What the labels mean
H and A describe home and away possession. D means dead ball. X is unknown or an ungraded interval. Nearest player is evidence of control, not proof of it. An invisible ball can remain live.
Pitch coordinates are metres on a 105 × 68 m field. Player numbers identify tracking slots, not named footballers. Frames are zero-based in this explorer; public source frames start at 1. Half-time is shown as X. Research temporal models reset at period boundaries; the unchanged reference voting windows can span a break.
Models you can inspect
Nearest-team distance, an interpretable state machine, a four-state HMM, and a hierarchy separating live/dead evidence from conditional team possession. Original implementations remain intact, alongside Phase 1 ablations and the official smoothing references.
HMM and hierarchy use offline temporal decoding: later observations can affect earlier predictions. They are not live broadcast systems. Reference variants smooth the nearest-team signal, not the unavailable competition inputs.
Read the scores carefully
MacroF1 averages H/A/D, including absent classes as zero. Ground-truth X frames are excluded. Prediction X counts as a frame miss; for spell scoring the official scorer forward-fills then back-fills X within valid runs. H/A possession spells are greedily matched at IoU thresholds 0.3 to 0.7. D is evaluated by frame precision and recall.
Boundary diagnostics match typed possession starts and ends within 25 frames (1 s at 25 fps), excluding period edges. MAE covers matched boundaries only; missed and extra counts show what it omits. Undefined precision, recall or MAE appears as “n/a”. SpellF1 and MacroF1 use the unchanged official scorer; boundary diagnostics are a separate research metric.
How to read the observation weight
This is a heuristic observation weight, not a calibrated confidence score. It combines ball visibility, the fraction of available tracking slots and a penalty for implausible ball speed. Inactive substitute columns can lower slot coverage even when all on-pitch players are visible. X marks unknown predictions; a high spell score can still include X frames because the official temporal scorer fills them.
What synthetic evidence can show
The v2 illustration uses role-based steering with independent player targets, bounded latent speed/acceleration, continuous passes and players repositioning during stoppages. It is a scripted simulation, not a model calibrated to real tactics. Measurement noise and identity swaps can still create observed jumps. The original synchronized/frozen fixtures remain labelled as legacy stress tests. Noise, visibility and events do not cover the diversity of real football. A single scenario is a debugging example, not a generalization estimate. Historical experiments use paired trials and whole-trial intervals.
Two worlds, one observation
The observability explorer gives two hidden states exactly the same tracking. Any deterministic tracking-only model must return the same prediction in both. Its balanced accuracy across the contradictory target interval cannot exceed 50%.
Optional event cues are hypothetical, generated independently of tracking through a latent action process and a noisy annotation process. They can miss, shift, duplicate, conflict or carry wrong categories and teams. A single pair can illustrate information gain; it cannot establish its reliability. Shared uninformative tags carry no information about which world is true.
Public tracking, without accuracy claims
The public demonstrations and public playground use the first 1,000 frames from Metrica Sports anonymized Sample Games 1 and 2, normalized through a separate CSV adapter. No independent possession truth is available here. Predictions are exploratory and all accuracy metrics are withheld. Playground loss/noise changes only the observations; the underlying public movement is retained. The v2 synthetic sequence is a single continuous passage with no halftime; legacy fixtures test halftime handling.
Provenance & reproducibility
Code is MIT licensed; public data retains its source terms. Metrica sample commit e706dd5; official reference commit f8f2629. The build checks the official sources and 408 frozen Phase 1 files. No restricted competition data, manual competition labels or public five-minute tuning labels power this explorer.
Changed playground settings execute the existing Python modules on your device using pinned Pyodide 0.29.2. The site serves a bundled Python runtime, NumPy and pandas; saved examples work while it loads. Browser numerical versions differ from the native lock and are recorded with each result.
Full provenance manifest · Source, tests & reproduction scripts
Research reports: Phase 0 · Phase 1 · Phase 2 · Motion v2 validation. Source data: Metrica sample-data · Official challenge.