Updated Sep 9, 2026 · Week 1 board.
Studio holds eight slots, and each one takes a single decision from you. Between them they carry 56 options. You can fill them in any order. One of the eight is the slot the board tells you to decide first, and it says so before you have touched anything.
Each slot is one question with a fixed set of answers
Target asks what the model predicts. Scope asks which games it learns from. Features asks what it reads about each of those games, and Transforms asks how those inputs get reshaped before any fitting starts.
Model asks which method does the fitting. Calibration asks whether the confidence it reports is honest. Eval asks how the record gets graded. Sizing asks how a decided pick turns into a stake.
| Slot | What it decides | Options | One of them |
|---|---|---|---|
| Target | What the model predicts | 8 | ATS side pick |
| Scope | Which games it learns from | 12 | Exclude season openers |
| Features | What it reads about a game | 14 | Team recent form |
| Transforms | How an input gets reshaped | 7 | Exponential time decay |
| Model | Which method does the fitting | 7 | Gradient boosted trees |
| Calibration | How raw output becomes a probability | 3 | Platt scaling |
| Eval | How the record gets graded | 2 | Sealed holdout eval |
| Sizing | How a pick turns into a stake | 3 | Flat stake sizing |
Features and Transforms take more than one pick each. The other six take exactly one. A slot you leave alone is still a decided slot by the time the model trains, and Studio decides it for you.
Target goes first because the other seven depend on it
Studio ranks the next slot by how much it narrows the board and how much it changes the answer. Target wins that ranking on an empty board. Deciding it alone cuts 88% of what is left, and the page prints that figure before you have chosen anything. Scope narrows a shade more at 92%, offered as the alternative.
The plainer reason to start at Target is that the other seven read differently for every answer. A player prop model and a spread model do not want the same inputs. Until Studio knows which one you are after, the ranking under every other option is a guess.
The sport buttons do that job one step earlier. Choose NFL and two of the fourteen Features options go dark, because those two read baseball. Twelve stay live, and the Features chip shows both numbers.
Two buttons will fill the rest for you
Auto-finish reaches into every slot still open, picks the strongest prior option in each, and hands back a full board in one press. Collapse best does the same job a slot at a time, so each choice arrives where you can watch it happen. Either way it is one undo.
Templates go further than both. Applying ATS classic pre-sets four of the eight and fills the remaining four with their best option, so the board arrives with nothing left to decide. That is the quickest route to a trained model and the slowest route to understanding one.
Six ready-made boards
Studio ships six templates, and each one is a starting board rather than a finished model. ATS classic and a flat-stakes baseline floor each pre-set four of the eight slots. Two prop templates and a daily-fantasy one pre-set three or four apiece, and a sealed calibrated-ensemble board fills four.
Applying any of them replaces the whole board in one click, and the slots the template skips take their best option. It is one undo, so trying a template costs nothing you cannot take back. Apply one, then apply another, and the difference between the two boards is the quickest lesson in what a slot is for.
What happens the moment a slot lands
Studio re-solves the whole board after every decision. Options that no longer fit go dark rather than disappearing, and the board will tell you which rule turned each one off. A form refuses you; this board tells you why.
Nothing you do here is final. Undo steps back one decision, and Un-collapse everything empties the board and hands back all 56 options. Wrecking a board costs one button to reset, so the cheapest way to learn the eight slots is to break a few of them on purpose.
The read-out panel keeps a running verdict while you work. On an empty board it says eight open and not trainable yet. The Train model button carries that same count and offers to fill whatever is still open before it runs.
One path through the eight, for a spread model
- Target: ATS side pick, which answers which side of the posted number covers.
- Scope: Exclude season openers, because a week 1 game has no in-season history behind it and week one breaks models for that reason.
- Features: Team recent form, then Travel and rest. This slot takes several picks, so add the second one after the first model trains.
- Transforms: Exponential time decay, so recent games count for more than old ones.
- Model: Gradient boosted trees, which carries a minimum sample floor and refuses to fit on thin history.
- Calibration: Platt scaling, which turns raw output into a probability you can size a bet from.
- Eval: Sealed holdout eval, which keeps a block of games sealed until you publish and grades on them once.
- Sizing: Flat stake sizing, because the same stake on every bet keeps a wrong probability from doing extra damage.
That path costs nothing to run. The estimate panel prices the training run at $0 and puts it on your own machine or a free lane, never on a paid GPU.
The Receipts Drawer
The desk read on the eight: a slot you skip is still decided, only not by you. Auto-finish is a fine way to get a first model trained and a poor way to learn what any of it does. Decide Target and Scope by hand, then let the board finish the other six.
Order matters less than the ranking implies. What matters is that Calibration and Sizing get read before money moves, because those two turn a number into a stake. Watch the read-out panel change each time a slot lands, because that panel is the only place the cost of a choice shows up before you train.
FAQ
Do I have to fill all eight slots myself? No. Auto-finish fills every open slot with its strongest prior option in one press, and Collapse best does the same a slot at a time. Each of those actions is a single undo.
Which slot should I decide first? Target. Studio ranks it top on an empty board, where deciding it cuts 88% of what is left. Every other slot means something different once the board knows what the model predicts.
Why did some options go dark when I picked NFL? Two of the fourteen Features options read baseball, so an NFL board hides them and shows twelve. Studio names the rule doing the hiding instead of leaving the option quietly missing.
What does it cost to train a board? Nothing. The estimate panel prices a run at $0 and puts it on your own device or a free lane. You do not need an account to press the button.
