ATLAS
Every market we price. Every one we turn down. And why.
A statistical model prices sports markets, compares itself to where they are trading, and publishes a call only on disagreement — before the game, then graded in public. The rejections are published too, which is the part nobody else shows.
The publish rule changed on 23 August — calls now come from the 0–4 point band. The record above spans the rule it replaced. Under the current rule: 0-1 on 1 graded. Both, in full
Priced and published cover the same days — every candidate since 2026-08-11, when the evaluated field became public and insert-only. That is 2.3% of what the model looked at. Every one it turned down, and why →
Breakeven against prediction-market pricing is about 50.5%, not the 52.4% a sportsbook demands — so the model is nominally ahead on a sample too small to separate skill from luck. Nothing here asks you to take that on faith: the record, the rejections and the price checker are free and stay free. 4 further calls were published and then voided as duplicates; they are excluded here and still visible on the record. Every price we quoted is checkable against the order book →
We caught our own publish rule selecting badly — and pulled every live call the same day.
ATLAS published a call only when the model disagreed with the market by enough. Then we measured what those disagreements actually were, and most of them were not the model seeing something — they were the market moving off the average while our number stayed put. The bar was selecting for the thing our model is worst at.
So we withdrew every market, published the reasoning, and resumed only under a narrower rule that refuses the disagreements we cannot defend. Then we wrote down, in advance, the measurement that would stop us again — and what result would trigger it.
That is the email. Not picks, not a newsletter: what we find when we test ourselves, including the times it goes against us. It has already gone against us twice.
The record and the calibration are public and free, signed up or not — nothing there is behind the address. The evaluated field is public too: every candidate, its status and why it was turned down. The per-candidate prices inside it are the part Desk pays for.
You don’t have to trust us. That’s the point.
Every call states a price and a time. The exchange’s order book history is public and needs no account to read. So if a price we quoted was never available when we quoted it, anyone can see that — including you, in about ten seconds.
The verification page shows every published call against the book, including the ones we can’t fully prove: nine of the earliest were migrated from a system that never recorded a freeze time, and they are labelled as such rather than quietly counted as clean.
Check the record yourself →How it works
- 01
Price the market independently
The model builds its own fair value for a game from the underlying statistics. It never starts from the market price, so it can disagree with it.
- 02
Publish only on disagreement
Most days most games clear nothing. When the model and the market are close, there is no call — and no call is published rather than manufactured. Every candidate priced is kept, including the ones turned down and the reason why.
- 03
Freeze it before the game
The call is recorded with its price, its side and its reasoning before the event starts. It cannot be edited or deleted afterwards.
- 04
Grade it in public
Wins and losses are posted at their true frequency, in the same format. A record that only shows the good days is a marketing page.
What a call looks like
Two real calls from the record — one that won, one that lost. Both were frozen before their game with the price and reasoning attached, and both are checkable against the order book.
Chicago Cubs at Seattle Mariners: OVER 7.5
WIN- Market price
- 50¢-100
- ATLAS said
- 57.0%
- Edge claimed
- +7.0 pts
- Frozen
- 2026-08-23 02:31 UTC
Cleveland Guardians at Los Angeles Angels: OVER 7.5
LOSS- Market price
- 54¢-117
- ATLAS said
- 57.3%
- Edge claimed
- +3.3 pts
- Frozen
- 2026-08-24 00:31 UTC
Nearly the same edge, opposite outcomes — both frozen before their game, both permanent. That pair is the whole argument: anyone can show you the winner, and the loss is the part you cannot fake. Which of the two a process produces more often is what the calibration work exists to answer.
Every call, wins and losses alike →Three ways to read it
- The record — free, always
- Every call ever made, the calibration status, and today’s matchup and side. Enough to watch a call land before paying for one.
- Edge — $29/mo
- Today’s calls in full: the line, the entry price, the expected value and the reasoning.
- Desk — $99/mo
- The whole evaluated field, rejects included — every candidate priced that day with the model’s probability beside the market’s, and why each one was refused. The unbiased sample rather than the selected one. This is the one to read if you price markets yourself and want the inputs rather than the conclusion.
On sharing: quote a call, screenshot the field, post a row to your group — that is fine, and you do not need to ask. What is not fine is republishing the whole daily feed as your own or reselling access. The record has always been public; if you want to show someone a call, sending them a link costs you nothing and costs us nothing.
Compare the plans →What ATLAS covers
- Published today — MLB full-game totals
- Every call in the record above is a major-league total. It is the only market with enough graded history here to be published at all.
- Priced but not published — NFL
- NFL markets are priced every day and appear in the public field alongside the baseball. Nothing from them reaches the record, and nothing will until they have been graded long enough to mean something.
- Why the list is short
- Each sport and each market type earns publication on its own graded evidence, not on another’s. Baseball totals working says nothing about a baseball moneyline — it is a different reading of the same model, and it has never been graded. So markets are added when they have a record, not when a season starts.
What you do with a call
A call names a market, a side and the price it was frozen at — for example “Boston Red Sox at Pittsburgh Pirates, OVER 7.5, 52¢”. These are prediction markets: contracts that settle at $1 if the outcome happens and $0 if it does not, so a price of 52¢ is the market saying 52%. ATLAS does not place trades, hold money, or receive anything from any venue.
The same games trade in several places — Polymarket, Kalshi, NoVig and others — and you can act wherever you already have an account. One caveat that matters: every entry price on the record is the Polymarket ask at the moment the call was frozen, because that order book is public and needs no account to read, which is what makes the record checkable at all. Prices elsewhere will differ, sometimes in your favour. The call is the game, the side and the line — the venue is yours to choose.
A sportsbook takes the other side of your bet. When you win consistently it has every reason to cut your limits or close your account, and it does. An exchange does not take the other side — you are trading against other participants, and the venue earns the same whether you win or lose. Nobody gets shut out for being right. What does cap you is how much is actually on the book, which is the next paragraph.
Whether those markets are open to you depends on where you live, and that is worth checking before you sign up for anything here. The record and the evaluated field stay free and readable either way — plenty of people follow the model without ever taking a position.
On size: these books are thin next to a sportsbook. Liquidity is shown on every row of the field, and it ranges from a few hundred dollars to six figures on the same day. A call sitting on a shallow book cannot absorb much before the price moves against you. ATLAS has not measured how much size its edge survives, and does not claim a number it has not tested.
Follow the record from here.