05Our DNA

Solution Architecture

How The Shrewdies architect football intelligence. Before a model is chosen, the whole design space goes on one board: every source of evidence, every engine, what breaks each in a real match, and the proof that settles it.

Two strands that never merge, evidence and decision, joined by five questions. Every study in this section answers one of them.
  • 16studies, one question each
  • 5families of decisions
  • 9calls, each forced by a real constraint
  • 1architecture fielded and measured

The whole design space, on one board

01 The map Where evidence comes from Tracking Video 02 The engine How evidence becomes one sequence One model Evidence + decoder Experts + gate Segment referee 03 The hard parts What breaks a design in a real match One clock Cameras Identities Missing ball Stoppages Events 04 Proof How a design earns belief Two metrics Confidence Open frontiers One source of truth On the pitch: the route we fielded One state on every frame
Four families of decisions, left to right, and the route we fielded lit across them. Point at a column to isolate it. Select it to open the family.

Possession inference has one output, a state on every frame, and many honest routes to it. We draw all of them before building any. The board reads left to right, in the order the decisions have to be made, and the lit route is the one that went on the pitch. The dashed box is where the next one starts: the single source of truth, one match state reconciled from the video frames and both tracking feeds.

One question, one study

Every study here answers one design question, and every concept has one home. A study states the question so that two engineers would disagree about it, puts the options side by side, names what breaks each one in football terms, and says how to judge it on matches the design has never seen. What the problem, the data and the rules are is told in Challenges. The methods and tools of the field are mapped in Solutions.

Our calls

Some studies carry a box marked Our call. It appears only where a hard constraint of the Initial Challenge decided the design: the competition rules, the labels that exist, the data as delivered, the days available, or the need to regenerate the exact file. Each box states the constraint, the call and what it bought. Where nothing forced a choice, the study is knowledge and nothing else.

Five families of decisions

  1. Diagram: Home, Away, Dead, Cheapest path.

    01 of 05

    On the pitch

    The architecture we fielded, and the calls that shaped it. Each call traces back to one hard constraint.

    • Football evidence on every frame, one exact decode per half
    • spellF1 0.340 and macroF1 0.693 on open labels, against 0.235 and 0.383
    • Under ten seconds on a laptop CPU, the same bytes on every run

    1 entry

  2. Diagram: Video frames, Tactical track, Broadcast track, One match state.

    02 of 05

    The map

    Where the evidence comes from: coordinates, pixels, or every witness reconciled into a single source of truth.

    • Three witnesses of one match, each filling the gaps of the others
    • Tracking is physics you can audit; video is perception you must own
    • Reconcile, never blend: every value keeps its witness

    3 entries

  3. Diagram: Evidence per frame, One coherent story, Home, Away, Dead ball.

    03 of 05

    The engine

    How evidence on single frames becomes one coherent story of the half.

    • Separate what a frame says from what the match means
    • The same scores, seven decoders, seven different stories
    • An ensemble is worth exactly the independence of its errors

    3 entries

  4. Diagram: First half, Second half, Gap, Offset.

    04 of 05

    The hard parts

    Six problems that decide whether a design survives a real match. None of them is about the model.

    • A missing ball is no evidence. It is never a stoppage
    • A slot is not a player, and a touch is not a turnover
    • Alive or dead, and which team, are two separate questions

    6 entries

  5. Diagram: macroF1, spellF1, Pareto front, Fielded.

    05 of 05

    Proof

    How a design earns belief: on whole matches it has never seen, by a rule written in advance.

    • Two official metrics that pull the dials in opposite directions
    • Explore on simulation, decide on open labels, check on the fixture
    • Nine questions the field has not settled, each with its decisive test

    3 entries