DeepStack is a forecasting game about the money behind artificial intelligence. Every question has a yes-or-no answer, a public document that settles it and a date. You put a number on the chance of yes. When the document is out, the answer is read, every number is scored by one rule, and the record is public. Nothing on the platform is money: there is nothing to buy, nothing to win in cash, and no number here is advice. The publication, The Buildout, is the depth under each question, not a requirement for playing.
These rules are the platform’s. They are revised only by the owner, and the date at the top is the date of the current version. What the platform keeps about you is on the privacy page; the words of the game are in the glossary.
A question
A question is a sentence with a yes-or-no answer, a rule that says how the answer is read, the public source that will be read, and a date. The rule is written before the question opens and does not change afterward. The source is a document anyone can open: a quarterly filing, a monthly release, a government series, a published auction result. The date is the day the question closes: a dated question closes at 00:00 UTC on that day, whatever the precision of its period, so nobody forecasts later than the house itself may.
A question resolves yes or no when the source is read against the rule. It resolves inconclusive when the source does not say, or says something the rule did not foresee; the result then states what was read, what was missing and what the house will do next. An inconclusive question counts for no one. A question can also be withdrawn before its date (below); a withdrawn question counts for no one either.
Many questions come in a ladder: the same figure against three thresholds, lowest first. A reading that clears a higher threshold clears every lower one, so your chance of yes can never rise up the ladder; the controls keep your numbers coherent, and each rung is scored on its own.
Every question belongs to a thesis, and every thesis has two argued sides. A yes strengthens the side that expected it and weakens the other; a no does the opposite; an inconclusive moves nothing. The record of a side is the record of an argument, never of the people quoted for it.
The three numbers and the blind mode
Each open question carries three numbers besides yours, on one scale from 0 (no) to 100 (yes). The readers’ number is the median of the latest number each signed-in reader gave the question; it shows once five readers have given one, with the count beside it, and in full color from 25. The Call is the house’s own number, recorded before the date with a reason that cites a figure, and revised only with a reason and no later than 72 hours before the date. The machines’ number is the median of the house’s forecasting models, which read the question’s page and nothing else.
The page is blind until you commit. Before you save a number, none of the three is on the page, in the address or in the code the browser receives. After you save, they are revealed one by one. The blind mode holds everywhere a number is shown: on the question, on the ladder, in the takes and in a duel. The record of your first number being blind is what the Blind call achievement rewards.
Your number and the Brier score
A number is a whole percent from 1 to 99: the chance, in your view, that the answer is yes. You can revise it as often as you like until the question closes; every number is kept with its time, and nothing is edited. The last number you gave before the close is the one the Brier score reads; all of them count for Points, each for the time it stood.
When a question resolves yes or no, every counting number gets a Brier score: the gap between the number, read as a fraction, and the answer, one for yes and zero for no, squared. Zero is perfect, 0.25 is what a coin flip earns every time, and a confident miss costs close to one. A 70 on a yes is 0.09; a 70 on a no is 0.49; a 95 is 0.0025 when right and 0.9025 when wrong. The Call and the machines are scored by the same rule on the same questions.
Your page shows your calibration: of the numbers you gave in each band of ten points, the share that came out yes. A reader whose 70s come true about seven times in ten is calibrated.
Points
Points reward being right, and being right early. On each resolved question, Points are 100 × (0.25 − Brier), integrated over the time your number stood. Each number you gave stands from the moment you gave it (or from the question’s opening, if you gave it before) to your next number or to the close. Its weight is those hours divided by the hours between the opening and the close. Points on the question are the sum, over your numbers, of weight × (0.25 − Brier) × 100, capped at 25 and floored at −75.
Time before your first number weighs nothing, so a right number on the first day is worth more than the same number on the last. A coin flip, 50, is worth zero whatever the answer. A question that resolves inconclusive, or is withdrawn, gives no Points to anyone.
| Your numbers | Answer | Points |
|---|---|---|
| 70 from the opening to the close | yes | 16 |
| 70 from the opening to the close | no | −24 |
| 50 for three quarters of the window, then 80 | yes | 5.25 |
| 99 from the opening | yes | 24.99 |
| 99 given halfway through the window | yes | 12.5 |
| 1 from the opening | yes | −73.01, the worst number stays above the floor |
Accumulated Points are the sum over every resolved question you answered, in every season. They set your level and your place in the standings. Points are a proper scoring rule: there is no number you can give that is not your own belief and still expect to do better.
Edge vs the Call
Edge vs the Call is one number: the Call’s mean Brier score minus yours, over the resolved questions that both you and the Call answered. Positive means you were closer to the answers than the house, negative that the house was closer. It shows from five such questions, and the count travels with it. A question on which the Call gave no number is left out of it.
Edge vs the Call measures you against the house, never against another reader, and never against a market. It is a reading of past questions, not a forecast of future ones.
The streak
The streak counts days. A day counts when you gave a first number on a question, or revised a number on a question that had new evidence since your last one: a new voice on a side, a new figure in the series. Until the platform marks new evidence on a question, every revision counts.
One missing day a calendar week, Monday to Sunday, is frozen: it does not break the run, and it does not add to it. The second missing day in a week ends the run. Your page shows the run alive now and the longest ever.
Levels
Levels follow accumulated Points: Reader from 0, Analyst from 100, Strategist from 400, Oracle from 1,200. A level is a name for how much scored forecasting stands behind a pseudonym, nothing more. It carries no privilege and buys nothing.
