE CALL stripped out. Most winnings is the same points with THE CALL counted twice, which is the ranking a prize pool would split. Winnings are shown as points; there is no cash on the board. Incomplete slates stay unranked on both.

You designate exactly one event as your CALL. That event scores double. From the tournament page it is a radio when you first submit a full slate; from Scroll it is the first event you save, which also enters you into that tournament. It is mandatory — a slate always has one — and it locks with that first save.

The one you are most confident about. Doubling an event doubles both the upside and the downside, so the expected value of a Call rises the further your probability sits from 50% — and it does not depend at all on what anyone else said. If you were told THE CALL is where you go contrarian, that is wrong, and we would rather say so than let you lose points to it.

Not once your slate is submitted. The one exception: if the event you Called is voided while it is still open, you re-designate it among the events that are still open. If it is voided after locking, re-picking is closed and your CALL bonus instead pays your own average score across the tournament, which neither rewards nor punishes you for the void.

The source named in the event before it opened, read against the threshold written in the same place. Every resolution is published in the tournament's resolution log with its timestamp, the source, and the reading. Full lifecycle on the rules page.

No. Leaderboard is not affiliated with, endorsed by, or a partner of Kalshi or Polymarket. Those names appear only as citations of a public market we may pin as a source. We do not operate an exchange, take or match orders, or send you there to trade. When we pin a public market, the card cites it as a source. We score our own game.

The event is voided, with a written reason, and scores for nobody. Voiding is the escape hatch and not the plan — the real defence is writing questions precise enough that ambiguity cannot arise. An outcome being surprising, unpopular, or bad for the standings is never grounds.

There is a 48-hour dispute window after each resolution; affected standings are marked provisional while a dispute is open. Write to theworldleaderboard@gmail.com with the event and what the source says. After the window closes the resolution is final, and later revisions by the source do not reopen it.

It can be extended once, never past the tournament's final resolution date. At that date, questions that can be read as date-bounded ("will X happen by 30 September?") resolve NO; anything that cannot be read that way is voided.

Create a group and Invite friends from the group page (or from Friends) — they land on the board without a code. You can still share the invite link for anyone not on Leaderboard yet. A group ranks its members against each other inside any tournament you have both entered — it is a filtered view of the same standings, not a separate contest with separate scoring. Pairwise friends are separate: add someone from Friends (or from a comment) and, after you both call an event, you see each other's probability — still hidden from the field until lock — on the question page and as avatars on the Scroll bar. From Friends you can also their stories lead your scroll, and open their profile to see likes, live calls, and stats. A like they press on Scroll shows as a named pin on that story for you. Their percent on a live call still waits until you have called the same event.

No. Groups have no stakes and there is no way to add any. Running a private cash pool through the site is specifically out of scope and always will be.

Anyone. Standings, your display name, your points, your track record and the ring around your photo (how many calls you have made, and how accurate they have been once they resolve) are public; that is the product. Your profile is at /me and anyone can open it from a name on standings. Optional job, area of expertise, and Substack / Linktree / website links you add on Account are public there too, as is how long you have spent reading stories grouped by topic (the full page is /reading). Your live probabilities stay hidden from the field until the event they belong to locks. If you comment after calling, that comment shows the number you had at that moment — a public mark, not your later slider moves. Accepted friends are the exception on the live number: once they have called the same event, they can see it — and they can see which stories you liked on the scroller and on your profile, and which you have called.

Your rank around the profile photo. Shape starts as a circle, then a spiked circle, a hexagon, a rounded triangle, a spiked hexagon — more calls, a bigger outline. Colour is average accuracy after events resolve — grey until three have settled. The ladders, and where you sit on them, are on the progression page. It is computed from predictions already on the account; nothing extra is stored.

Minutes you spend on Scroll stories and full-screen questions that have THE STORY or THE ARGUMENT to read, grouped by topic — China, politics, oil. A card with no coverage does not count. The clock runs while that text is on screen and the tab is in front; a fast swipe does not count. Education articles are not counted. Totals are public on your profile and on the reading page. It is a record of what you have been getting sharper on, not a score.

To browse, read events, and look at standings, no. To submit a slate, join a group, or comment, yes — a score has to belong to someone. Sign-in is by Google or by an emailed link; there is no password to choose or forget.

Yes, on your account page, at any time. It is the name that appears in standings and beside your comments, so it changes everywhere at once — including on results from tournaments already finished.

Yes. Background on your account page has two optional fields — job, and area of expertise — that appear on your profile for anyone to read. Leave a field blank to take it down.

Only ones about your own events: an event you forecast has resolved, been corrected, or been voided, and the occasional notice that a voided CALL needs re-picking. No marketing. You can switch them off on your account page and still get sign-in links, which are not optional because they are how you get in.

Yes, from your account page, immediately and without asking anyone. It erases your predictions, entries, group memberships and comment text, and removes you from every standing. It cannot be undone. You can download a copy of everything first from the same page. Details in the privacy policy.

Session cookies keep you signed in; those are strictly necessary. Optional product analytics (PostHog) lives in this browser and does not run until you accept it — the first-visit bar, or the switch on your account page. Nothing follows you onto other sites. Details in the privacy policy.

Yes. Every tournament is free to enter. There is nothing to buy, no paid tier, and no way to spend money that improves your position. That is true of Leaderboard as it stands today — it is an early product, and paid formats are intended later. If one arrives it gets its own terms and its own announcement, and it will not be applied to a tournament you already entered.

No. There is no stake, no wager, no odds, no payout tied to an outcome, and nothing to cash out. You are being scored on accuracy, not paid on a result.

Rewards after resolution are intended and their form is not settled yet, so the honest answer is that we have not decided. What we can say is that the scoring does not change for them, and that any tournament you have already entered stays on the terms you entered it under.