05Family 01 of 05

On the pitch

One architecture went from the board to the pitch. This page is the why: the constraints it was drawn under and the call each one forced. The what, with its numbers, is the study below.

The constraint The call it forced Offline, one laptop, days Coordinates carry the match No labels on the match Evidence, then a decoder No labels, exact rerun Viterbi writes the sequence Ball untracked on 41% Silence is no evidence A team, not an identity Reason about teams Automatic tags, no answers Tags calibrate themselves Public window: check only Open labels pick the profile Organisers rerun the file Deterministic, CPU-only Days, and an exact rerun One witness, for now
Constraint on the left, the call it forced on the right. Select a call to read the study that argues it.
  • 1architecture fielded and measured
  • 4hard constraints
  • 9calls they forced

Constraints first

We did not choose an architecture and then defend it. We wrote down what the Initial Challenge would not allow, and the design space shrank to one column. The rules themselves are read as engineering requirements in The rules are the architecture.

Four hard constraints down the side, five candidate designs across. Each constraint strikes what cannot meet it. One column is left standing.

From constraints to bytes

The same four constraints are visible in the running system. Inputs are read-only and nothing leaves the machine. The container is pinned and has no network. Two runs give the same bytes. The whole match fits inside the time a laptop CPU allows.

The fielded system as a running thing. Each constraint is pinned to the part it shaped: offline inputs, an exact rerun, a laptop's worth of compute, and the gates the file passes before it leaves.

How an architect chooses here

  1. Write the constraints before the options. Compute, labels, time, reproducibility. They are inputs to the design, on the board first.
  2. Strike, then choose. Every option that violates one is out, however fashionable. The choice is made among what is left.
  3. Name the constraint in every call. If no constraint forced a choice, it is a preference, and it is not written up as a call.
  4. Choose away from the match. A design selected on the data it is graded on has proved nothing.

The architecture that survived, its evidence, its decode and its measured result are one study: The Shrewdies decoder.

In this family1

  • The architecture we fielded, end to end: football evidence on every frame, an exact Viterbi decode per half, and every parameter chosen away from the match.

    • Fielded
    • Geospatial
    • Viterbi
    • CPU-only