Achievements
Each achievement is earned once, at the instant it happens, and is never taken away. The list of version one:
- Blind call: a first number on a question with the case still closed, so you committed before reading either side.
- Both sides: the case opened before your first number on the question.
- Studied: a Part of The Buildout read to its end on a question before the question closed.
- Changed my mind: a revision that lowered your Brier score on a question that later resolved.
- Early: the number that counted was given 30 days or more before the close, and was on the right side.
- Against the crowd: the number that counted stood 20 points or more from the readers’ median you saw, from five readers, and was on the right side.
- Night owl: you saw a result on the day it resolved.
Achievements measure how you played, not how well: Early and Against the crowd ask for the right side, the rest ask only for the act. None adds Points.
Seasons
A season is a stretch of the calendar with its own standings. Season 0 runs from October 28 to December 31, 2026, and follows the earnings season. A question counts in the season in which it resolved, whatever season it opened in. There is no prize in Season 0: the standings are the record, and the end of the season brings each player a summary of their own numbers.
Any later season with a prize is announced before it opens, with its own rules on this page.
Standings and leagues
The standings list every player with three or more resolved questions in the season, by Points, then by Edge vs the Call (a player without one last), then by the number of resolved questions, then by pseudonym. The house, the readers’ median and the machines are measured by the same Brier score on the Scorecard, on their own rows; they are never in the players’ standings.
Sector standings are derived from the same scores: each player’s Points on the questions of one sector, with the same floor of three. They are the house’s and are not created by anyone.
A private league is a standings table for a group: friends, a desk, a class. Anyone with an account can start one, with a name and an invite code shown once; the code admits up to 500 members, and the owner can draw a new one, which stops the old. A league lives in the season it was made in and lists a member from one resolved question. Its table is shown to its members only; a university league shows its table in the open. An account may start at most 20 leagues, and leaving one is a click. Leagues carry no prize.
Takes
A take is up to 280 characters of why, written on a question after your own number is saved. It can cite up to five things from the question’s dossier with a tap: a voice, a claim, a chart, a source, a series. One take per player per question; writing again replaces the words and the citations and keeps the votes.
Takes are read only by players who have put a number on the question, so the blind mode holds. They are ordered by useful votes, then by the writer’s accumulated Points, then oldest first. Each player has one useful vote per take, never on their own, and can take it back. Each player can report a take once.
A take says why the answer will be yes or no. A take that reads as a recommendation to buy, sell or short anything is refused when it is written. Reported takes go to a review queue; a take that breaks the rules is hidden, and a hidden take is shown to no one and takes no vote. Takes are published under the pseudonym and are moderated; the rules of conduct are on the About page.
Follows
You can follow players (by pseudonym), theses, sectors and companies, up to 500 in all. Following a sector puts its questions into your Today’s five; following a player or a thesis puts their moves into your feed: a number with the one it replaced, a resolution that moved a thesis you follow, a take on a question you answered. Nobody is told who follows them.
Duels
A duel is one question between two players. You start it from your own number and send the link; the other player puts a number before seeing yours, so the blind mode holds. When the question resolves, the lower Brier score wins; equal scores are a tie; an inconclusive or withdrawn question settles nothing. The record between the two of you is kept and shown to both. A duel carries no Points and no prize.
Duels open with the second drop of the season; until then the link says so.
Scouts
A scout brings evidence to a question: a link and one line on what it shows, up to ten a day. The desk reviews each one. If it enters, the question’s dossier credits you by pseudonym and the source enters the house’s record. A scout is not a take and carries no vote.
Scouts open with the second drop of the season.
Results Night
Results Night is the evening the answers land. Each night has a page that follows the questions closing that day: the countdown, the three numbers, and each result as it is read, with the Brier score of every number beside it.
When the source is a quarterly filing, the company’s release comes out first and the filing follows, usually within ten days. On the night, the reading from the release is shown as provisional, with that word on it. The official result is recorded from the document the rule names, and that is the reading that scores. If the document contradicts the release, the official result corrects the provisional one, and the correction is a line in the result for everyone to read.
Withdrawn before the date
The house can withdraw a question before its date: when the rule turns out not to be readable against the source, when a mistake is found in the question, or when the owner vetoes a question within 24 hours of its approval. A withdrawn question is marked withdrawn before the date, with the reason. It counts for no one: no Brier score, no Points, no Edge vs the Call, no duel settled. A question is never withdrawn after its date, and a result once published is corrected in the open, never removed.
Playing without an account
You can play without an account. Your first number creates a random id of 32 characters in this browser; it and your numbers are kept under it, and nothing else is asked. After your third number, the platform offers to keep them: signing in, by passkey or by a link sent to your email, moves the guest’s numbers to the account with their original times, so what you forecast as a guest scores as yours. A guest’s numbers are never in the readers’ number or in the standings until they are claimed.
A player chooses a pseudonym of 3 to 24 letters, digits, underscores and hyphens, never an email address. The pseudonym is the only name the platform ever shows.
Who plays, and what a number is
DeepStack is for people 18 or older. Signing in confirms that you are 18 or older. We score claims, not people: a person quoted in a Part is never scored, and a player is scored only by their own choice, under the pseudonym they picked. There is no money in the game, nothing to buy, and no currency with a value outside it.
A number is your reading of a public question, not a trade and not advice. The Call, the machines and the readers’ number are published identically to everyone and know nothing about you. Nothing on the platform tells you what to buy or sell, and nothing on it is connected to an account that could.
Analysis and opinion, not investment advice.