05On the pitch

The Shrewdies decoder

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
Two evidence lanes, kept apart, feed one exact decode per half. The lower lane ran once, away from the match: it explored on simulated matches and let open DFL labels pick the profile.
Output
199,319 frames at 25 fps, each H, A or D
Runtime
Under 10 s on a laptop CPU
Determinism
Byte-identical output on every run
spellF1
0.340 against 0.235 for the organisers' baseline method
macroF1
0.693 against 0.383, on the same open DFL match

The architecture in one line

Read the match the way a match official would, then let an exact algorithm write the single most coherent possession story for each half. Evidence and decision are separate machines, and nothing is tuned on the match being graded.

Layer What it does The design study behind it
Observation Tracking and tactical tags. A missing ball stays missing. Geospatial-only inference
Evidence Scores home, away and dead on every frame, in two lanes. Evidence, then a decoder
Decision One exact decode per half over home, away, dead. Seven ways to decode a match
Selection Profiles from a Pareto front, the final one picked on open labels. Two metrics, one verdict
Delivery Pinned container, no network, read-only inputs, 15 behaviour tests. Reproducibility

Why each layer looks the way it does, constraint by constraint, is the subject of the On the pitch family page.

The evidence, exactly

Every frame scores three hypotheses. Each signal has a physical meaning.

Signal Football meaning Supports
A player within 1.5 m of a ball moving slower than 7 m/s Someone can actually play it That player’s team
Ball more than 1 m beyond a line Out of play Dead
Players almost still for 3 s Play has stopped Dead
Ball placed and motionless for 1.5 s Free kick, goal kick, kick-off Dead
Start of a tagged set piece A restart is coming Dead
A fast ball inside the pitch Play is live Not dead
A team-naming tag code A hint about which side That side, weighted

Three design rules govern how those signals are used. Each has its own study, so each is one line here.

  • A missing ball scores nothing. Possession stays with the last team in control: When the ball goes missing.
  • Alive-or-dead and which-team never share evidence. A tag cannot move the edge of a stoppage, and a unit test guarantees it: Alive or dead, then who.
  • Each tag code weighs what it proves. Its weight is its measured agreement with visible ball control: Events are not possession.

The decode, exactly

One Viterbi pass per half chooses the sequence that maximises evidence minus a fixed cost for every change of state. The fielded profile sets the team switch cost at 4.23 and the alive-or-dead switch cost at 9.43.

A change of state is bought only when the evidence for it exceeds the toll. A touch never pays. Sustained control pays after about four frames. A stoppage has to pay the higher toll. The curves are schematic.

Seven dials were searched: the two switch costs and five evidence weights. The control radius was not. It stays at 1.5 m, a physical constant, because a search left free pushed it to a value no footballer would recognise.

What it scored

The fielded profile against the organisers' baseline method, on an openly labelled DFL match with the ball hidden on 41% of frames, scored by the official code.

The fielded profile wins on both official metrics (the full table is on the Initial Challenge page). The baseline method invents possessions, 570 against a truth of 402. The decoder errs the other way, 249: turnovers hidden inside long tracking gaps surface late. The two other frozen profiles scored spellF1 0.249 and 0.286 on the same match, and how the three were produced and chosen is in Two metrics, one verdict.

What comes next

The architecture has clear seams, and each takes a stronger part without touching the rest. The first is the hidden ball: the single source of truth reconciles the video frames and both tracking feeds into one match state, so fewer frames reach the decoder silent. The rest of what we are building is on the roadmap.

Our read

An architecture is proven by what it refuses to do. This one refuses to invent a stoppage when the ball disappears, to let a tag stretch a dead ball, and to choose anything on the match it is graded on.

Our call

ConstraintReproducibility

Why it is deterministic and CPU-only

The organisers regenerate the submitted file themselves. The decode has no randomness, the dependencies are two pinned libraries, and the container is pinned by digest and runs without a network. The same bytes come out on every run, in under ten seconds.

Sources

  1. Technical report: Who has the ball?github.com
  2. Code: Shrewdies/MTAH2027, initial-challengegithub.com
  3. Official evaluation (spellF1, macroF1)github.com
  4. Bassek et al. (2025), DFL open tracking and event data, Scientific Data 12:195nature.com
  5. Vidal-Codina et al. (2022), automatic event detection from tracking datadoi.org