How it works · the science
Tracking in. Understanding out.
One foundation model reads the game the way no product before it could — from the raw movement of all 22 players. Here’s the whole pipeline, honestly: what goes in, what it learns, and the four things it hands back.
01 — The inputTracking data
It reads positions, not highlights.
Every player, every second — position and motion, frame by frame. The model reads the sequence as it forms, not a frozen snapshot.
02 — What it learnsValue surface · xT
From raw movement, it learns where the game is actually won — the few spaces that decide it. Move across the surface to probe.
Z17
Fig. Z17 — terminal zoneValue surface · xT
— move across the surface to probe
03 — IntakeFootage → understanding
Feed it video. Get back understanding.
Raw broadcast in; positions, events and value out. This is how a club with no tracking deal still plugs in — and why the pitch’s built-in symmetries (a left-wing move is the right, mirrored) let it learn all this from scarce data.
04 — What it hands back
One model. Four answers.
01
Structured concepts
Presses, overloads, shape — pulled out in coach language.
02
State valuation
How dangerous is this moment? Expected Threat, learned.
03
Action valuation
Was that the right decision — for every player, every touch.
04
Retrieval
Turn any moment into a fingerprint; find every one like it.
Built on the lineage of Expected Threat (Karun Singh, 2018)Equivariance, per DeepMind · TacticAI (Nature, 2024)
Bring your data
Trained on open data. Sharpened on yours.
We train the model on simulation and open data, so it works out of the box. Plug in your own tracking — on your machines, your data never leaving the building — and it reads your season the same way.
Become a design partner