Research · the lab notebook
We build in the open.
A research lab is only as good as what it publishes. This is how 17z reads football — the methods, the results on open data, and the beliefs underneath — written for the analyst and the scientist both. Not a product page. A notebook.
Representation
Before it answers anything,
it learns the shape of the game.
Method — state spaceA match as a trajectory
A match is a path, not a box score.
The model represents a game as a trajectory through a learned space — where it’s been, and where it’s heading next.
Method — the space of styles
Every way football can be played,
as one continuous space.
Writing
The notebook.
Each post is a method, a result, or a belief — worked through in public, with the reasoning and the uncertainty on the table.
01MethodWhy we train on fake data firstSimulation + open data builds a model that works before a single club shares theirs. The idea, and why it beats the cold start.Soon →02MethodShazam for football momentsTurning a passage of play into a fingerprint, then finding every one like it across a season — no tagging, no tape.Soon →03MethodThe mirror trick: teaching a model that left = rightGeometric equivariance, in plain language — and why it roughly doubles what scarce football data can teach.Soon →04ResultExpected Threat, reproduced on open dataA homage to Karun Singh’s xT — and a public check that our learned value surface holds up.Soon →05BeliefThe coastline of a matchMeasure the game finer and there is always more there. Why the edge lives in the micro.Soon →