05Proof

Open frontiers

Nine questions possession inference has not settled, each with the experiment that would settle it. This is where the next gains are.

  • Open questions
  • Experiments
  • Field
A settled core, the edge of what is known, and nine open questions beyond it. The first, possession through long ball gaps, carries the most weight.
Questions
Nine, each with a decisive test
Common requirement
Whole labelled matches, held out
Hardest
Possession through long ball gaps

The design question

Which questions are still open for everyone? Published results answer them one dataset at a time, under one provider’s definitions. None has a general answer yet. Each is stated with the evidence that would close it, and with the study that frames it.

The nine

Question What would settle it Framed in
Does ball-free inference recover possession through long gaps? Labelled matches with simulated tracking gaps The missing ball
Can a second witness fill tactical ball gaps? Independent sync, calibration and ball visibility Single source of truth
Two heads or three classes for alive, dead and team? The same evidence, both designs, held-out matches Alive or dead
Do identities need repair at all? Permutation and fragmentation tests Players without names
Does a graph network beat geometry with gradient boosting? Match-separated comparisons and ablations Geospatial-only
Does a decoder lift both official metrics? Spell-aware comparison on held-out matches Seven ways to decode
Are event spotters precise enough for boundaries? Boundary error measured in frames Events
Does calibration survive a change of provider? Reliability diagrams on another provider’s data Confidence
Do providers mean the same thing by possession? Published definitions, one match annotated both ways Two metrics

The first frontier, as an experiment

Hide the ball in windows of growing length on a labelled match, and plot how often possession is still right. The three curves are hypotheses, drawn as guesses. Nobody has published the measured ones.

What the questions share

  • The unit is the match. Every test needs whole matches held out. Frame-level splits answer nothing.
  • The condition is the gap. A result on fully tracked balls says little about the frames that decide the score.
  • The metric is the spell. Gains in frame accuracy that cost possessions are losses.

Where the leverage is

The first two questions concern the missing ball, and they carry the most weight. What we are building toward them is on the roadmap.

Our read

The frontier is the missing ball. Whoever shows, on held-out matches with realistic gaps, that hidden turnovers can be recovered will move the field more than any new backbone.

Sources

  1. Official evaluation (spellF1, macroF1)github.com
  2. PathCRF: ball-free possession paths from player trajectoriesgithub.com
  3. Bassek et al. (2025), DFL open tracking and event data, Scientific Data 12:195nature.com